As you might recall we discussed the paper by Gliozzo, et. al. (ACL-2005) on domain kernels for word sense disambiguation during our meetings of Oct 14 and Oct 28.
We had a few questions on that work that Mahesh has kindly written up, so I thought I would post those here "for the record" so to speak. We are still working on getting answers, so if you have any thoughts that would be most appreciated. The following is all by Mahesh Joshi: First, a brief summary of our understanding of the paper ... ========================================================================== Domain kernels for WSD is a method to evaluate similarity among contexts of the ambiguous words, using domain knowledge gained from external unlabeled corpora. The process of gaining external knowledge involves initially creating a conventional term-by-document (k-by-n matrix) vector space model out of unlabeled corpus. The process of LSA converts this matrix to a domain matrix of size k-by-k'. This domain matrix is used to transform each k-by-1 document (column) vector in the vector space model to a 1-by-k' domain vector (row) for that document. Thus instead of the document being represented in a k dimensional feature space, it is now represented in a much smaller k' dimensional feature space. Since there were n such documents in the vector space model, there will be n such k' dimensional domain vectors, after applying equation (1) to all documents. (So a matrix view of this would correspond to n documents along the rows and k' dimensions along columns, thus corresponding to a n-by-k' size matrix, right? For reference, I will use the term "document domain vector matrix" for this matrix.) On the the other hand, domain vectors for words are directly obtained from the domain matrix (the rows). So all domain vectors (for documents or for words) have k' dimensions. Since the vector space model is created from unlabeled corpus, the rows of the "document domain vector matrix" do not have a direct correspondence to the contexts of the ambiguous word, as the contexts come from the labeled lexical samples. The terms t-sub-i and t-sub-j in formula (4) refer to the 2 contexts of ambiguous word and similarity between these contexts is to be evaluated. ========================================================================== With this understanding, the question that we were discussing was as follows: For evaluating formula (4), we will require the domain vectors for the 2 contexts of the ambiguous word. We were not sure how these domain vectors for the contexts are computed. Could you please let us know the same? There does not seem to be a one-to-one correspondence between these contexts and the rows of the "document domain vector matrix" and the domain matrix gives the domain vectors only for the words. Are the domain vectors (from the domain matrix) for the words of the context added together to create the domain vectors for the contexts? Also, in this case what happens if some word in the context does not have a corresponding domain vector in the domain matrix? Or is there some other method used for the computation of the context domain vectors? ------------------------ Yahoo! Groups Sponsor --------------------~--> DonorsChoose.org helps at-risk students succeed. Fund a student project today! http://us.click.yahoo.com/9.ZgmA/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/
