Information Retrieval based on Paraphrase

Text Retrieval systems based on ranking use similarity as an approximation to relevance. Most of these systems ignore word meaning. We assume that some measure of paraphrase would be a better similarity measure. We develop a concept of paraphrase based on Meaning-Text Theory and implement an approximation to the ideal using the Longman Dictionary of Contemporary English (LDOCE). The performance of the new system is assessed using recall and precision averages on two standard collections. We discuss the results and conclude that we could improve performance if only the restricted vocabulary of the dictionary entries had a particular property. We then propose a technique for modifying the entries using statistical methods recently used in information retrieval.