Re: [ba-unrev-talk] Automated Email-based CHAT-FAQ Expert System
I just spent the last couple of days hooking up a relational database to
KnownSpace http://www.knownspace.org and got it working, sortof. Couple
dozen bugs left, but hey, what the heck. (01)
KnownSpace has an email client. It also has lots of little agents that do
things when events are generated, like when a new email is received. It
includes a Latent Semantic Analysis agent that can play nice with
keywords. I am thinking that KnownSpace (Java, Apache license) would make
a great starting place. BTW: as of this morning, there exists
http://sourceforge.net/projects/knownspace although since it just went up
today, nothing is there yet. (02)
Jack (03)
At 03:10 PM 1/2/2002 -0800, Eric Armstrong wrote:
>This message encasulates a recent email exchange in which
>I thought out more of the concepts involved in an automated,
>interactive FAQ system:
>
>Doug wrote:
>
> > I'm always happy to answer questions directly, if it will save people
>time.
> > I'm glad you asked.
>
>Yeah. It is amazing how what's on the top of your head for one person is
>
>miles away for someone else. Like Eric Raymond wrote on the subject of
>open source software: "With enough eyes, all bugs are shallow."
>
>I would absolutely LOVE to work on a system that monitored email
>traffic, storing responses to questions, and which made it possible for
>people to discover answers through automated queries.
>
>I figure that such a system would make a best try at finding an answer,
>but that people would always be involved to interpret the question,
>answer it, and make the system smarter.
>
>There is some ontology work using topic maps and the like that would
>be spectacularly useful for such an undertaking. My dream job. Good
>for a year or two, at least.
>
>Doug wrote:
>
> > Sounds great.
>
>It would be a great project to work on. I'd set it up as a filter
>on the company mailing list, with a companion screen that
>people could use to modify the ontological underpinnings of the
>system.
>
>That way, the automated system could take a stab at replying
>whenever it could, saving people from having to repeat themselves
>too often. (That would be the idea, at least.)
>
>Clearly, the system has a lot of potential benefit. Who do we know
>that would want to fund such a thing?
>
>Doug wrote:
>
> > Nobody else has attempted to solve this problem already?
> > I'd love to dump my tips into a system that everyone could
> > update/contribute to.
>
>That would be a great place to start building a database of
>automatically-
>answerable questions.
>
>This is a problem that is poised on the cusp of becoming solvable. There
>
>have been some attempts at intelligent customer service systems, but so
>far they have fallen rather flat.
>
>The tack I recommend -- an integrated email conversation with an
>answerbot sitting in and being educated, has never really been tried.
>It is an approach I have been recommending for most of my professional
>career -- that an integrated man/machine systems will always outperform
>either one by itself.
>
>Then, too, the recent work on ontologies has a tremendous capability
>for expediting such a system. For example, if a user asks:
>
> "I put floppy #2 in the drive, but I need to back up and put in #1
>first,
> How do I do that?"
>
>With a keyword-driven system, the machine sees "floppy" and
>"back up", and gives the user instructions for copying files. Not very
>helpful!
>
>But an ontology-based system potentially has the capacity to recognize
>the context as "floppy in drive" and the "back up" as meaning "go back
>to previous step". It can then deduce that the instructions are:
>
> * put the computer where you can see it
> * turn on the lights
> * find the slot where you put the floppy in
> * press the button on the slot to eject the floppy
> * find the floppy labeled "#1"
> * insert it
>
>Doug Lenat's cyc system, for example, has enough "common sense" to
>know that "back up" means eject, in this case. Even so, capturing the
>ontology-information necessary to recognize and successfully answer
>a query like that is no easy task. And trying to anticipate every
>possible question in advance is both prohibitively expensive and
>impractical.
>
>The system I propose would have 4 components:
> * Users
> who make queries via email
>
> * The AnswerBot
> which listens to queries, and which attempts to answer
> when it can
>
> * Experts
> who answer them when the AnswerBot can't,
> or who correct the AnswerBot's misguided attempts
> (and who act as users in other areas)
>
> * Ontology Librarians
> who add ontology-information that make it possible for the
> system to find information it already has, but didn't know
> how to relate to the specific query.
>
> One way to do that is by importing the topic map versions of
> the stuff Lenat has released. (Some material is free. More
> can be purchased.) Another way is by adding additional
> ontological structure: For example the fact that "back up"
> can mean "go back a step" as well as "copy files", depending
> on context.
>
> Note:
> Ideally, Experts will answer with pointers into an indexed FAQ,
> which will make it easy for the ontology librarians to recognize
> that the information already exists. Otherwise, they can capture
> the information as a new nugget for the FAQ.
> [Purple numbers, anyone?]
>
> In addition to maintaining the ontology library, these folks use
> their human intelligence to recognize which reponses were
> helpful, and create new info tidbits (nuggets/tips) for the system.
>
> So in a way they are documentation / customer support /
> expert-system-administration gurus. (04)