[xwiki-devs] [DISCUSSION] Handling document translations in Solr Search
Hi devs, This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches: 1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing. 2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better. - The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query. If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title). -- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand. 3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory. Currently I have implemented option 1) in our existing Solr integration, which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content. During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.) On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2). All these problems basically come from the fact that there is no way to specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers. Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime - at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values). I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there) I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically. Any input on this would be highly appreciated, specially if others have more experience with Solr setups. Thanks, Eduard ---------- [1] http://markmail.org/message/kaxaka7lsbgo57ms [2] http://lucidworks.lucidimagination.com/display/lweug/Multilingual+Indexing+a... [3] http://drupalcity.de/session/language-specific-and-multilingual-full-text-se... [4] http://drupal.org/project/apachesolr_multilingual [5] http://stackoverflow.com/questions/4064880/create-new-core-directories-in-so... [6] http://info.basistech.com/blog/bid/171842/Indexing-Strategies-for-Multilingu...
Hello Eduard, it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields. (maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated? There's one catch with this approach which is new to me but seems to be quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See: http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar... The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Currently I have implemented option 1) in our existing Solr integration, which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language. However, indicating which language it is matched into is probably useful... Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2).
Generating sounds like an easy approach to me.
All these problems basically come from the fact that there is no way to specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime - at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values).
How would that look at query time?
I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema. Those that want particular fields and tunings need to write their own schema. The same UI could also include whether to include a phonetic track or not (then require reindexing). hope it helps. paul
Hi Paul, I was counting on your feedback :) On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names). Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml). Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice. Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :) Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to be quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar... The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr integration, which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running. Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead. Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way to specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime - at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values).
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own schema.
The same UI could also include whether to include a phonetic track or not (then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas. Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
Hi devs, Any other input on this matter? To summarize a bit, if we go with the multiple fields for each language, we end up with an index like: English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content The Solr configuration is generated by some XWiki UI that returns a zip that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance. This operation is to be performed each time an admin changes the indexed languages (rarely or even only once). Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in. Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy. One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized. Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.). WDYT? Thanks, Eduard On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected]>wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to be quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar... The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr integration, which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way to specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own schema.
The same UI could also include whether to include a phonetic track or not (then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
Hi Edy, On 11/26/2012 07:25 PM, Eduard Moraru wrote:
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a zip that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance.
IMHO this is a "must", not a "could" for embedded instances.
This operation is to be performed each time an admin changes the indexed languages (rarely or even only once).
There could be a reminder in the admin language UI.
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
But that's an exotic use case already I think. The common case is to query for the context language only.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
I think for applications this complexity/implementations details would benefit being hidden behind a "query builder" interface of some sort, WDYT ? Jerome
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected]>wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing. 2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better. I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query. Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title). You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand. True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to be quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar... The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory. Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr integration, which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content. During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.) True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2). Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way to specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers. Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values). How would that look at query time?
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically. The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own schema.
The same UI could also include whether to include a phonetic track or not (then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
Hi Jerome, Thanks for the reply. Please read below... On Tue, Nov 27, 2012 at 5:27 PM, Jerome Velociter <[email protected]>wrote:
Hi Edy,
On 11/26/2012 07:25 PM, Eduard Moraru wrote:
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a zip that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance.
IMHO this is a "must", not a "could" for embedded instances.
This operation is to be performed each
time an admin changes the indexed languages (rarely or even only once).
There could be a reminder in the admin language UI.
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
But that's an exotic use case already I think. The common case is to query for the context language only.
Even so, the developer would still have to do something like "title_${xcontext.language}:something AND doccontent_${xcontext.language}:somethingElse ..." which is not really nice.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
I think for applications this complexity/implementations details would benefit being hidden behind a "query builder" interface of some sort, WDYT ?
I also thought about something in that direction. The current query logic uses XWiki`s query API, so the query processing part would have to be done inside the query.execute() method. Something like expanding multilingual fields: "title:something AND doccontent:somethingElse AND lang:en" becomes "(doccontent_en:something OR doccontent_fr:something OR ...) AND (title_en:something OR title_fr:something OR ...) AND lang:en" (based on the currently supported -- or just indexed -- languages) Maybe we could also allow a parameter that disables this default action for queries that want to be explicit and don`t want this field expansion. I wonder if such a simple solution would fix our query issues or if it will create new ones that we don`t yet know about :) Thanks, Eduard
Jerome
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected]
wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC
project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3 different approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on
the
fields, having to resort to (at best) basic tokenization and
lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an
english-centered
chain of analyzers); *_fr of type text_fr (focused on french, etc.),
thus
making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you
want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open
Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields
(ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern,
requiring you to know beforehand which languages you want to support
(for
example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to be
quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.**org/mod_mbox/lucene-solr-user/** 201211.mbox/%3Czarafa.**509ccb61.698a.**1d02345614818807@mail.** openindex.io%3E<http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%[email protected]%3E> The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each
language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and
that
you have a nice separation - The disadvantage is that it's difficult to get it right
(administrative
issues) and then you also have the (considerable) problem of having to
fix
the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider
the
others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance,
we
could provide a way to generate the config files and write them to the
data
directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr integration,
which is also more or less compatible with our existing Lucene queries,
but
I would like to find a better solution that actually analyses the
content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query
complexity,
user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but
retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3]
about how Drupal handles its Solr integration and, particularly, a
plugin
[4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and
the
ui/extension gives you a zip with the configuration you need to use in
your
(remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to
edit
xml files and could also still be useful for other cases, not only
option
2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way to
specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that
would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when
you
have the lang information (in the solrDocument to be indexed), but when
you
perform the query, you can not analyze the query text since you do not
know
the language of the field you're querying (it was determined at runtime
-
at index time) and thus do not know what operations to apply to the
query
(to reduce it to the same form as the indexed values).
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem that
elaborates on the complexities and limitations of each options. (Ignore
the
Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable
while
not restrictive. I will probably also send a mail on the Solr list to
get
some more input from there, but I get the feeling that whatever
solution we
choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars),
since
it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own
schema.
The same UI could also include whether to include a phonetic track or not (then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with
Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul
______________________________**_________________ devs mailing list [email protected] http://lists.xwiki.org/**mailman/listinfo/devs<http://lists.xwiki.org/mailman/listinfo/devs>
______________________________**_________________
devs mailing list [email protected] http://lists.xwiki.org/**mailman/listinfo/devs<http://lists.xwiki.org/mailman/listinfo/devs>
______________________________**_________________ devs mailing list [email protected] http://lists.xwiki.org/**mailman/listinfo/devs<http://lists.xwiki.org/mailman/listinfo/devs>
Le 28 nov. 2012 à 16:52, Eduard Moraru a écrit :
"title:something AND doccontent:somethingElse AND lang:en" becomes "(doccontent_en:something OR doccontent_fr:something OR ...) AND (title_en:something OR title_fr:something OR ...) AND lang:en" (based on the currently supported -- or just indexed -- languages)
Just a methodological note: we should not be programming queries by combining strings but by combining Solr query objects which are rich, flexible, and easy to manipulate. However, speaking this way in mails is by far the best way! The mechanism you describe above is exactly how I would do it: my query expansion start with a q parameter and a context. Based on the context a weighted list of languages is extracted (context.language, then browser languages). I give the q parameter to the query-parser, with a whitespace analyzer. I obtain generally a boolean-query (otherwise I wrap it). Then I recursively go through the components of the boolean-queries. Each TermQuery of the default field is expanded into a disjunction of the TermQuery in each of the language (with their weight and with the right analyzer). Same for each phrasequery, wildcardquery, and fuzzyquery (not sure). Finding the analyzer defined in the Solr schema is not too hard. I can share code if you wish. Paul
Hi Edy, I'm not a huge fan of the title_fr title_en title_morelanguages approach as indeed it seems to be quite complex at the query level. I was more leaning towards multiple indexes if we can query them globally but I understand this is complex too. Now let's see the use cases that are hugely important: 1/ Make sure that if you decide your wiki is monolingual: - the indexing uses the specific language analyzer - make sure the query uses the specific language analyzer - make sure the search looks in all content even if the language setting of the document is wrongly set (consider all documents being of the specific language) 2/ Allow is a wiki is multi-lingual: - search in the language you decide (maybe the UI should display a language choice for the query) - search in content that is analyzed in the proper language when the content is declared in this language - allow to specify if you want to restrict your search to documents declared in the language of your query, versus search more widely in all documents accross languages. If you search in only the language of the query only one document can show up but it should point to the right translation that matches, if you search in multiple languages then you can show individual translations. - allow technical users to search for all documents across all languages (where the language analysis does not really matter)
From an admin point of view it makes good sense for the admin to be able to specify in a multilingual wiki which language analysis should be activated, and then have this transmitted to SOLR to properly configure the engine. Reindexing is ok when changing the configuration.
I believe in the end wether you use multiple fields with _fr _en or multiple SOLR cores, as long as you can query accross SOLR cores is a bit the same. If you cannot run a query merging multiple indexes then the first solution is kind of absolutely necessary as it would be the only one allowing to search across all languages. Maybe a solution would be to create one index per language and index ALL content regardless of it's language using the language analyzer of that index. This would allow to have better results even though the users have badly tagged the language of a document, and it's only the job of the UI to limit the search to only the language of the query, or all documents. So you could have a configuration in the admin that says: 1/ Create an English Index 2/ Create an additional French index The UI would allow to search in English and French, + would add a language restriction for the documents. In the future if we are able to "detect" the language of the documents we could add a lucene field with the "detected" language instead of the "provided" language of the documents, therefore increasing the quality of searches only on documents of a specific language. This later solution would be the only one that would really work on file attachements as we have no information about the specific language of file attachements (or even XWiki objects) which are attached to the main document and not to the translated document. This later issues shows that a search on "only french content" should still include the attachements because we have no idea if the attachments are "french" or "english". Ludovic 2012/11/26 Eduard Moraru <[email protected]>
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a zip that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance. This operation is to be performed each time an admin changes the indexed languages (rarely or even only once).
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected]
wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3
different
approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a new language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if
not
hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to be quite important to implement this approach: the idf should be modified,
the
Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar...
The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr
integration,
which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way
to
specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem that elaborates on the complexities and limitations of each options.
(Ignore
the
Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own schema.
The same UI could also include whether to include a phonetic track or
not
(then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost
Hi Ludovic, Thanks for the reply. Please read below... On Tue, Nov 27, 2012 at 5:44 PM, Ludovic Dubost <[email protected]> wrote:
Hi Edy,
I'm not a huge fan of the title_fr title_en title_morelanguages approach as indeed it seems to be quite complex at the query level. I was more leaning towards multiple indexes if we can query them globally but I understand this is complex too.
Now let's see the use cases that are hugely important:
1/ Make sure that if you decide your wiki is monolingual: - the indexing uses the specific language analyzer - make sure the query uses the specific language analyzer - make sure the search looks in all content even if the language setting of the document is wrongly set (consider all documents being of the specific language)
You mean that, if the wiki is monolingual, we should ignore the language filter and hardcode it to "All languages", right? However, what would be the advantage of this? Why would be want to pollute the results with irrelevant documents (caused by a probable recent configuration change that went from multi-lingual to mono-lignual)? Wasn't that the whole reason why the admin switched to mono-lingual?
2/ Allow is a wiki is multi-lingual: - search in the language you decide (maybe the UI should display a language choice for the query)
We already support this, by using the "Filtered Search" option and selecting the language. - search in content that is analyzed in the proper language when the
content is declared in this language - allow to specify if you want to restrict your search to documents declared in the language of your query, versus search more widely in all documents accross languages. If you search in only the language of the query only one document can show up but it should point to the right translation that matches, if you search in multiple languages then you can show individual translations.
I think this one is the same as the first bullet above.
- allow technical users to search for all documents across all languages (where the language analysis does not really matter)
Do you mean as an API? What exactly do you mean by "language analysis does not really matter"? Any Example?
From an admin point of view it makes good sense for the admin to be able to specify in a multilingual wiki which language analysis should be activated, and then have this transmitted to SOLR to properly configure the engine. Reindexing is ok when changing the configuration.
I believe in the end wether you use multiple fields with _fr _en or multiple SOLR cores, as long as you can query accross SOLR cores is a bit the same. If you cannot run a query merging multiple indexes then the first solution is kind of absolutely necessary as it would be the only one allowing to search across all languages.
Maybe a solution would be to create one index per language and index ALL content regardless of it's language using the language analyzer of that index. This would allow to have better results even though the users have badly tagged the language of a document, and it's only the job of the UI to limit the search to only the language of the query, or all documents.
So you could have a configuration in the admin that says:
1/ Create an English Index 2/ Create an additional French index
The UI would allow to search in English and French, + would add a language restriction for the documents.
Applying the language specific analyzers (for Chinese, for example) to all the documents will just create a mess for all the documents that do not match the analyzer's language. I`m not sure the results for the badly-indexed languages will make any sense to users. Also, this is very similar to the multi-core approach (one core per language), just that you also add documents that are indexed with the wrong analyzers. We have the same problem regarding merging relevance scores across indexes (cores) that is a big turn-off for the original multi-core approach.
In the future if we are able to "detect" the language of the documents we could add a lucene field with the "detected" language instead of the "provided" language of the documents, therefore increasing the quality of searches only on documents of a specific language.
In the previous discussions (on the GSoC thread) we agreed that language in XWiki is known before-hard, so no recognition is required, at least not at document level.
This later solution would be the only one that would really work on file attachements as we have no information about the specific language of file attachements (or even XWiki objects) which are attached to the main document and not to the translated document.
Yes, this is a problem right now. AFAIU, the plan [1] is to support translated objects and maybe attachments as well. Until then, we could either: 1) Use the original document's language to index the attachment's content 2) Use a language detection library to try to detect the attachment content's language and index it accordingly. The above could also be applied for objects and their properties. ---------- [1] http://jira.xwiki.org/browse/XWIKI-69
This later issues shows that a search on "only french content" should still include the attachements because we have no idea if the attachments are "french" or "english".
(The paragraphs below discuss on what currently exists and what could be done, ignoring the possible language detection mentioned above) Right now a document also indexes the object's properties in a field called "objcontent". I do this for all translations, thus duplicating the field's value in all translations. I can do the same for attachments. The purpose is, indeed, to be able to find document translations based on hits in their objects/attachments. If a language filter is used and there is a hit in an object, only one document is returned. If there are no language filters, all translations will be returned. However, if we search for the object/property/attachment itself, it will only be assigned to one language: the language of the original document. This means that if we search for all languages, the object itself will be found too (there is no language filter used). If we add a language filter that is different from the object/property/attachment's original document language, the object/property/attachment will not be found. Maybe we can come up with some processing of the query in the search application, that applies the language filter only for documents: ((-type:"OBJECT" OR -type:"OBJECT_PROPERTY" OR -type:"ATTACHMENT") OR lang:"<userSelectedLanguage>") -- writing it like this because the default operand is AND in the query filter clause that we use in the Search application. The problem with this is that that, when a language filter is used, the object/property/attachments that are now included in the results might not have the specified language and will pollute the results. Thanks, Eduard
Ludovic
2012/11/26 Eduard Moraru <[email protected]>
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a zip that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance. This operation is to be performed each time an admin changes the indexed languages (rarely or even only once).
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected]
wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> wrote:
Hello Eduard,
it's nice of you to see you take this further.
This issue has already been previously [1] discussed during the GSoC project, but I am not particularly happy with the chosen approach. When handling multiple languages, there are generally[2][3] 3
different
approaches:
1) Indexing the content in a single field (like title, doccontent, etc.) - This has the advantage that queries are clear and fast - The disadvantage is that you can not run very well tuned analyzers on the fields, having to resort to (at best) basic tokenization and lowercasing.
2) Indexing the content in multiple fields, one field for each language (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) - This has the advantage that you can easily specify (as dynamic fields) that *_en fields are of type text_en (and analyzed by an english-centered chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
- The disadvantage is that querying such a schema is a pain. If you want all the results in all languages, you end up with a big and expensive query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
If you want just some language, you have to read the right fields (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
-- Also, the schema.xml definition is a static one in this concern, requiring you to know beforehand which languages you want to support (for example when defining the default fields to search for). Adding a
new
language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems to
be
quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar...
The solution sketched there sounds easy but I have not tried it.
3) Indexing the content in different Solr cores (indexes), one for each language. Each core requires it's on directory and configuration files. - The advantage is that queries are clean to write (like option 1) and that you have a nice separation - The disadvantage is that it's difficult to get it right (administrative issues) and then you also have the (considerable) problem of having to fix the relevancy score of a query result that has entries from different cores; each core has it's own relevancy computed and does not consider the others. - To make it even worst, it seems that you can not [5] also push to a remote Solr instance the configuration files when creating a new core programatically. However, if we are running an embedded Solr instance, we could provide a way to generate the config files and write them to the data directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
Currently I have implemented option 1) in our existing Solr
integration,
which is also more or less compatible with our existing Lucene queries, but I would like to find a better solution that actually analyses the content.
During GSoC, option 2) was preferred but the implementation did not consider practical reasons like the ones described above (query complexity, user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
On a related note, I have also watched an interesting presentation [3] about how Drupal handles its Solr integration and, particularly, a plugin [4] that handles the multilingual aspect. The idea seen there is that you have this UI that helps you generate configuration files, depending you your needs. For instance, you (admin) check that you need search for language English, French and German and the ui/extension gives you a zip with the configuration you need to use in your (remote or embedded) solr instance. The configuration for each language comes preset with the analyzers you should use for it and the additional resources (stopwords.txt, synonims.txt, etc.). This approach helps with avoiding the need for admins to be forced to edit xml files and could also still be useful for other cases, not only option 2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
All these problems basically come from the fact that there is no way
to
specify in the schema.xml that, based on the value of a field (like the field "lang" that stores the document language), you want to run this or that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
at index time) and thus do not know what operations to apply to the query (to reduce it to the same form as the indexed values).
Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that would call other analyzers at runtime, based on the value of the lang field. However, this solution could only be applied at index time, when you have the lang information (in the solrDocument to be indexed), but when you perform the query, you can not analyze the query text since you do not know the language of the field you're querying (it was determined at runtime
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
I have also read another interesting analysis [6] on this problem
that
elaborates on the complexities and limitations of each options. (Ignore the Rosette stuff mentioned there)
I have been thinking about this for some time now, but the solution is probably somewhere in between, finding an option that is acceptable while not restrictive. I will probably also send a mail on the Solr list to get some more input from there, but I get the feeling that whatever solution we choose, it will most likely require the users to at least copy (or even edit) some files into some directories (configurations and/or jars), since it does not seem to be easy/possible to do everything on-the-fly, programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their own schema.
The same UI could also include whether to include a phonetic track or
not
(then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
2012/11/28 Eduard Moraru <[email protected]> > Hi Ludovic, > > Thanks for the reply. Please read below... > > On Tue, Nov 27, 2012 at 5:44 PM, Ludovic Dubost <[email protected]> wrote: > > > Hi Edy, > > > > I'm not a huge fan of the title_fr title_en title_morelanguages approach > as > > indeed it seems to be quite complex at the query level. I was more > leaning > > towards multiple indexes if we can query them globally but I understand > > this is complex too. > > > > Now let's see the use cases that are hugely important: > > > > 1/ Make sure that if you decide your wiki is monolingual: > > - the indexing uses the specific language analyzer > > - make sure the query uses the specific language analyzer > > - make sure the search looks in all content even if the language > setting > > of the document is wrongly set (consider all documents being of the > > specific language) > > > > You mean that, if the wiki is monolingual, we should ignore the language > filter and hardcode it to "All languages", right? > > However, what would be the advantage of this? Why would be want to pollute > the results with irrelevant documents (caused by a probable recent > configuration change that went from multi-lingual to mono-lignual)? Wasn't > that the whole reason why the admin switched to mono-lingual? > If the user is monolingual, we can safely ignore the language setting of each document and only the "main" document will be shown anyway to the user. So we should make sure we search on ALL documents that are available to the user. The user might have to "reindex" to make sure this is properly taken into account by the search engine. What is important here, is that even if the wiki is set to "fr", then even if documents have "en" for the main language they will still show up in the search. The opposite would be bad. > > > > 2/ Allow is a wiki is multi-lingual: > > - search in the language you decide (maybe the UI should display a > > language choice for the query) > > > > We already support this, by using the "Filtered Search" option and > selecting the language. > Sure. I was just repeating what is important. > > - search in content that is analyzed in the proper language when the > > content is declared in this language > > - allow to specify if you want to restrict your search to documents > > declared in the language of your query, versus search more widely in all > > documents accross languages. If you search in only the language of the > > query only one document can show up but it should point to the right > > translation that matches, if you search in multiple languages then you > can > > show individual translations. > > > > I think this one is the same as the first bullet above. > > No I think it's different. The user can decide to make his search in "French" but look in the "French" + "English" dataset. For me if both documents with the same page name match, both should come out separately. > > > - allow technical users to search for all documents across all > languages > > (where the language analysis does not really matter) > > > > Do you mean as an API? > Not specifically as an API > > What exactly do you mean by "language analysis does not really matter"? Any > Example? > I mean here that as a technical user your objective is to make sure your search spans ALL content in the wiki. In this case you don't care about stemming and such. The standard UI could take this into account by choosing "any language search" and with data set "all languages". We just need to make sure that this won't exclude any content of the search (Here is an example of a case where such exclusion might occur. Suppose you have done only "French" and "English" indexing and there is "German" content also in your wiki but since you have not asked for German search in your wiki you don't have title_de and content_de fields or don't have a german specific solr index (in the other method), then your german content would not be indexed at all ?) > > > > > > From an admin point of view it makes good sense for the admin to be able > to > > specify in a multilingual wiki which language analysis should be > activated, > > and then have this transmitted to SOLR to properly configure the engine. > > Reindexing is ok when changing the configuration. > > > > I believe in the end wether you use multiple fields with _fr _en or > > multiple SOLR cores, as long as you can query accross SOLR cores is a bit > > the same. If you cannot run a query merging multiple indexes then the > first > > solution is kind of absolutely necessary as it would be the only one > > allowing to search across all languages. > > > > Maybe a solution would be to create one index per language and index ALL > > content regardless of it's language using the language analyzer of that > > index. This would allow to have better results even though the users have > > badly tagged the language of a document, and it's only the job of the UI > to > > limit the search to only the language of the query, or all documents. > > > > So you could have a configuration in the admin that says: > > > > 1/ Create an English Index > > 2/ Create an additional French index > > > > The UI would allow to search in English and French, + would add a > language > > restriction for the documents. > > > > Applying the language specific analyzers (for Chinese, for example) to all > the documents will just create a mess for all the documents that do not > match the analyzer's language. I`m not sure the results for the > badly-indexed languages will make any sense to users. > > I undestand the issue here, but in most cases the user will say "french search" on "french content", he will only expand to non french content if he was not satisfied by his search. What is just allowed here is to also search in "french" inside all content. That would cover content that would have the bad language setting as well as any other content. The results might be a bit noisy but I don't think it's a big issue. > Also, this is very similar to the multi-core approach (one core per > language), just that you also add documents that are indexed with the wrong > analyzers. We have the same problem regarding merging relevance scores > across indexes (cores) that is a big turn-off for the original multi-core > approach. > > This is a more serious issue. If it's hard to merge the results spanning over multiple cores this could be a showstopper. However the solution of having only one Lucene document for all languages is not so cool either as it would make it difficult to know which ones of the languages has matched and present them separately with separate scores. It's really the core issue to decide on. What are the benefits and drawbacks of the different solutions. For each solution is there something in the UI that you cannot do ? So far I've heard: 1/ Presenting different scores for documents in different languages with the same doc name if the title_fr,content_fr method is used 2/ Merging scores accross indexes in multicore approach Other ? Can we list them in a wiki page ? > > > > In the future if we are able to "detect" the language of the documents we > > could add a lucene field with the "detected" language instead of the > > "provided" language of the documents, therefore increasing the quality of > > searches only on documents of a specific language. > > > > In the previous discussions (on the GSoC thread) we agreed that language in > XWiki is known before-hard, so no recognition is required, at least not at > document level. > > Let's forget this > > > > > This later solution would be the only one that would really work on file > > attachements as we have no information about the specific language of > file > > attachements (or even XWiki objects) which are attached to the main > > document and not to the translated document. > > > > Yes, this is a problem right now. AFAIU, the plan [1] is to support > translated objects and maybe attachments as well. Until then, we could > either: > 1) Use the original document's language to index the attachment's content > This is not a good solution. If I understand correctly we could not end up not searching in a french attachment because the original document is in marked "en". I'm for Paul's solution to index objects and attachments in each translation (if we have separate entities for translated documents). I understand that in the title_fr,content_fr approach this problem does not happen. > 2) Use a language detection library to try to detect the attachment > content's language and index it accordingly. > > Not sure we can for now > The above could also be applied for objects and their properties. > ---------- > [1] http://jira.xwiki.org/browse/XWIKI-69 > > > > > > This later issues shows that a search on "only french content" should > still > > include the attachements because we have no idea if the attachments are > > "french" or "english". > > > > (The paragraphs below discuss on what currently exists and what could be > done, ignoring the possible language detection mentioned above) > > Right now a document also indexes the object's properties in a field called > "objcontent". I do this for all translations, thus duplicating the field's > value in all translations. I can do the same for attachments. The purpose > is, indeed, to be able to find document translations based on hits in their > objects/attachments. If a language filter is used and there is a hit in an > object, only one document is returned. If there are no language filters, > all translations will be returned. > It seems we have to do this for now > > However, if we search for the object/property/attachment itself, it will > only be assigned to one language: the language of the original document. > This means that if we search for all languages, the object itself will be > found too (there is no language filter used). If we add a language filter > that is different from the object/property/attachment's original document > language, the object/property/attachment will not be found. > > Maybe we can come up with some processing of the query in the search > application, that applies the language filter only for documents: > > ((-type:"OBJECT" OR -type:"OBJECT_PROPERTY" OR -type:"ATTACHMENT") OR > lang:"<userSelectedLanguage>") -- writing it like this because the default > operand is AND in the query filter clause that we use in the Search > application. > > The problem with this is that that, when a language filter is used, the > object/property/attachments that are now included in the results might not > have the specified language and will pollute the results. > I'm not sure I understand. We have an "objcontent" field for each translation that has the full content of objects and properties, but individual object fields we don't have in each translations ? The more I see all the issues, the more I lean towards a separate index per language solution. The reason I do is that the main need is for a non English user to have very relevant results in his own language. Therefore we need to make sure that all content that the users have published has the chance to be analyzed using the non English language analyzer. So indexing all objects and attachments with the relevant language analyzer is the solution. This is also why I proposed to index all content in this specific index regardless of the language declared, which would only be used in the UI to limit searches to the specific language. In this view: 0/ There would be a language specific index per language with the objects and attachments indexed only in the language of the index 1/ The user chooses the language in which he searches 2/ Automatically that sets the index to be used to be the "french" index 3/ Automatically that presets to limit the span of the search to declared "french" documents 4/ The user can decide to go for non french documents at his own risks knowing that the results might be weird because of wrong analysis (this is what happens today with english analysis over french documents) The benefit here is that you don't have a merging score over multiple index issue, since you would never have to do a search across multiple indexes. Searches are still simple to write. By default results are quite relevant since you limit the search the french declared documents (this would be the same as limiting your search to title_fr, content_fr) and still cover what needs to be covered (objects and attachments). Another benefit is that this falls back gracefully to monolingual as you just have to have one index in the language declared for the monolingual wiki. The drawback is that the indexing is more costly and there is duplicated content in the index. Howerver it is the Admin that say which languages he wants available and he takes responsibility of the ressources this needs. Could this solution work ? Ludovic > > > Thanks, > Eduard > > > > Ludovic > > > > > > > > 2012/11/26 Eduard Moraru <[email protected]> > > > > > Hi devs, > > > > > > Any other input on this matter? > > > > > > To summarize a bit, if we go with the multiple fields for each > language, > > we > > > end up with an index like: > > > > > > English version: > > > id: xwiki:Main.SomeDocument_en > > > language: en > > > space: Main > > > title_en: XWiki document > > > doccontent_en: This is some content > > > > > > French version: > > > id: xwiki:Main.SomeDocument_fr > > > language: fr > > > space: Main > > > title_fr: XWiki document > > > doccontent_fr: This is some content > > > > > > The Solr configuration is generated by some XWiki UI that returns a zip > > > that the admin has to unpack in his (remote) Solr instance. This could > be > > > automated for the embedded instance. This operation is to be performed > > each > > > time an admin changes the indexed languages (rarely or even only once). > > > > > > Querying such a schema is a bit tricky when you are interested in more > > than > > > one language, because you have to add all the clauses (title_en, > > title_fr, > > > etc.) specific to the languages you are interested in. > > > > > > Some extra fields might also be added like title_ws (for whitespace > > > tokenization only) that have various approaches to the indexing > > operation, > > > with the aim of improving the relevancy. > > > > > > One solution to simplify the query for API clients would be to use > fields > > > like "title" and "doccontent" and to put as values very lightly (or not > > at > > > all) analyzed content, as Paul suggested. This would allow applications > > to > > > write simple (and backwards compatible maybe) queries that will still > > work, > > > but will not catch some of the nuances of specific languages. As far as > > > I`ve seen until now, applications are not very interested in nuances, > but > > > rather in filtering the results, a task for which this solution might > be > > > well suited. Of course, nothing stops applications from using the *new* > > and > > > more expressive fields that are properly analized. > > > > > > Thus, the search application will be the major beneficiary of these > > > analyzed fields (title_en, title_fr, etc.), while still allowing > > > applications to get their job done (trough generic, but less/not > analized > > > fields like "title", "doccontent", etc.). > > > > > > WDYT? > > > > > > Thanks, > > > Eduard > > > > > > > > > > > > > > > On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru <[email protected] > > > >wrote: > > > > > > > Hi Paul, > > > > > > > > I was counting on your feedback :) > > > > > > > > On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]> > > > wrote: > > > > > > > >> > > > >> Hello Eduard, > > > >> > > > >> it's nice of you to see you take this further. > > > >> > > > >> > This issue has already been previously [1] discussed during the > GSoC > > > >> > project, but I am not particularly happy with the chosen approach. > > > >> > When handling multiple languages, there are generally[2][3] 3 > > > different > > > >> > approaches: > > > >> > > > > >> > 1) Indexing the content in a single field (like title, doccontent, > > > etc.) > > > >> > - This has the advantage that queries are clear and fast > > > >> > - The disadvantage is that you can not run very well tuned > analyzers > > > on > > > >> the > > > >> > fields, having to resort to (at best) basic tokenization and > > > >> lowercasing. > > > >> > > > > >> > 2) Indexing the content in multiple fields, one field for each > > > language > > > >> > (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) > > > >> > - This has the advantage that you can easily specify (as dynamic > > > fields) > > > >> > that *_en fields are of type text_en (and analyzed by an > > > >> english-centered > > > >> > chain of analyzers); *_fr of type text_fr (focused on french, > etc.), > > > >> thus > > > >> > making the results much better. > > > >> > > > >> I would add one more field here: title_ws and text_ws where the full > > > text > > > >> is analyzed just as words (using the whitespace-tokenizer?). > > > >> A match there would even be preferred to a match in the below > > > text-fields. > > > >> > > > >> (maybe that would be called title and text?) > > > >> > > > >> > - The disadvantage is that querying such a schema is a pain. If > you > > > want > > > >> > all the results in all languages, you end up with a big and > > expensive > > > >> > query. > > > >> > > > >> Why is this an issue? > > > >> Dismax does it for you for free (thanks to the "form" parameter that > > > >> gives weight to each of the fields). > > > >> This is an issue only if you start to have more than 100 languages > or > > > >> so... > > > >> Lucene, the underlying engine of solr, handles thousands of clauses > > in a > > > >> query without an issue (this is how prefix-queries are handled... > they > > > are > > > >> expanded to a query for any of the term that matches the prefix, a > > > setting > > > >> deep somewhere, which is about 2000 avoids this to explode). > > > >> > > > > > > > > Sure, Solr is great when you want to do simple queries like "XWiki > Open > > > > Source", however, since in XWiki we also expose the Solr/Lucene query > > > APIs > > > > to the platform, there will be (as as it is currently with Lucene) a > > lot > > > of > > > > extensions wanting to do search using this API. These extensions > (like > > > the > > > > search suggest for example, rest search, etc) want to do something > like > > > > "title:'Open Source' AND type:document AND doccontent:XWiki". Because > > > > option 2) is so messy in it's fields, it would mean that the > extension > > > > would have to come up with a query like "title_en:'Open Source' AND > > > > type:document AND doccontent_en:XWiki" (assuming that it is only > > limited > > > to > > > > the current -- english or whatever -- language; what happens if it > > wants > > > to > > > > do that no matter what language? It will have to specify each > > combination > > > > possible because we can't use generic field names). > > > > > > > > Solr's approach works for using it in your web application's search > > > input, > > > > in a specific usecase, where you have precisely specified the default > > > > search fields and their boosts inside your schema.xml. However, as a > > > search > > > > API, using option 2) you are making the life of anyone else wanting > to > > > use > > > > the Solr search API really hard. Also, your search application will > > work > > > > nicely when the user enters a simple query in the input field, but an > > > > advanced user will suffer the same fate when trying to write an > > advanced > > > > query, thus not relying on the default query (computed by solr based > on > > > > schema.xml). > > > > > > > > Also, based on your note above regarding improvements like title_ws > and > > > > such, again, all of these are very well suited for the search > > application > > > > use case, together with the default query that you configure in > > > schema.xml, > > > > making the search results perform really well. However, what does all > > > these > > > > fields mean to another extension wanting to do search? Will it have > to > > > > handle all these implementation details to query for title, content > and > > > > such? I`m not sure how well this would work in practice. > > > > > > > > Unrealistic idea(?): perhaps we should come up with an abstract > search > > > > language (Solr/Lucene clone) that parses the searched fields andhides > > the > > > > complexities of all the indexed fields, allowing to write simple > > queries > > > > like "title:XWiki", while this gets translated to "title_en:XWiki OR > > > > title_fr:XWiki OR title_de:XWiki..." :) > > > > > > > > Am I approaching this wrong by trying to have both a > tweakable/tweaked > > > > search application AND a search API? Are the two not compatible? Do > we > > > have > > > > to sacrifice search result performance (no language-specific stuff) > to > > be > > > > able to have a usable API? > > > > > > > > > > > >> > If you want just some language, you have to read the right fields > > > >> > (ex title_en) instead of just getting a clear field name (title). > > > >> > > > >> You have to be careful, this is really only if you want to be > > specific. > > > >> In this case, it is likely that you also do not want so much > stemming. > > > >> My experience, which was before dismax on curriki.org, has made it > so > > > >> that any query that is a bit specific is likely to not desire > > stemming. > > > >> > > > > > > > > Can you please elaborate on this? I`m not sure I understand the > > problem. > > > > > > > > > > > >> > > > >> > -- Also, the schema.xml definition is a static one in this > concern, > > > >> > requiring you to know beforehand which languages you want to > support > > > >> (for > > > >> > example when defining the default fields to search for). Adding a > > new > > > >> > language requires you to start editing the xml files by hand. > > > >> > > > >> True but the available languages are almost all hand-coded. > > > >> You could generate the schema.xml based on the available languages > if > > > not > > > >> hand-generated? > > > >> > > > > > > > > Basically I would have to output a zip with schema.xml, > solrconfig.xml > > > and > > > > then all the resources specific to all the selected languages > > (stopwords, > > > > synonims, etc) for the languages that we can provide out of the box. > > For > > > > other languages, the admin would have to get dirty with the xmls. > > > > > > > > > > > >> > > > >> There's one catch with this approach which is new to me but seems to > > be > > > >> quite important to implement this approach: the idf should be > > modified, > > > the > > > >> Similarity class should be, so that the total number of documents is > > the > > > >> total number of documents having that language. > > > >> See: > > > >> > > > >> > > > > > > http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%[email protected]%3E > > > >> The solution sketched there sounds easy but I have not tried it. > > > >> > > > >> > 3) Indexing the content in different Solr cores (indexes), one for > > > each > > > >> > language. Each core requires it's on directory and configuration > > > files. > > > >> > - The advantage is that queries are clean to write (like option 1) > > and > > > >> that > > > >> > you have a nice separation > > > >> > - The disadvantage is that it's difficult to get it right > > > >> (administrative > > > >> > issues) and then you also have the (considerable) problem of > having > > to > > > >> fix > > > >> > the relevancy score of a query result that has entries from > > different > > > >> > cores; each core has it's own relevancy computed and does not > > consider > > > >> the > > > >> > others. > > > >> > - To make it even worst, it seems that you can not [5] also push > to > > a > > > >> > remote Solr instance the configuration files when creating a new > > core > > > >> > programatically. However, if we are running an embedded Solr > > instance, > > > >> we > > > >> > could provide a way to generate the config files and write them to > > the > > > >> data > > > >> > directory. > > > >> > > > >> Post-processing results is very very very dangerous as performance > is > > at > > > >> risk (e.g. if a core does not answer)... I would tend to avoid that > as > > > much > > > >> as possible. > > > >> > > > > > > > > Not really related, but this reminds me about the post processing > that > > I > > > > do for checking view rights over the returned result, but that's > > another > > > > discussion that we will probably need to have :) > > > > > > > > > > > >> > > > >> > Currently I have implemented option 1) in our existing Solr > > > integration, > > > >> > which is also more or less compatible with our existing Lucene > > > queries, > > > >> but > > > >> > I would like to find a better solution that actually analyses the > > > >> content. > > > >> > > > > >> > During GSoC, option 2) was preferred but the implementation did > not > > > >> > consider practical reasons like the ones described above (query > > > >> complexity, > > > >> > user configuration, etc.) > > > >> > > > >> True, Savitha surfed the possibility of having different solr > > documents > > > >> per language. > > > >> I still could not be sure that this was not showing the document > match > > > >> single in one language. > > > >> > > > >> However, indicating which language it is matched into is probably > > > >> useful... > > > >> > > > > > > > > Already doing that. > > > > > > > > > > > >> Funnily, cross-language-retrieval is a mature research field but > > > >> retrieval for multilanguage user is not so! > > > >> > > > >> > On a related note, I have also watched an interesting presentation > > [3] > > > >> > about how Drupal handles its Solr integration and, particularly, a > > > >> plugin > > > >> > [4] that handles the multilingual aspect. > > > >> > The idea seen there is that you have this UI that helps you > generate > > > >> > configuration files, depending you your needs. For instance, you > > > (admin) > > > >> > check that you need search for language English, French and German > > and > > > >> the > > > >> > ui/extension gives you a zip with the configuration you need to > use > > in > > > >> your > > > >> > (remote or embedded) solr instance. The configuration for each > > > language > > > >> > comes preset with the analyzers you should use for it and the > > > additional > > > >> > resources (stopwords.txt, synonims.txt, etc.). > > > >> > This approach helps with avoiding the need for admins to be forced > > to > > > >> edit > > > >> > xml files and could also still be useful for other cases, not only > > > >> option > > > >> > 2). > > > >> > > > >> Generating sounds like an easy approach to me. > > > >> > > > > > > > > Yes, however I don`t like the fact that we can not do everything from > > the > > > > webapp and the admin needs to access the filesystem to install the > > given > > > > configuration on the embedded/remote solr directory. Lucene does not > > have > > > > this problem now. It just works with XWiki and everything is done > from > > > > XWiki UI. I feel that losing this commodity will not be very well > > > received > > > > by users that now have some new install steps to get XWiki running. > > > > > > > > Well, of course, for the embedded solr version, we could handle it > like > > > we > > > > do now and push the files directly from the webapp to the filesystem. > > > Since > > > > embedded will be default, it should be OK and avoid the extra install > > > step. > > > > Users with a remote solr machine should have the option to get the > zip > > > > instead. > > > > > > > > Not sure if we can apply the new configuration without a restart, but > > > I`ll > > > > have to look more into it. I know the multi-core architecture > supports > > > > something like this but will have to see the details. > > > > > > > > > > > >> > > > >> > All these problems basically come from the fact that there is no > way > > > to > > > >> > specify in the schema.xml that, based on the value of a field > (like > > > the > > > >> > field "lang" that stores the document language), you want to run > > this > > > or > > > >> > that group of analyzers. > > > >> > > > >> Well, this is possible with ThreadLocal but is not necessarily a > good > > > >> idea. > > > >> Also, it is very common that users formulate queries without > > formulating > > > >> their language and thus you need to "or" the user's queries through > > > >> multiple languages (e.g. given by the browser). > > > >> > > > >> > Perhaps a solution would be a custom kind of "AggregatorAnalyzer" > > that > > > >> > would call other analyzers at runtime, based on the value of the > > lang > > > >> > field. However, this solution could only be applied at index time, > > > when > > > >> you > > > >> > have the lang information (in the solrDocument to be indexed), but > > > when > > > >> you > > > >> > perform the query, you can not analyze the query text since you do > > not > > > >> know > > > >> > the language of the field you're querying (it was determined at > > > runtime > > > >> - > > > >> > at index time) and thus do not know what operations to apply to > the > > > >> query > > > >> > (to reduce it to the same form as the indexed values). > > > >> > > > >> How would that look at query time? > > > >> > > > > > > > > That's what I was saying, that at query time, the searched term will > > not > > > > get analyzed by the right chain. When you search for a single > language, > > > you > > > > could add that language as a query filter and then you could apply > the > > > > right chain, but when searching in 2 or more (or no, meaning all) > > > languages > > > > you are stuck. > > > > > > > >> > > > >> > I have also read another interesting analysis [6] on this problem > > that > > > >> > elaborates on the complexities and limitations of each options. > > > (Ignore > > > >> the > > > >> > Rosette stuff mentioned there) > > > >> > > > > >> > I have been thinking about this for some time now, but the > solution > > is > > > >> > probably somewhere in between, finding an option that is > acceptable > > > >> while > > > >> > not restrictive. I will probably also send a mail on the Solr list > > to > > > >> get > > > >> > some more input from there, but I get the feeling that whatever > > > >> solution we > > > >> > choose, it will most likely require the users to at least copy (or > > > even > > > >> > edit) some files into some directories (configurations and/or > jars), > > > >> since > > > >> > it does not seem to be easy/possible to do everything on-the-fly, > > > >> > programatically. > > > >> > > > >> The only hard step is when changing the supported languages, I > think. > > > >> In this case, when automatically generating the index, you need to > > warn > > > >> the user. > > > >> The admin UI should have a checkbox "use generated schema" or a > > textarea > > > >> for the schema. > > > >> > > > > > > > > Please see above regarding configuration generation. Basically, since > > we > > > > are going to support both embedded and remote solr instances, we > could > > > > support things like editing the schema from XWiki only for the > embedded > > > > instance, but not for the remote one. We might end up having separate > > UIs > > > > for each case, since we might want to exploit the flexibility of the > > > > embedded one as much as possible. > > > > > > > > > > > >> > > > >> Those that want particular fields and tunings need to write their > own > > > >> schema. > > > >> > > > >> The same UI could also include whether to include a phonetic track > or > > > not > > > >> (then require reindexing). > > > > > > > > > > > >> hope it helps. > > > >> > > > > > > > > Yes, very helpful so far. I`m counting on your expertise with > > Lucene/Solr > > > > on the details. My current approach is a practical one without > previous > > > > experience on the topic, so I`m still doing mostly guesswork in some > > > areas. > > > > > > > > Thanks, > > > > Eduard > > > > > > > > > > > >> paul > > > >> _______________________________________________ > > > >> devs mailing list > > > >> [email protected] > > > >> http://lists.xwiki.org/mailman/listinfo/devs > > > >> > > > > > > > > > > > _______________________________________________ > > > devs mailing list > > > [email protected] > > > http://lists.xwiki.org/mailman/listinfo/devs > > > > > > > > > > > -- > > Ludovic Dubost > > Founder and CEO > > Blog: http://blog.ludovic.org/ > > XWiki: http://www.xwiki.com > > Skype: ldubost GTalk: ldubost > > _______________________________________________ > > devs mailing list > > [email protected] > > http://lists.xwiki.org/mailman/listinfo/devs > > > _______________________________________________ > devs mailing list > [email protected] > http://lists.xwiki.org/mailman/listinfo/devs > -- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost
Hi, On Wed, Nov 28, 2012 at 5:35 PM, Ludovic Dubost <[email protected]> wrote:
2012/11/28 Eduard Moraru <[email protected]>
Hi Ludovic,
Thanks for the reply. Please read below...
On Tue, Nov 27, 2012 at 5:44 PM, Ludovic Dubost <[email protected]> wrote:
Hi Edy,
I'm not a huge fan of the title_fr title_en title_morelanguages approach as indeed it seems to be quite complex at the query level. I was more leaning towards multiple indexes if we can query them globally but I understand this is complex too.
Now let's see the use cases that are hugely important:
1/ Make sure that if you decide your wiki is monolingual: - the indexing uses the specific language analyzer - make sure the query uses the specific language analyzer - make sure the search looks in all content even if the language setting of the document is wrongly set (consider all documents being of the specific language)
You mean that, if the wiki is monolingual, we should ignore the language filter and hardcode it to "All languages", right?
However, what would be the advantage of this? Why would be want to pollute the results with irrelevant documents (caused by a probable recent configuration change that went from multi-lingual to mono-lignual)? Wasn't that the whole reason why the admin switched to mono-lingual?
If the user is monolingual, we can safely ignore the language setting of each document and only the "main" document will be shown anyway to the user. So we should make sure we search on ALL documents that are available to the user.
The user might have to "reindex" to make sure this is properly taken into account by the search engine.
What is important here, is that even if the wiki is set to "fr", then even if documents have "en" for the main language they will still show up in the search. The opposite would be bad.
Why/How can you end up with documents of other languages than the current default one if the wiki is mono-lingual?
2/ Allow is a wiki is multi-lingual: - search in the language you decide (maybe the UI should display a language choice for the query)
We already support this, by using the "Filtered Search" option and selecting the language.
Sure. I was just repeating what is important.
- search in content that is analyzed in the proper language when the
content is declared in this language - allow to specify if you want to restrict your search to documents declared in the language of your query, versus search more widely in
all
documents accross languages. If you search in only the language of the query only one document can show up but it should point to the right translation that matches, if you search in multiple languages then you can show individual translations.
I think this one is the same as the first bullet above.
No I think it's different. The user can decide to make his search in "French" but look in the "French" + "English" dataset. For me if both documents with the same page name match, both should come out separately.
It's the same as in the UI allows you to specify which language to search in, but it also allows you to specify that you want to search in "All languages", that is what I meant.
- allow technical users to search for all documents across all languages (where the language analysis does not really matter)
Do you mean as an API?
Not specifically as an API
What exactly do you mean by "language analysis does not really matter"?
Any
Example?
I mean here that as a technical user your objective is to make sure your search spans ALL content in the wiki. In this case you don't care about stemming and such. The standard UI could take this into account by choosing "any language search" and with data set "all languages". We just need to make sure that this won't exclude any content of the search
(Here is an example of a case where such exclusion might occur. Suppose you have done only "French" and "English" indexing and there is "German" content also in your wiki but since you have not asked for German search in your wiki you don't have title_de and content_de fields or don't have a german specific solr index (in the other method), then your german content would not be indexed at all ?)
I understand now what you mean. Well, to avoid this, we can use only one notion and that is the one of "supported languages". All the supported languages should be indexed in order for the Search feature to work as expected (and not have the case where you have a document that is not indexed because of its language). I am not sure if there are many examples where admins would choose to index just some of the supported languages and not index all of them. WDYT?
From an admin point of view it makes good sense for the admin to be
able
to
specify in a multilingual wiki which language analysis should be activated, and then have this transmitted to SOLR to properly configure the engine. Reindexing is ok when changing the configuration.
I believe in the end wether you use multiple fields with _fr _en or multiple SOLR cores, as long as you can query accross SOLR cores is a bit the same. If you cannot run a query merging multiple indexes then the first solution is kind of absolutely necessary as it would be the only one allowing to search across all languages.
Maybe a solution would be to create one index per language and index ALL content regardless of it's language using the language analyzer of that index. This would allow to have better results even though the users have badly tagged the language of a document, and it's only the job of the UI to limit the search to only the language of the query, or all documents.
So you could have a configuration in the admin that says:
1/ Create an English Index 2/ Create an additional French index
The UI would allow to search in English and French, + would add a language restriction for the documents.
Applying the language specific analyzers (for Chinese, for example) to all the documents will just create a mess for all the documents that do not match the analyzer's language. I`m not sure the results for the badly-indexed languages will make any sense to users.
I undestand the issue here, but in most cases the user will say "french search" on "french content", he will only expand to non french content if he was not satisfied by his search. What is just allowed here is to also search in "french" inside all content. That would cover content that would have the bad language setting as well as any other content. The results might be a bit noisy but I don't think it's a big issue.
What I think you are actually asking is that all languages are queried, but that results in the requested language get boosted/elevated and thus are displayed before all other language results. Perhaps this can be achieved by using the lang field as a query field and not as a query filter. This means that it will influence the score. Couple that with a higher boost for the lang field and results in the requested language should be scored better than all the rest (because there will always be a hit for the lang field resulting in a better score). There might still be some results from other languages that could have a better score than results from the requested language (because they just have more hits on diferent fields), but it should work in most cases. However, I`m not sure we want this. Users might results from other languages as a bug. Problems might also be harder to debug. Instead, we could do something like Google and if we see a small number of results, we could suggest that the user searches in all languages. In any case, I`m not sure this is such a big issue right now, since it's basically an optimisation that we could choose to do, or let the user do it himself, since he has the tools (UI).
Also, this is very similar to the multi-core approach (one core per language), just that you also add documents that are indexed with the wrong analyzers. We have the same problem regarding merging relevance scores across indexes (cores) that is a big turn-off for the original multi-core approach.
This is a more serious issue. If it's hard to merge the results spanning over multiple cores this could be a showstopper. However the solution of having only one Lucene document for all languages is not so cool either as it would make it difficult to know which ones of the languages has matched and present them separately with separate scores.
I agree. The one-document-for-all-languages solution was proposed by Paul, but I don`t think it's the way to go. We are currently only considering the one-document-per-language direction.
It's really the core issue to decide on. What are the benefits and drawbacks of the different solutions. For each solution is there something in the UI that you cannot do ? So far I've heard:
1/ Presenting different scores for documents in different languages with the same doc name if the title_fr,content_fr method is used
Since we are not considering the one-document-for-all-languages solution, this is not an issue. 2/ Merging scores accross indexes in multicore approach
Other ? Can we list them in a wiki page ?
In the future if we are able to "detect" the language of the documents
we
could add a lucene field with the "detected" language instead of the "provided" language of the documents, therefore increasing the quality of searches only on documents of a specific language.
In the previous discussions (on the GSoC thread) we agreed that language in XWiki is known before-hard, so no recognition is required, at least not at document level.
Let's forget this
This later solution would be the only one that would really work on
file
attachements as we have no information about the specific language of file attachements (or even XWiki objects) which are attached to the main document and not to the translated document.
Yes, this is a problem right now. AFAIU, the plan [1] is to support translated objects and maybe attachments as well. Until then, we could either: 1) Use the original document's language to index the attachment's content
This is not a good solution. If I understand correctly we could not end up not searching in a french attachment because the original document is in marked "en". I'm for Paul's solution to index objects and attachments in each translation (if we have separate entities for translated documents).
Yep, we already do this for objects, just need to do it for attachments as well.
I understand that in the title_fr,content_fr approach this problem does not happen.
This problem is not related to how we handle multiple translations (separate indexes or not), as long as each translation is a separate entity. Basically this last part is the problem. If the French translation is a separate entity from the original document (i.e. that is in English), any object/attachment *index field* of the English original version will need to be duplicated into the French translation as well, or the French translation risks not to get a hit on the object/attachment.
2) Use a language detection library to try to detect the attachment content's language and index it accordingly.
Not sure we can for now
The above could also be applied for objects and their properties. ---------- [1] http://jira.xwiki.org/browse/XWIKI-69
This later issues shows that a search on "only french content" should
still
include the attachements because we have no idea if the attachments are "french" or "english".
(The paragraphs below discuss on what currently exists and what could be done, ignoring the possible language detection mentioned above)
Right now a document also indexes the object's properties in a field called "objcontent". I do this for all translations, thus duplicating the field's value in all translations. I can do the same for attachments. The purpose is, indeed, to be able to find document translations based on hits in their objects/attachments. If a language filter is used and there is a hit in an object, only one document is returned. If there are no language filters, all translations will be returned.
It seems we have to do this for now
However, if we search for the object/property/attachment itself, it will only be assigned to one language: the language of the original document. This means that if we search for all languages, the object itself will be found too (there is no language filter used). If we add a language filter that is different from the object/property/attachment's original document language, the object/property/attachment will not be found.
Maybe we can come up with some processing of the query in the search application, that applies the language filter only for documents:
((-type:"OBJECT" OR -type:"OBJECT_PROPERTY" OR -type:"ATTACHMENT") OR lang:"<userSelectedLanguage>") -- writing it like this because the
default
operand is AND in the query filter clause that we use in the Search application.
The problem with this is that that, when a language filter is used, the object/property/attachments that are now included in the results might not have the specified language and will pollute the results.
I'm not sure I understand. We have an "objcontent" field for each translation that has the full content of objects and properties, but individual object fields we don't have in each translations ?
"objcontent" stores the properties (format: "propName:propValue") of each object inside the original document (multi-valued field). Besides XWikiDocuments, we also index Objects, Properties and Attachments as Lucene/Solr first-class documents. Each of these entries has the wiki, space, name and lang fields set from the document to which they belong to. This is what I meant above with "if we search for the object/property/attachment itself" So to reiterate, the idea was that if you want to search for an indexed Object (type:"OBJECT"), you will *have* to avoid setting a language filter, or you might not find the object you are looking for, since it is indexed under the language of original document. --- While writing this down, I just thought of an elegant idea to fix this. The lang field could be set to multi-valued. This means that, when we index objects, properties and attachments, we could set in the lang field all the values corresponding to all the existing translations of the owning document. Example: An Object: id: xwiki:Main.Page^XWiki.Class[0] <-- (object reference) class: XWiki.Class wiki: xwiki space: Main name: Page lang: en <-- proposal to make it multi-valued would look like this, stored like a list of values. fr de type: OBJECT Note that this solution would also affect document languages as well (since they share the same schema), but we will just put one value and it will not affect queries. Queries will still be written like: "lang:en" If we apply this solution, even if a language filter exists in the user's query, it will still hit the object because the lang field of the object contains the value.
The more I see all the issues, the more I lean towards a separate index per language solution.
Again, these issues are not related to this, just to the fact that Document translations are first-class Lucene documents.
The reason I do is that the main need is for a non English user to have very relevant results in his own language. Therefore we need to make sure that all content that the users have published has the chance to be analyzed using the non English language analyzer. So indexing all objects and attachments with the relevant language analyzer is the solution. This is also why I proposed to index all content in this specific index regardless of the language declared, which would only be used in the UI to limit searches to the specific language.
In this view:
0/ There would be a language specific index per language with the objects and attachments indexed only in the language of the index 1/ The user chooses the language in which he searches 2/ Automatically that sets the index to be used to be the "french" index 3/ Automatically that presets to limit the span of the search to declared "french" documents 4/ The user can decide to go for non french documents at his own risks knowing that the results might be weird because of wrong analysis (this is what happens today with english analysis over french documents)
The benefit here is that you don't have a merging score over multiple index issue, since you would never have to do a search across multiple indexes. Searches are still simple to write. By default results are quite relevant since you limit the search the french declared documents (this would be the same as limiting your search to title_fr, content_fr) and still cover what needs to be covered (objects and attachments). Another benefit is that this falls back gracefully to monolingual as you just have to have one index in the language declared for the monolingual wiki. The drawback is that the indexing is more costly and there is duplicated content in the index. Howerver it is the Admin that say which languages he wants available and he takes responsibility of the ressources this needs.
Could this solution work ?
Unfortunately I am still not convinced by this approach. Besides the complexities of managing multiple cores (each with it's own schema and config files), the user is exposed to a lot of unnecessary badly indexed data that will make him stay away from the search feature, as it has happened with Lucene so far. I believe this discussion thread has come up with some nice solutions to most of our problems and that the multiple-fields direction is one that can give us relevant results, properly indexed content for all languages (even when searching in all languages) and a good solution for objects, attachments and properties (though, again, this is not related to this specific choice). I will try to make a summary of the ideas from this thread and put them into a wiki page that documents the design/progress of the multi-lingual related work. Thank you very much for now and, of course, this does not mean that the discussion is over in any way :) -Eduard
Ludovic
Thanks, Eduard
Ludovic
2012/11/26 Eduard Moraru <[email protected]>
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a
zip
that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance. This operation is to be performed each time an admin changes the indexed languages (rarely or even only once).
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru < [email protected]
wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht <[email protected]
wrote:
Hello Eduard,
it's nice of you to see you take this further.
> This issue has already been previously [1] discussed during the
GSoC
> project, but I am not particularly happy with the chosen approach. > When handling multiple languages, there are generally[2][3] 3 different > approaches: > > 1) Indexing the content in a single field (like title, doccontent, etc.) > - This has the advantage that queries are clear and fast > - The disadvantage is that you can not run very well tuned analyzers on the > fields, having to resort to (at best) basic tokenization and lowercasing. > > 2) Indexing the content in multiple fields, one field for each language > (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) > - This has the advantage that you can easily specify (as dynamic fields) > that *_en fields are of type text_en (and analyzed by an english-centered > chain of analyzers); *_fr of type text_fr (focused on french, etc.), thus > making the results much better.
I would add one more field here: title_ws and text_ws where the full text is analyzed just as words (using the whitespace-tokenizer?). A match there would even be preferred to a match in the below text-fields.
(maybe that would be called title and text?)
> - The disadvantage is that querying such a schema is a pain. If you want > all the results in all languages, you end up with a big and expensive > query.
Why is this an issue? Dismax does it for you for free (thanks to the "form" parameter that gives weight to each of the fields). This is an issue only if you start to have more than 100 languages or so... Lucene, the underlying engine of solr, handles thousands of clauses in a query without an issue (this is how prefix-queries are handled... they are expanded to a query for any of the term that matches the prefix, a setting deep somewhere, which is about 2000 avoids this to explode).
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
> If you want just some language, you have to read the right fields > (ex title_en) instead of just getting a clear field name (title).
You have to be careful, this is really only if you want to be specific. In this case, it is likely that you also do not want so much stemming. My experience, which was before dismax on curriki.org, has made it so that any query that is a bit specific is likely to not desire stemming.
Can you please elaborate on this? I`m not sure I understand the problem.
> -- Also, the schema.xml definition is a static one in this
concern,
> requiring you to know beforehand which languages you want to support (for > example when defining the default fields to search for). Adding a new > language requires you to start editing the xml files by hand.
True but the available languages are almost all hand-coded. You could generate the schema.xml based on the available languages if not hand-generated?
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
There's one catch with this approach which is new to me but seems
to be
quite important to implement this approach: the idf should be modified, the Similarity class should be, so that the total number of documents is the total number of documents having that language. See:
The solution sketched there sounds easy but I have not tried it.
> 3) Indexing the content in different Solr cores (indexes), one for each > language. Each core requires it's on directory and configuration files. > - The advantage is that queries are clean to write (like option
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar... 1)
and
that > you have a nice separation > - The disadvantage is that it's difficult to get it right (administrative > issues) and then you also have the (considerable) problem of having to fix > the relevancy score of a query result that has entries from different > cores; each core has it's own relevancy computed and does not consider the > others. > - To make it even worst, it seems that you can not [5] also push to a > remote Solr instance the configuration files when creating a new core > programatically. However, if we are running an embedded Solr instance, we > could provide a way to generate the config files and write them to the data > directory.
Post-processing results is very very very dangerous as performance is at risk (e.g. if a core does not answer)... I would tend to avoid that as much as possible.
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
> Currently I have implemented option 1) in our existing Solr
integration,
> which is also more or less compatible with our existing Lucene queries, but > I would like to find a better solution that actually analyses the content. > > During GSoC, option 2) was preferred but the implementation did not > consider practical reasons like the ones described above (query complexity, > user configuration, etc.)
True, Savitha surfed the possibility of having different solr documents per language. I still could not be sure that this was not showing the document match single in one language.
However, indicating which language it is matched into is probably useful...
Already doing that.
Funnily, cross-language-retrieval is a mature research field but retrieval for multilanguage user is not so!
> On a related note, I have also watched an interesting presentation [3] > about how Drupal handles its Solr integration and, particularly, a plugin > [4] that handles the multilingual aspect. > The idea seen there is that you have this UI that helps you generate > configuration files, depending you your needs. For instance, you (admin) > check that you need search for language English, French and German and the > ui/extension gives you a zip with the configuration you need to use in your > (remote or embedded) solr instance. The configuration for each language > comes preset with the analyzers you should use for it and the additional > resources (stopwords.txt, synonims.txt, etc.). > This approach helps with avoiding the need for admins to be forced to edit > xml files and could also still be useful for other cases, not only option > 2).
Generating sounds like an easy approach to me.
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
> All these problems basically come from the fact that there is no
way to
> specify in the schema.xml that, based on the value of a field (like the > field "lang" that stores the document language), you want to run this or > that group of analyzers.
Well, this is possible with ThreadLocal but is not necessarily a good idea. Also, it is very common that users formulate queries without formulating their language and thus you need to "or" the user's queries through multiple languages (e.g. given by the browser).
> Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that > would call other analyzers at runtime, based on the value of the lang > field. However, this solution could only be applied at index time, when you > have the lang information (in the solrDocument to be indexed), but when you > perform the query, you can not analyze the query text since you do not know > the language of the field you're querying (it was determined at runtime - > at index time) and thus do not know what operations to apply to the query > (to reduce it to the same form as the indexed values).
How would that look at query time?
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
> I have also read another interesting analysis [6] on this
problem that
> elaborates on the complexities and limitations of each options. (Ignore the > Rosette stuff mentioned there) > > I have been thinking about this for some time now, but the solution is > probably somewhere in between, finding an option that is acceptable while > not restrictive. I will probably also send a mail on the Solr list to get > some more input from there, but I get the feeling that whatever solution we > choose, it will most likely require the users to at least copy (or even > edit) some files into some directories (configurations and/or jars), since > it does not seem to be easy/possible to do everything on-the-fly, > programatically.
The only hard step is when changing the supported languages, I think. In this case, when automatically generating the index, you need to warn the user. The admin UI should have a checkbox "use generated schema" or a textarea for the schema.
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
Those that want particular fields and tunings need to write their
own
schema.
The same UI could also include whether to include a phonetic track or not (then require reindexing).
hope it helps.
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
paul _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
On Thu, Nov 29, 2012 at 3:47 PM, Eduard Moraru <[email protected]> wrote:
Hi,
On Wed, Nov 28, 2012 at 5:35 PM, Ludovic Dubost <[email protected]> wrote:
2012/11/28 Eduard Moraru <[email protected]>
Hi Ludovic,
Thanks for the reply. Please read below...
On Tue, Nov 27, 2012 at 5:44 PM, Ludovic Dubost <[email protected]> wrote:
Hi Edy,
I'm not a huge fan of the title_fr title_en title_morelanguages approach as indeed it seems to be quite complex at the query level. I was more leaning towards multiple indexes if we can query them globally but I understand this is complex too.
Now let's see the use cases that are hugely important:
1/ Make sure that if you decide your wiki is monolingual: - the indexing uses the specific language analyzer - make sure the query uses the specific language analyzer - make sure the search looks in all content even if the language setting of the document is wrongly set (consider all documents being of the specific language)
You mean that, if the wiki is monolingual, we should ignore the language filter and hardcode it to "All languages", right?
However, what would be the advantage of this? Why would be want to pollute the results with irrelevant documents (caused by a probable recent configuration change that went from multi-lingual to mono-lignual)? Wasn't that the whole reason why the admin switched to mono-lingual?
If the user is monolingual, we can safely ignore the language setting of each document and only the "main" document will be shown anyway to the user. So we should make sure we search on ALL documents that are available to the user.
The user might have to "reindex" to make sure this is properly taken into account by the search engine.
What is important here, is that even if the wiki is set to "fr", then even if documents have "en" for the main language they will still show up in the search. The opposite would be bad.
Why/How can you end up with documents of other languages than the current default one if the wiki is mono-lingual?
When importing a XAR for example.
2/ Allow is a wiki is multi-lingual: - search in the language you decide (maybe the UI should display a language choice for the query)
We already support this, by using the "Filtered Search" option and selecting the language.
Sure. I was just repeating what is important.
- search in content that is analyzed in the proper language when the
content is declared in this language - allow to specify if you want to restrict your search to documents declared in the language of your query, versus search more widely in
all
documents accross languages. If you search in only the language of
the
query only one document can show up but it should point to the right translation that matches, if you search in multiple languages then you can show individual translations.
I think this one is the same as the first bullet above.
No I think it's different. The user can decide to make his search in "French" but look in the "French" + "English" dataset. For me if both documents with the same page name match, both should come out separately.
It's the same as in the UI allows you to specify which language to search in, but it also allows you to specify that you want to search in "All languages", that is what I meant.
- allow technical users to search for all documents across all languages (where the language analysis does not really matter)
Do you mean as an API?
Not specifically as an API
What exactly do you mean by "language analysis does not really matter"?
Any
Example?
I mean here that as a technical user your objective is to make sure your search spans ALL content in the wiki. In this case you don't care about stemming and such. The standard UI could take this into account by choosing "any language search" and with data set "all languages". We just need to make sure that this won't exclude any content of the search
(Here is an example of a case where such exclusion might occur. Suppose
you
have done only "French" and "English" indexing and there is "German" content also in your wiki but since you have not asked for German search in your wiki you don't have title_de and content_de fields or don't have a german specific solr index (in the other method), then your german content would not be indexed at all ?)
I understand now what you mean. Well, to avoid this, we can use only one notion and that is the one of "supported languages". All the supported languages should be indexed in order for the Search feature to work as expected (and not have the case where you have a document that is not indexed because of its language). I am not sure if there are many examples where admins would choose to index just some of the supported languages and not index all of them.
WDYT?
From an admin point of view it makes good sense for the admin to be
able
to
specify in a multilingual wiki which language analysis should be activated, and then have this transmitted to SOLR to properly configure the engine. Reindexing is ok when changing the configuration.
I believe in the end wether you use multiple fields with _fr _en or multiple SOLR cores, as long as you can query accross SOLR cores is a bit the same. If you cannot run a query merging multiple indexes then the first solution is kind of absolutely necessary as it would be the only one allowing to search across all languages.
Maybe a solution would be to create one index per language and index ALL content regardless of it's language using the language analyzer of
that
index. This would allow to have better results even though the users have badly tagged the language of a document, and it's only the job of the UI to limit the search to only the language of the query, or all documents.
So you could have a configuration in the admin that says:
1/ Create an English Index 2/ Create an additional French index
The UI would allow to search in English and French, + would add a language restriction for the documents.
Applying the language specific analyzers (for Chinese, for example) to all the documents will just create a mess for all the documents that do not match the analyzer's language. I`m not sure the results for the badly-indexed languages will make any sense to users.
I undestand the issue here, but in most cases the user will say "french search" on "french content", he will only expand to non french content if he was not satisfied by his search. What is just allowed here is to also search in "french" inside all content. That would cover content that would have the bad language setting as well as any other content. The results might be a bit noisy but I don't think it's a big issue.
What I think you are actually asking is that all languages are queried, but that results in the requested language get boosted/elevated and thus are displayed before all other language results.
Perhaps this can be achieved by using the lang field as a query field and not as a query filter. This means that it will influence the score. Couple that with a higher boost for the lang field and results in the requested language should be scored better than all the rest (because there will always be a hit for the lang field resulting in a better score). There might still be some results from other languages that could have a better score than results from the requested language (because they just have more hits on diferent fields), but it should work in most cases.
However, I`m not sure we want this. Users might results from other languages as a bug. Problems might also be harder to debug.
Instead, we could do something like Google and if we see a small number of results, we could suggest that the user searches in all languages.
In any case, I`m not sure this is such a big issue right now, since it's basically an optimisation that we could choose to do, or let the user do it himself, since he has the tools (UI).
Also, this is very similar to the multi-core approach (one core per language), just that you also add documents that are indexed with the wrong analyzers. We have the same problem regarding merging relevance scores across indexes (cores) that is a big turn-off for the original
multi-core
approach.
This is a more serious issue. If it's hard to merge the results spanning over multiple cores this could be a showstopper. However the solution of having only one Lucene document for all languages is not so cool either as it would make it difficult to know which ones of the languages has matched and present them separately with separate scores.
I agree. The one-document-for-all-languages solution was proposed by Paul, but I don`t think it's the way to go. We are currently only considering the one-document-per-language direction.
It's really the core issue to decide on. What are the benefits and drawbacks of the different solutions. For each solution is there something in the UI that you cannot do ? So far I've heard:
1/ Presenting different scores for documents in different languages with the same doc name if the title_fr,content_fr method is used
Since we are not considering the one-document-for-all-languages solution, this is not an issue.
2/ Merging scores accross indexes in multicore approach
Other ? Can we list them in a wiki page ?
In the future if we are able to "detect" the language of the
documents
we
could add a lucene field with the "detected" language instead of the "provided" language of the documents, therefore increasing the quality of searches only on documents of a specific language.
In the previous discussions (on the GSoC thread) we agreed that language in XWiki is known before-hard, so no recognition is required, at least not at document level.
Let's forget this
This later solution would be the only one that would really work on
file
attachements as we have no information about the specific language of file attachements (or even XWiki objects) which are attached to the main document and not to the translated document.
Yes, this is a problem right now. AFAIU, the plan [1] is to support translated objects and maybe attachments as well. Until then, we could either: 1) Use the original document's language to index the attachment's content
This is not a good solution. If I understand correctly we could not end up not searching in a french attachment because the original document is in marked "en". I'm for Paul's solution to index objects and attachments in each translation (if we have separate entities for translated documents).
Yep, we already do this for objects, just need to do it for attachments as well.
I understand that in the title_fr,content_fr approach this problem does not happen.
This problem is not related to how we handle multiple translations (separate indexes or not), as long as each translation is a separate entity. Basically this last part is the problem. If the French translation is a separate entity from the original document (i.e. that is in English), any object/attachment *index field* of the English original version will need to be duplicated into the French translation as well, or the French translation risks not to get a hit on the object/attachment.
2) Use a language detection library to try to detect the attachment content's language and index it accordingly.
Not sure we can for now
The above could also be applied for objects and their properties. ---------- [1] http://jira.xwiki.org/browse/XWIKI-69
This later issues shows that a search on "only french content" should
still
include the attachements because we have no idea if the attachments
are
"french" or "english".
(The paragraphs below discuss on what currently exists and what could be done, ignoring the possible language detection mentioned above)
Right now a document also indexes the object's properties in a field called "objcontent". I do this for all translations, thus duplicating the field's value in all translations. I can do the same for attachments. The purpose is, indeed, to be able to find document translations based on hits in their objects/attachments. If a language filter is used and there is a hit in an object, only one document is returned. If there are no language filters, all translations will be returned.
It seems we have to do this for now
However, if we search for the object/property/attachment itself, it
will
only be assigned to one language: the language of the original document. This means that if we search for all languages, the object itself will be found too (there is no language filter used). If we add a language filter that is different from the object/property/attachment's original document language, the object/property/attachment will not be found.
Maybe we can come up with some processing of the query in the search application, that applies the language filter only for documents:
((-type:"OBJECT" OR -type:"OBJECT_PROPERTY" OR -type:"ATTACHMENT") OR lang:"<userSelectedLanguage>") -- writing it like this because the default operand is AND in the query filter clause that we use in the Search application.
The problem with this is that that, when a language filter is used, the object/property/attachments that are now included in the results might not have the specified language and will pollute the results.
I'm not sure I understand. We have an "objcontent" field for each translation that has the full content of objects and properties, but individual object fields we don't have in each translations ?
"objcontent" stores the properties (format: "propName:propValue") of each object inside the original document (multi-valued field).
Besides XWikiDocuments, we also index Objects, Properties and Attachments as Lucene/Solr first-class documents. Each of these entries has the wiki, space, name and lang fields set from the document to which they belong to. This is what I meant above with "if we search for the object/property/attachment itself"
So to reiterate, the idea was that if you want to search for an indexed Object (type:"OBJECT"), you will *have* to avoid setting a language filter, or you might not find the object you are looking for, since it is indexed under the language of original document.
---
While writing this down, I just thought of an elegant idea to fix this. The lang field could be set to multi-valued. This means that, when we index objects, properties and attachments, we could set in the lang field all the values corresponding to all the existing translations of the owning document. Example:
An Object: id: xwiki:Main.Page^XWiki.Class[0] <-- (object reference) class: XWiki.Class wiki: xwiki space: Main name: Page lang: en <-- proposal to make it multi-valued would look like this, stored like a list of values. fr de type: OBJECT
Note that this solution would also affect document languages as well (since they share the same schema), but we will just put one value and it will not affect queries. Queries will still be written like: "lang:en"
If we apply this solution, even if a language filter exists in the user's query, it will still hit the object because the lang field of the object contains the value.
The more I see all the issues, the more I lean towards a separate index per language solution.
Again, these issues are not related to this, just to the fact that Document translations are first-class Lucene documents.
The reason I do is that the main need is for a non English user to have very relevant results in his own language. Therefore we need to make sure that all content that the users have published has the chance to be analyzed using the non English language analyzer. So indexing all objects and attachments with the relevant language analyzer is the solution. This is also why I proposed to index all content in this specific index regardless of the language declared, which would only be used in the UI to limit searches to the specific language.
In this view:
0/ There would be a language specific index per language with the objects and attachments indexed only in the language of the index 1/ The user chooses the language in which he searches 2/ Automatically that sets the index to be used to be the "french" index 3/ Automatically that presets to limit the span of the search to declared "french" documents 4/ The user can decide to go for non french documents at his own risks knowing that the results might be weird because of wrong analysis (this is what happens today with english analysis over french documents)
The benefit here is that you don't have a merging score over multiple index issue, since you would never have to do a search across multiple indexes. Searches are still simple to write. By default results are quite relevant since you limit the search the french declared documents (this would be the same as limiting your search to title_fr, content_fr) and still cover what needs to be covered (objects and attachments). Another benefit is that this falls back gracefully to monolingual as you just have to have one index in the language declared for the monolingual wiki. The drawback is that the indexing is more costly and there is duplicated content in the index. Howerver it is the Admin that say which languages he wants available and he takes responsibility of the ressources this needs.
Could this solution work ?
Unfortunately I am still not convinced by this approach. Besides the complexities of managing multiple cores (each with it's own schema and config files), the user is exposed to a lot of unnecessary badly indexed data that will make him stay away from the search feature, as it has happened with Lucene so far.
I believe this discussion thread has come up with some nice solutions to most of our problems and that the multiple-fields direction is one that can give us relevant results, properly indexed content for all languages (even when searching in all languages) and a good solution for objects, attachments and properties (though, again, this is not related to this specific choice).
I will try to make a summary of the ideas from this thread and put them into a wiki page that documents the design/progress of the multi-lingual related work.
Thank you very much for now and, of course, this does not mean that the discussion is over in any way :) -Eduard
Ludovic
Thanks, Eduard
Ludovic
2012/11/26 Eduard Moraru <[email protected]>
Hi devs,
Any other input on this matter?
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
The Solr configuration is generated by some XWiki UI that returns a
zip
that the admin has to unpack in his (remote) Solr instance. This could be automated for the embedded instance. This operation is to be performed each time an admin changes the indexed languages (rarely or even only once).
Querying such a schema is a bit tricky when you are interested in more than one language, because you have to add all the clauses (title_en, title_fr, etc.) specific to the languages you are interested in.
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy.
One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will
still
work,
but will not catch some of the nuances of specific languages. As far as I`ve seen until now, applications are not very interested in nuances, but rather in filtering the results, a task for which this solution might be well suited. Of course, nothing stops applications from using the *new* and more expressive fields that are properly analized.
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
WDYT?
Thanks, Eduard
On Wed, Nov 21, 2012 at 10:49 PM, Eduard Moraru < [email protected]
wrote:
Hi Paul,
I was counting on your feedback :)
On Wed, Nov 21, 2012 at 3:04 PM, Paul Libbrecht < [email protected]
wrote:
> > Hello Eduard, > > it's nice of you to see you take this further. > > > This issue has already been previously [1] discussed during
the GSoC
> > project, but I am not particularly happy with the chosen approach. > > When handling multiple languages, there are generally[2][3] 3 different > > approaches: > > > > 1) Indexing the content in a single field (like title, doccontent, etc.) > > - This has the advantage that queries are clear and fast > > - The disadvantage is that you can not run very well tuned analyzers on > the > > fields, having to resort to (at best) basic tokenization and > lowercasing. > > > > 2) Indexing the content in multiple fields, one field for each language > > (like title_en, title_fr, doccontent_en, doccontent_fr, etc.) > > - This has the advantage that you can easily specify (as dynamic fields) > > that *_en fields are of type text_en (and analyzed by an > english-centered > > chain of analyzers); *_fr of type text_fr (focused on french, etc.), > thus > > making the results much better. > > I would add one more field here: title_ws and text_ws where the full text > is analyzed just as words (using the whitespace-tokenizer?). > A match there would even be preferred to a match in the below text-fields. > > (maybe that would be called title and text?) > > > - The disadvantage is that querying such a schema is a pain. If you want > > all the results in all languages, you end up with a big and expensive > > query. > > Why is this an issue? > Dismax does it for you for free (thanks to the "form" parameter that > gives weight to each of the fields). > This is an issue only if you start to have more than 100 languages or > so... > Lucene, the underlying engine of solr, handles thousands of clauses in a > query without an issue (this is how prefix-queries are handled... they are > expanded to a query for any of the term that matches the prefix, a setting > deep somewhere, which is about 2000 avoids this to explode). >
Sure, Solr is great when you want to do simple queries like "XWiki Open Source", however, since in XWiki we also expose the Solr/Lucene query APIs to the platform, there will be (as as it is currently with Lucene) a lot of extensions wanting to do search using this API. These extensions (like the search suggest for example, rest search, etc) want to do something like "title:'Open Source' AND type:document AND doccontent:XWiki". Because option 2) is so messy in it's fields, it would mean that the extension would have to come up with a query like "title_en:'Open Source' AND type:document AND doccontent_en:XWiki" (assuming that it is only limited to the current -- english or whatever -- language; what happens if it wants to do that no matter what language? It will have to specify each combination possible because we can't use generic field names).
Solr's approach works for using it in your web application's search input, in a specific usecase, where you have precisely specified the default search fields and their boosts inside your schema.xml. However, as a search API, using option 2) you are making the life of anyone else wanting to use the Solr search API really hard. Also, your search application will work nicely when the user enters a simple query in the input field, but an advanced user will suffer the same fate when trying to write an advanced query, thus not relying on the default query (computed by solr based on schema.xml).
Also, based on your note above regarding improvements like title_ws and such, again, all of these are very well suited for the search application use case, together with the default query that you configure in schema.xml, making the search results perform really well. However, what does all these fields mean to another extension wanting to do search? Will it have to handle all these implementation details to query for title, content and such? I`m not sure how well this would work in practice.
Unrealistic idea(?): perhaps we should come up with an abstract search language (Solr/Lucene clone) that parses the searched fields andhides the complexities of all the indexed fields, allowing to write simple queries like "title:XWiki", while this gets translated to "title_en:XWiki OR title_fr:XWiki OR title_de:XWiki..." :)
Am I approaching this wrong by trying to have both a tweakable/tweaked search application AND a search API? Are the two not compatible? Do we have to sacrifice search result performance (no language-specific stuff) to be able to have a usable API?
> > If you want just some language, you have to read the right fields > > (ex title_en) instead of just getting a clear field name (title). > > You have to be careful, this is really only if you want to be specific. > In this case, it is likely that you also do not want so much stemming. > My experience, which was before dismax on curriki.org, has made it so > that any query that is a bit specific is likely to not desire stemming. >
Can you please elaborate on this? I`m not sure I understand the problem.
> > > -- Also, the schema.xml definition is a static one in this concern, > > requiring you to know beforehand which languages you want to support > (for > > example when defining the default fields to search for). Adding a new > > language requires you to start editing the xml files by hand. > > True but the available languages are almost all hand-coded. > You could generate the schema.xml based on the available languages if not > hand-generated? >
Basically I would have to output a zip with schema.xml, solrconfig.xml and then all the resources specific to all the selected languages (stopwords, synonims, etc) for the languages that we can provide out of the box. For other languages, the admin would have to get dirty with the xmls.
> > There's one catch with this approach which is new to me but seems to be > quite important to implement this approach: the idf should be modified, the > Similarity class should be, so that the total number of documents is the > total number of documents having that language. > See: > >
http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201211.mbox/%3Czar...
> The solution sketched there sounds easy but I have not tried it. > > > 3) Indexing the content in different Solr cores (indexes), one for each > > language. Each core requires it's on directory and configuration files. > > - The advantage is that queries are clean to write (like option
and
> that > > you have a nice separation > > - The disadvantage is that it's difficult to get it right > (administrative > > issues) and then you also have the (considerable) problem of having to > fix > > the relevancy score of a query result that has entries from different > > cores; each core has it's own relevancy computed and does not consider > the > > others. > > - To make it even worst, it seems that you can not [5] also push to a > > remote Solr instance the configuration files when creating a new core > > programatically. However, if we are running an embedded Solr instance, > we > > could provide a way to generate the config files and write them to the > data > > directory. > > Post-processing results is very very very dangerous as performance is at > risk (e.g. if a core does not answer)... I would tend to avoid that as much > as possible. >
Not really related, but this reminds me about the post processing that I do for checking view rights over the returned result, but that's another discussion that we will probably need to have :)
> > > Currently I have implemented option 1) in our existing Solr integration, > > which is also more or less compatible with our existing Lucene queries, > but > > I would like to find a better solution that actually analyses the > content. > > > > During GSoC, option 2) was preferred but the implementation did not > > consider practical reasons like the ones described above (query > complexity, > > user configuration, etc.) > > True, Savitha surfed the possibility of having different solr documents > per language. > I still could not be sure that this was not showing the document match > single in one language. > > However, indicating which language it is matched into is probably > useful... >
Already doing that.
> Funnily, cross-language-retrieval is a mature research field but > retrieval for multilanguage user is not so! > > > On a related note, I have also watched an interesting presentation [3] > > about how Drupal handles its Solr integration and, particularly, a > plugin > > [4] that handles the multilingual aspect. > > The idea seen there is that you have this UI that helps you generate > > configuration files, depending you your needs. For instance, you (admin) > > check that you need search for language English, French and German and > the > > ui/extension gives you a zip with the configuration you need to use in > your > > (remote or embedded) solr instance. The configuration for each language > > comes preset with the analyzers you should use for it and the additional > > resources (stopwords.txt, synonims.txt, etc.). > > This approach helps with avoiding the need for admins to be forced to > edit > > xml files and could also still be useful for other cases, not only > option > > 2). > > Generating sounds like an easy approach to me. >
Yes, however I don`t like the fact that we can not do everything from the webapp and the admin needs to access the filesystem to install the given configuration on the embedded/remote solr directory. Lucene does not have this problem now. It just works with XWiki and everything is done from XWiki UI. I feel that losing this commodity will not be very well received by users that now have some new install steps to get XWiki running.
Well, of course, for the embedded solr version, we could handle it like we do now and push the files directly from the webapp to the filesystem. Since embedded will be default, it should be OK and avoid the extra install step. Users with a remote solr machine should have the option to get the zip instead.
Not sure if we can apply the new configuration without a restart, but I`ll have to look more into it. I know the multi-core architecture supports something like this but will have to see the details.
> > > All these problems basically come from the fact that there is no way to > > specify in the schema.xml that, based on the value of a field (like the > > field "lang" that stores the document language), you want to run this or > > that group of analyzers. > > Well, this is possible with ThreadLocal but is not necessarily a good > idea. > Also, it is very common that users formulate queries without formulating > their language and thus you need to "or" the user's queries through > multiple languages (e.g. given by the browser). > > > Perhaps a solution would be a custom kind of "AggregatorAnalyzer" that > > would call other analyzers at runtime, based on the value of the lang > > field. However, this solution could only be applied at index time, when > you > > have the lang information (in the solrDocument to be indexed), but when > you > > perform the query, you can not analyze the query text since you do not > know > > the language of the field you're querying (it was determined at runtime > - > > at index time) and thus do not know what operations to apply to the > query > > (to reduce it to the same form as the indexed values). > > How would that look at query time? >
That's what I was saying, that at query time, the searched term will not get analyzed by the right chain. When you search for a single language, you could add that language as a query filter and then you could apply the right chain, but when searching in 2 or more (or no, meaning all) languages you are stuck.
> > > I have also read another interesting analysis [6] on this problem that > > elaborates on the complexities and limitations of each options. (Ignore > the > > Rosette stuff mentioned there) > > > > I have been thinking about this for some time now, but the solution is > > probably somewhere in between, finding an option that is acceptable > while > > not restrictive. I will probably also send a mail on the Solr list to > get > > some more input from there, but I get the feeling that whatever > solution we > > choose, it will most likely require the users to at least copy (or even > > edit) some files into some directories (configurations and/or jars), > since > > it does not seem to be easy/possible to do everything on-the-fly, > > programatically. > > The only hard step is when changing the supported languages, I think. > In this case, when automatically generating the index, you need to warn > the user. > The admin UI should have a checkbox "use generated schema" or a textarea > for the schema. >
Please see above regarding configuration generation. Basically, since we are going to support both embedded and remote solr instances, we could support things like editing the schema from XWiki only for the embedded instance, but not for the remote one. We might end up having separate UIs for each case, since we might want to exploit the flexibility of the embedded one as much as possible.
> > Those that want particular fields and tunings need to write their own > schema. > > The same UI could also include whether to include a phonetic track or not > (then require reindexing).
> hope it helps. >
Yes, very helpful so far. I`m counting on your expertise with Lucene/Solr on the details. My current approach is a practical one without previous experience on the topic, so I`m still doing mostly guesswork in some areas.
Thanks, Eduard
> paul > _______________________________________________ > devs mailing list > [email protected] > http://lists.xwiki.org/mailman/listinfo/devs >
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Ludovic Dubost Founder and CEO Blog: http://blog.ludovic.org/ XWiki: http://www.xwiki.com Skype: ldubost GTalk: ldubost _______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
_______________________________________________ devs mailing list [email protected] http://lists.xwiki.org/mailman/listinfo/devs
-- Thomas Mortagne
Why/How can you end up with documents of other languages than the current default one if the wiki is mono-lingual?
by changing mind.
No I think it's different. The user can decide to make his search in "French" but look in the "French" + "English" dataset. For me if both documents with the same page name match, both should come out separately.
This can seriously be a problem considering that many browsers come, by default with many supported languages. paul
To summarize a bit, if we go with the multiple fields for each language, we end up with an index like:
English version: id: xwiki:Main.SomeDocument_en language: en space: Main title_en: XWiki document doccontent_en: This is some content
French version: id: xwiki:Main.SomeDocument_fr language: fr space: Main title_fr: XWiki document doccontent_fr: This is some content
Careful Eduard, that means that searching for common words (e.g "direction") would show you the same XWiki document as two different search results. I think you would want to combine the two in one Solr document. This is the old problem of Savitha which, I think, she has not solved. I've always believed that the fact that these two documents exist implies that they both have the same meaning and are translations of each other. I like the idea to carry the whole analysis of a language defined by the document language (it could be an object-field also).
Some extra fields might also be added like title_ws (for whitespace tokenization only) that have various approaches to the indexing operation, with the aim of improving the relevancy. One solution to simplify the query for API clients would be to use fields like "title" and "doccontent" and to put as values very lightly (or not at all) analyzed content, as Paul suggested. This would allow applications to write simple (and backwards compatible maybe) queries that will still work, but will not catch some of the nuances of specific languages.
It's a good idea to go backwards compatible with such a predictable behaviour as the whitespace analyzer. Le 27 nov. 2012 à 16:27, Jerome Velociter a écrit :
Thus, the search application will be the major beneficiary of these analyzed fields (title_en, title_fr, etc.), while still allowing applications to get their job done (trough generic, but less/not analized fields like "title", "doccontent", etc.).
I think for applications this complexity/implementations details would benefit being hidden behind a "query builder" interface of some sort, WDYT ?
Absolutely. Note also that such a query-expander (I believe this is the normal term) already exists within the EDismax. I'd add the expand-along-language function (where text becomes, if multlingual and browser that gives the languages en ro fr, text_en^3 text_ro^2 text_fr^1. Le 27 nov. 2012 à 16:44, Ludovic Dubost a écrit :
Maybe a solution would be to create one index per language and index ALL content regardless of it's language using the language analyzer of that index.
I fear that this will bring zillions of false positive because many people have a long list of supported language. Stemming is quite aggressive sometimes... for example searching for sitting will find all "sit", but this should not be the case if choosing French alone (then only the gathering of people is meant). If a browser indicates fr de as languages, and searches for sitting, this would find documents with attachments that contain this "sit" in any language, even though searching for English was not activated.
This later solution would be the only one that would really work on file attachements as we have no information about the specific language of file attachements (or even XWiki objects) which are attached to the main document and not to the translated document.
It is not entirely true that object fields and attachments do not carry a language. But I agree that there may be installations where the admin would prefer that attachments and objects are considered multilingual. Note that your solution is also doable with the multi-field approach of above and does not require several indices. There's the wiki language which one could apply to attachments and objects. There could be object properties doing this (e.g. Curriki has this), it would need to be customizable. There's the language of documents of sections in several file formats (e.g. in PDF or word files): while the current extractors do not honour this (I think), it could be used to switch analyzers. Again an option? ("index attachments in all languages", "index object fields in all languages") paul
participants (5)
-
Eduard Moraru -
Jerome Velociter -
Ludovic Dubost -
Paul Libbrecht -
Thomas Mortagne