ADAPTING SEARCH ACCORDING TO QUERIES: Experimental studies

WWW information access is based on search engines. Most of the users consider a single search engine that they will use whatever their queries are. Additionally, current search engines follow a general framework and they treat all the queries in the same way. In this paper, we present experiments that show that considering variability both in terms of systems and in terms of queries would improve the results of a search. We based our experiments on the data from the international evaluation forum TREC (200 topics, several Gbytes of documents). This forum simulates search engine usage and gives access to topics, document collections, expected documents and system answers. We show that high precision can be improved by automatically deciding which system should be used for a given query. In our approach, system selection is based on query clustering; queries are clustered considering linguistic features that are automatically extracted from queries. We show that high precision (P@5) is improved by 8.9% for P@5 over the best system for TREC 5 for example. We also show that considering a larger set of collections, P@5 is improved of about 5%, when considering a learning/testing evaluation framework.