Boosting Weak Ranking Functions to Enhance Passage Retrieval for
Question Answering.
Nicolas Usunier, Massih-Reza Amini, Patrick Gallinari
Laboratoire d'Informatique Paris 6
8, Rue du capitaine Scott
75015 Paris
We investigate the problem of passage retrieval for QA systems. We
adopt a machine learning approach and apply to QA a boosting algorithm
initially proposed for ranking a set of objects by combining baseline
ranking functions. The system operates in two steps. For a given
question, it first retrieves passages using a conventional search
engine and assigns each passage a series of scores. It then ranks the
returned passages using a weighted feature combination. Weights
express the feature importance for ranking and are learned to maximize
the number of top ranked relevant passages over a training set. We
empirically show using questions from the TREC-11 question/answering
track and the Aquaint collection that the proposed algorithm
significantly increases both coverage and precision with respect to a
conventional IR system.