Today (Sept 30) we discussed the ACL 2005 paper by Patrick Pantel entitled Inducing Ontological Co-occurrence Vectors. Thanks very much to Anagha for leading the discussion and guiding us through this interesting but complex work.
During the discusssion, I speculated that it might be interesting to run Patrick's CBC demo on the "chair" example in the paper, just to see what sort of output we get. I encourage you to to the same, it's sort of fun, and you can see more detail than I will show here. The demo is at the following site, and I used the TREC-2002 data (it gives you a choice of several corpora). http://www.isi.edu/~pantel/Content/Demos/LexSem/cbc.htm In looking at WordNet, I think the "chair" diagram in the paper is either different than WordNet, or we interpreted it wrongly. Note that the parent of the "chair" sense is presiding_officer, not chairwoman. (The diagram in the paper suggests that chairwoman is the parent of the "chair" sense, which seemed a bit strange to us). Anyway, this is what WordNet 2.0 thinks about it: president, chairman, chairwoman, chair, chairperson => presiding officer => leader => person, individual, someone, somebody, mortal, human, soul => organism, being => living thing, animate thing => object, physical object => entity In any case, when you run CBC on each of the words in the "chair" sense, you get the following: chair ------ chair, sofa, table (N615) - 0.463049 meet, hold, arrive (V1326) - 0.169118 co-chairman, co-chairmen, co-chair (N1070) - 0.147566 chairwoman ---------- vice-chairman, vice chairman, chairmen (N1072) - 0.3424 president --------- leader, President, member (N210) - 0.497118 chairman -------- leader, President, member (N210) - 0.433116 chairperson ----------- vice-chairman, vice chairman, chairmen (N1072) - 0.342201 ------------------------------- So, it looks to me like *perhaps*... vice-chairman, vice chairman, chairmen, leader, President, member would be propogated up in the committee based method, and that nothing would be propogated up in the shared propogation method (since the latter requires the features to be shared amongst 3 or more of the words). Does that seem reasonable? I am vaguely surprised by this outcome, so perhaps I have messed it up! BTW, I understand that we did not arrive at a definite interpretation of the committee propogation, so perhaps I have totally missed the point. :) Also, there are actually more than 3 features for each word, I think it just shows that many by default - in the web interface you can click down and see more, but that started to get a bit messy. But, perhaps that would lead to more results for the shared propogation method. In any case, this was a very interesting discussion, and I enjoyed it very much. Please do keep thinking about this and related methods, as I said I like them very much because they represent a merger of distributional methods (like SenseClusters) with WordNet (WordNet-Similarity). We saw this last year in the McCarthy paper from ACL-2004, and I think this is the sort of work we can expect to see in future from others and perhaps even from us! Ted -- Ted Pedersen http://www.d.umn.edu/~tpederse ------------------------ Yahoo! Groups Sponsor --------------------~--> DonorsChoose.org helps at-risk students succeed. Fund a student project today! http://us.click.yahoo.com/G7lQgA/FpQLAA/HwKMAA/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/
