Statistical Selection of Exact Answers (MultiText Experiments for TREC 2002).

Abstract

Introduction For TREC 2002, the MultiText Group concentrated on the QA track. We also submitted runs for the Web track. Building on the work of previous years, our TREC 2002 QA system takes a statistical approach to answer selection, supported by a a lightweight parser that performs question categorization and query generation. Answer candidates are selected from passages retrieved by an algorithm that identi es short text fragments containing weighted combinations of query terms. If the parser is able to assign one of a predetermined set of question categories to the question, the system employs a nite-state pattern recognizer to extract answer candidates. Otherwise, oneto ve-word n-grams from the passages are used. Despite the simplicity of the approach, our best QA run returned correct answers to 37% of the questions. Our basic question answering strategy is an extension of the technique we used for both TREC 2000 and 2001 [5]. In past years, our system ranked individual terms