A user reported some confusion about the Senseval-1 data, where there were words for which we had training data, but no test data. That was a bit of a puzzle, but Saif seems to have dug to the bottom of it!
========================================================================== >From Saif... Heres what I think is happening with the Senseval-1 instances of "amaze". The original Senseval-1 test data has instances corresponding to amaze-v (but none for amaze-a). The training data has instances for both (amaze-v and amaze-a). However, the training data for both amaze-v and amaze-a are in a single file -- amaze.cor. Further if we look at the original Senseval-1 data we see that the instances are not explicitly said to belong to "amaze-v" or "amaze-a" but in general to "amaze". So at first look it seems as if we have a training file for amaze and a test file for amaze -- so alls fine. But in reality, the original data has training data for "amaze-a" and no test data for "amaze-a". Why so, I do not know. The data we have on Ted's page is Senseval-1 data in Senseval-2 data format. this was created using the Sval1to2 package using the --split_on_POS option. This way we have separate files for each part of speech. So while we have training files for "amaze-v" and "amaze-a", we have test file for just "amaze-v". As mentioned earlier the original data itself gives no explicit indication as to which part of speech a particular instance corresponds to. So when we use the --split_on_POS option, the program Sval1to2.pl actually looks up the sensemap file which has the following entries for amaze: amaze 512617 aston a amazing amaze 512619 astonv v amaze Instances tagged with sense-id 512617 go into amaze-a training file while instances tagged with sense-id 512619 go into amaze-v training file. This explains what happens to the data during the conversion process from Senseval-1 format to Senseval-2 format but we still do not know why we do not have test instances for amaze-a in the original test data. My guess is that when the Senseval-1 test and training files were created they did not explicitly distinguish between the parts of speech. So as far as they were concerned they had a test and training file for "amaze". Either they did not realize that the test data had no instances of amaze-a or they made a conscious decision not to have any. -Saif ------------------------ Yahoo! Groups Sponsor --------------------~--> Give the gift of life to a sick child. Support St. Jude Children's Research Hospital's 'Thanks & Giving.' http://us.click.yahoo.com/h8TXDC/6WnJAA/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/
