Abstract
Esfinge is a general domain Portuguese question answering system, which uses the information available in the Web as an additional resource when searching for answers. The experiments described in this paper show that the Web helped more with this year’s questions than in previous years and that using the Web to support answers also enables the system to answer correctly or at least to find relevant documents for more questions. In this third participation in CLEF, the main goals were (i) to check whether a database of word co-occurrences could improve the answer scoring algorithm and (ii) to get higher benefits from the use of a named entity recognizer that was used last year in sub-optimal conditions. 2006 results were slightly better than last year’s, even though the question set included more definitions of type Que é X? (What is X?), with which the system had the worst results in previous participations.