Hi, Ben!

I'm not dead yet --- I've been in the midst of a couple of papers (one of which was just accepted yesterday to the International Conference on Weblogs and Social Media) and a grant proposal on comptranslation and other Human Language Technologies (HLT) due Mon 05 Feb 2007. I will have a bit more time to reanimate this blog and to follow up on interesting discussions after that's done.

As for Bayesian theories in discourse and dialogue, I was just talking with one of the participants in the HLT proposal about that yesterday. The state of the field, as you may know, tends more towards case-based reasoning (CBR) and analogy, but there is a trend in AI towards graphical models as causal representation. Here are specific Bayesian models that may be relevant:

• Charniak and Goldman. There's a paper in Artificial Intelligence on a Bayesian network-based system for plan recognition.


• The above is a precursor to the work of Horvitz, Heckerman, Cadez and Kadie of MS Research on related subjects (some of their applications, such as Lumiere, was billed as a "decision network for the Microsoft Office Assistant", which at the time was much reviled due to the high annoyance factor of Clippy). There's a Communications of the ACM article about this c. March, 1995 and some tech reports on Heckerman's publications page.


• Since then, Bayesian networks have been used in user interfaces ranging from facial context tracking (Ashutosh Garg, an Illinois Ph.D. student at the time, gave a demo of this at IJCAI-2001 amid Gates's keynote talk) to web clickstream mining (the one by the above four authors was an application either for MS Small Business Server or the commerce server edition of IIS) to user modeling (Horvitz gave a demo of a prototype Outlook 2007 or Commerce Server app at UAI-2002 in Edmonton, that tracks "most probable time that someone will return from AFK").


• Google is presently leading the field in certain aspects of information extraction, e.g., gist summarization, fusion of data sources for query answering (QA), etc. Peter Norvig gave a talk at IJCAI-2005 in Edinburgh titled "What to do with all the world's data".

BTW, there are of course many, many more applications than those developed by Microsoft and Google; these are just the ones I know of that spring immediately to mind. I did not go to IJCAI-2007 in Hyderabad last month, though I did send martin_samuel, so you might get a better state-of-the-field update from him this spring. I'll check with him.

I've been reading Pinker's The Language Instinct, Melamed's Empirical Methods for Exploiting Parallel Texts (reviewed here by Ted Pedersen and here by Constantin Orasan), and Gee's An Introduction to Discourse Analysis. When I've had a chance to digest the ideas and synthesize them more, and discuss them with my colleagues who have much more formal linguistics and cognitive linguistics background than I do, I will be able to give you a more useful answer.

What do you mean by "decisioning"? There are of course decision theoretic models of shared beliefs/desires/intentions as well as shared goals and responsibilities. Some of these have a probabilistic foundation.

(continued)

Katia Sycara at CMU does work on dialogue-driven negotiation, and you might take a look at the work of her (even) more Bayesian colleagues, Moshe Tennenholtz and Sarit Kraus.

I have an interest in Bayesian generative approaches to interactive fiction (IF), but it's on the back burner for now because of the information extraction (IE) and machine translation (MT) focus of our current work. IIRC, Mark Kantrowitz did some interesting work on IF at CMU in the early 1990s when he first started working on his Ph.D.

Thanks for the IPCC link! I'll look at it in about 55 hours. :-)

--
Banazir

I see you pointed out Katia Sycara ... sorry, got distracted.

BTW: googling for Amir Evenchik didn't bring any homepage, but did bring up "Multiple Criteria Group Decision Support Through the Use of ..." - "Reaching agreements. through argumentation: a logical model and implementation." (PDF) ... that explicates (yet again) what I'm trying to integrate with popular web practices.