Pattern Training Phase Pattern Application Phase

The syntactic information latent in any coherent text can be exploited to overcome some inadequacies of keyword-based retrieval and make information retrieval more eeective. We have earlier quantitatively demonstrated how syntactic information is useful in ltering out irrelevant documents. We have implemented a system which exploits a rich syntactic representation of supertags in a exible manner to lter documents for relevance. The system has been tested on a large collection of newswire sentences, and achieves recall and precision gures of 88% and 97% for ltering out irrelevant documents. Its performance and modularity makes it a promising postprocessing addition to any Information Retrieval system. In this paper we examine how the performance of this system is aaected by varying the context provided to the system and show that the experimentally determined optimal size of the context validates linguistic intuitions .