Landauer Multiple meanings in LSA 1 Single representations of multiple meanings in Latent Semantic Analysis

WW XX YY ZZ YY is kleeper than WW. YY is kleeper than ZZ. Is YY kleeper than XX? Is "kleeper" ambiguous? Latent Semantic Analysis (LSA) is a psychological model and computational simulation intended to mimic and help explain the way that humans learn and represent the meaning of words, text, and other knowledge. In this chapter I briefly describe the underlying theoretical and computational machinery of LSA, review some of the surprising things it is able to do, and discuss some of its limitations and possibilities for future development. I will concentrate on what LSA has to say about multiple word meanings, where it succeeds and fails, and what is needed to fix it. For researchers and theorists concerned with word meanings and ambiguity, the most important implication of the LSA theory is that it questions the idea that different senses of a word have separate and discrete representations that are individually disambiguated. Instead, it represents a word meaning as a single point in a very high dimensional semantic space. In LSA, the acquisition of a word meaning is an irreversible mathematical melding of the meanings of all the contexts in which it has been encountered. In comprehension, words are not disambiguated by sense one at a time; their different effects in different contexts are merely determined by how they combine. Thus, a word has a different "sense" every time it is used.