Probabilistic Latent Semantic Indexing-英文文献.pdfVIP

Probabilistic Latent Semantic Indexing Pro ceedings of the Twenty-Second Annual International SIGIR Conference on Research and Development in Information Retrieval Thomas Hofmann International Computer Science Institute, Berkeley, CA EECS Department, CS Division, UC Berkeley hofmann@cs.b Abstract return the relevant data in some amenable representation, e.g., in form of a ranked list of relevant do cuments. Many Probabilis tic Latent Semantic Indexing is a novel approach retrieval metho ds are based on simple word matching strate- to automated do cument indexing which is based on a sta- gies to determine the rank of relevance of a do cument with tistical latent class mo del for factor analysis of count data. resp ect to a query. Yet, it is well known that literal term Fitted from a training corpus of text do cuments by a gen- matching has severe drawbacks, mainly due to the ambiva- eralization of the Exp ectation Maximization algorithm, the lence of words and their unavoidable lack of precision as well utilized mo del is able to deal with domain{sp eci c synonymy as due to p ersonal style and individu al dierences in word as well as with p olysemous words. In contrast to standard usage. Latent Semantic Indexing (LSI) by Singular Value Decom- Latent Semantic Analysis (LSA) [1] is an approach to p osition, the probabilisti c variant has a solid statistical foun- automatic indexing and information retrieval that attempts dation and denes a prop er generative data mo del. Retrieval to overcome these problems