A robust multivariate reranking algorithm for Question Answering enrichment

Cognitive research indicates that multimedia information have benefits over text-only information by means of intuitive and vivid expression. In this paper, we introduce a novel multimedia Question Answering approach to enrich text answers with image and video information. Especially, we propose a robust multivariate reranking algorithm which adopts a squared hinge loss function for learning ranking functions from preference information. Given a question and the community-contributed answer, our approach determines whether it should perform enrichment, for which a list of relevant and high-quality web-scale multimedia data are enriched and ranked automatically by reranking algorithm. Different from some efforts that attempt to directly use image and video data to replace text answers, enriching text in this way may be more feasible and useful to deal with more complex questions. The empirical experiments conducted on a wide variety of daily questions show the effectiveness of our approach.