12 Jul
2005
12 Jul
'05
8:35 p.m.
Hi Jie,
Thanks for your email.
For XWiki the way I would do it:
- Have the FAQBot hosted on a wiki page. This page would have a text input
field for letting users enter queries and it would list answers on the page.
We should also provide other interfaces like IRC/chat/mailing list.
- Have some macros that can be used in the same manner as the {excerpt}
macro is used by Confluence. One macro would be used to designate a
question, another one for an answer and yet another one for a srai.
- Have some pages that are AIML sets. Mark them using a XWiki object class
and load them when the FAQBot page is loaded.
- Create on the fly AIML sets by aggregating all information scattered in
the different pages (using the macros)
About the reinforcement. It could be done by different ways. One idea:
- On the FAQ bot page we could have a "teach" button where we would ask the
user to enter his question and the answer he would have liked.
- All these user questions/answers will be added automatically to a User
AIML set (a xwiki page containing AIML). This AIML set will also be loaded
by the FAQBot page (same as the other AIML sets).
- Whenever this user AIML set page is changed a diff email will be sent to
the XWiki dev mailing list. The idea is that XWiki community will provide
oversight about new questions/answers. The reason for the oversight is
double:
- to verify users do not enter stupid questions/answers
- to possibly refactor the existing AIML sets or pages containing custom
tags by removing the user-entered information and placing it where it is
more appropriate.
Again, all this should be packaged as a XWiki application. Jeremi Joslin
should be able to help you about the format (he's the one who's created the
notion of XWiki applications).
Last, we need to start deciding on the project directory structure in SVN.
Could you please come up with a proposal? Also, we're doing a generic FAQBot
so I'd like it to remain as separate from XWiki as possible. The generic
part and the XWiki specific ones should be cleanly separated in this
directory structure. Could you please come up with an architecture diagram
fro this?
Thanks
-Vincent
PS: I'm cc-ing the xwiki-dev mailing list. They may have some
ideas/remarks/feedback.
PPS: I'd need a formal updated proposition from you listing deliverables and
planning so that we're clean WRT Google. Including the architecture diagram
in the proposal would be nice.
-----Original Message-----
From: Jie Tang [mailto:jietang@fas.harvard.edu]
Sent: lundi 11 juillet 2005 08:01
To: Vincent Massol
Subject: RE: [SOC] Faqbot (ping?)
Hi Vincent,
Hope everything is going well with the baby, and congratulations. Take
whatever
time you need to get settled...I'll be ready whenever you are.
First off, I still think that the Faqbot is something I'd like to do. I've
been
thinking over how to implement some of the ideas for the bot this weekend,
in
addition to building some elementary aiml sets.
The idea behind the Xwiki plugin is that an xwiki page can have a section
designated as an faq (through some specially formatted comment?). When
this
page is updated by a user, the new faq content is run through the macro,
which
will transform the faq text into aiml expressions, which can then be
loaded by
the faqbot dynamically. Thus, in theory it wouldn't require a webmaster to
run
any extra utilities to have the faqbot incorporate the extra knowledge.
The
algorithm could have the bot read in the question part of the FAQ, looking
for
keywords that it already knows how to process. It would then try to
generate a
bunch of questions similar in wording to the new question, to allow the
bot to
provide answers to questions not worded exactly the same as the original
query.
It would associate the questions with the proper answer. The answer part
of the
FAQ would have to be processed too, to find key phrases which the bot
already
understands. This set of phrases would also be associated with the new faq
entry, and the bot would then ask one of these associated questions
depending
on the users response.
The tough part would be to figure out how exactly each entry refers to the
other; i.e., where the bot should go next after receiving a certain
response
from the user. One possibility would be extra formatting and markup,
describing
which other faq entries to look at for related knowledge, what to do if
so-and-so is the response, etc., within the XWiki page. While this would
be
easier to implement, the more markup is required the less useful this will
be
over writing new aiml files entirely.
The other idea is to implement some sort of reinforcement/learning in the
bot.
(this is more appealing to me, though I'm unsure how feasible it is).
Ideally,
a user would be able to rate the usefulness of the bot's responses to
their
queries, or suggest new responses; the bot would then be able to adjust
its
responses accordingly. This could be accomplished through creating
entirely new
aiml files, modifying existing aiml files (i.e., using a <random> element
and
weight certain responses more than others), or an exta system which causes
the
bot to learn and unlearn aiml. This would mimic a simple neural network
algorithm, implemented in aiml.
I've a feeling that I'll be working on the aiml sets for awhile. I look
forward
to hearing your thoughts on the new ideas?
Thanks, and take care,
Jie Tang
[snip]