Verb-based relations and conceptual proximity in information retrieval

In many text retrieval systems, nouns are considered the primary information-bearing words, with verbs being generally considered of little value. Yet a natural language query often contains one or more verbs and these verbs often specify significant relationships between the query concepts. In this project we investigate the effect of treating verbs as relations between noun-specified concepts. The primary question is whether explicitly searching for these relationships can significantly improve the effectiveness of an information retrieval system. The following five research questions are addressed: (1) How often does more than one concept occur in a query?, (2) How often do verb-based relations occur in queries?, (3) What effect does explicitly looking for relations have on retrieval effectiveness?, (4) How sensitive is retrieval effectiveness to the proximity window in which the concepts and relations occur?, and (5) How can we identify queries that would benefit significantly from using explicit relation searches?; To investigate the first two questions, we analyze 200 queries used in the Text REtrieval Conference (TREC) experiments and find that multiple concepts and verb-based relations are common. To investigate the last three questions, we create concept-relation-concept query graphs and compare the retrieval effectiveness of searches generated from these graphs using Boolean AND and various proximity operators, both with and without the relation. We find that Boolean AND performs significantly better than any proximity operator, both for searches that contain relations and searches that do not. We also find that, in general, retrieval effectiveness is significantly degraded by including a relation in a search and that the effect of a relation on any given graph is not predictable. There may be specific applications for which verb-based relations can provide significant benefit. We describe possible characteristics of such applications.