Evaluation of utility of LSA for word sense discrimination.pdfVIP

Evaluation of Utility of LSA for Word Sense Discrimination Esther Levin Mehrbod Sharifi Jerry Ball Dept. of Computer Science Dept. of Computer Science Air Force Research Laboratory City College of New York City College of New York 6030 S Kent Street NY, NY 10031 NY, NY 10031 Mesa, AZ 85212-6061 esther@cs.ccny.cuny mehrbod@ Jerry.Ball@mesa.afmc.af.m .edu il The critical step of the LSA algorithm is to com- Abstract pute the singular value decomposition (SVD) of the normalized co-occurrence matrix. If the matri- The goal of the on-going project de- ces comprising the SVD are permuted such that the scribed in this paper is evaluation of the singular values are in decreasing order, they can be utility of Latent Semantic Analysis (LSA) truncated to a much lower rank. According to Lan- for unsupervised word sense discrimina- dauer and Dumais (1997), it is this dimensionality tion. The hypothesis is that LSA can be reduction step, the combining of surface informa- used to compute context vectors for am- tion into a deeper abstraction that captures the mu- biguous words that can be clustered to- tual implications of words and passages and gether –