Combining the Ranked Output from Fulltext and Summary Indexes

This paper demonstrates the advantage of combining a ranked output from a fulltext index and another from an index generated using generic summary texts. Our main ndings are as follows: (1) The combination of a fulltext-based run and a summary-based run can signi cantly outperform either alone, especially when the fulltext evidence is emphasized. (2) The primary role of the summary index in the above combination is to promote some of the relevant documents already retrieved by the fulltext index, rather than to discover new relevant documents. (3) Moreover, though the summary quality can make a di erence for this combination approach, even simple extracts such as 5% \lead" may give signi cant improvements.