User information extraction for personalized dialogue systems

Abstract

We propose a method to extract user information in a structured form for personalized dialogue systems. Assuming that user information can be represented as a quadruple 〈predicate-argument structure, entity, attribute category, topic〉, we focus on solving problems in extracting predicate argument structures from question-answer pairs in which arguments and predicates are frequently omitted, and in estimating attribute categories related to user behavior which a method using only content words cannot distinguish. Experimental results show that the proposed method significantly outperformed baseline methods and was able to extract user information with 81.2% precision and 58.1% recall.