Learned Vector-Space Models for Document Retrieval

Bellcore's Latent Semantic Indexing (LSI) system and HNC's MatchPlus system represent two attempts to model and exploit the inter-relationships among terms to improve information retrieval. Most information retrieval methods depend on exact matches between words in users' queries and words in documents. Typically, documents containing one or more query words are returned to the user. Such methods will, however, fail to retrieve relevant materials that do not share words with users' queries. One reason for this is that the standard retrieval models (e.g., Boolean, standard vector, probabilistic) treat words as if they are pairwise orthogonal or independent, although it is quite obvious that they are not. Consider, for example, the terms "automobile", "car", "driver", and "elephant". The terms "automobile" and "car" are synonyms, "driver" is a related concept, and "elephant" is pretty much unrelated. In most retrieval systems the query "automobile" is no more likely to retrieve an article about cars than one about elephants, if neither author uses precisely the term automobile. It would be preferable, however, if a query about automobiles would retrieve articles about cars and even articles about drivers to a lesser extent. A central theme of both the LSI and MatchPlus methods is that term-term inter-relationships like these can be explicitly modeled in the representation and automatically used to improve retrieval.