The minutes and discussion regarding applications of WSD has been very interesting, especially for those of us who didn't attend the workshop in Barcelona. Thanks to all of you who have contributed to this!
At the risk of repeating what might have already been said at that event, here are a few misc thoughts about WSD and applications. Lexical sample word sense disambiguation seems most useful to me in terms of studying the properties and behavior of a particular word or sense. I personally find this fascinating, and seeing how you can figure out which meaning of "line" or "train" was intended in a particular context still seems like a very worthwhile pursuit. It's about understanding a word, or a particular sort of feature, or the nature of a particular sense inventory or sense distinction. But as a practical matter, I can't think of applications based on lexical sample disambiguation, simply because to achieve any coverage you end up with thousands of classifiers, each of which know alot about a particular word (but not how those words interact with each other). Training each of those classifiers is hard and requires lots of data, and of course there is the issue of fixing upon a sense inventory that might not be the best choice for all applications. The other significant potential downside to lexical sample style disambiguation is that it can easily turn into an exercise in machine learning methodology, where the end result are accuracy values in a table, rather than contributing something to an application. This isn't all bad, in that word sense disambiguation has told us lots about machine learning methodology (merits of naive bayes, co-training, svms, decision trees, etc). But it doesn't really lend itself to integration in applications as best I can tell. However, let's imagine for a moment that there was something like the Brill Tagger for WSD - free code, easy to install, use, and understand, and that it performs all-words sense disambiguation reasonably well (whatever that might mean). I think then applications would come out of the woodwork. One example, I was at AAAI this July (part of the reason why I wasn't at ACL) and I talked to various folks who really seemed to want a system that disambiguates all content words in text. Given the availability of such a tool, they would have at least tried it out, and maybe some clever new use of WSD would come from that. My own experience with some software we've released is that people quickly discover ways to use it that I could have never predicted, and I think reaching out to potential users of WSD with tools they can use is really what we need to resolve this issue of applications and WSD. My conclusion? If we want to deal with applications, we need to focus more on all-words disambiguation. In short, we need to have tools like the Brill Tagger available for WSD. Maybe not perfect, maybe not even the best possible result, but easy to use and try out. Finally, I would be remiss if I didn't mention that I think data-driven clustering approaches are a very important future directions. But that's another topic. The point of this note is that I still think there is a place for "traditional" WSD, but that we need to focus on getting our tools and ideas in the hands of application developers so they can be the ones to decide when and how to use WSD, and whether or not it helps. Cordially, Ted PS To practice what I preach, one of my higher priorities these days is to revamp our lexical sample SenseRelate system to handle all words disambiguation (http://senserelate.sourceforge.net) -- Ted Pedersen http://www.d.umn.edu/~tpederse ------------------------ Yahoo! Groups Sponsor --------------------~--> Make a clean sweep of pop-up ads. Yahoo! Companion Toolbar. Now with Pop-Up Blocker. Get it for free! http://us.click.yahoo.com/L5YrjA/eSIIAA/yQLSAA/x3XolB/TM --------------------------------------------------------------------~-> Yahoo! Groups Links <*> To visit your group on the web, go to: http://groups.yahoo.com/group/nlpatumd/ <*> To unsubscribe from this group, send an email to: [EMAIL PROTECTED] <*> Your use of Yahoo! Groups is subject to: http://docs.yahoo.com/info/terms/
