Learning a Flexible Question Classifier

To generate high quality answers, many question taxonomies designed for modern question answering systems are getting more and more complex and fine-grained. Furthermore, without concrete context some questions are ambiguous and are difficult to be correctly labeled by question classifier, even by people manually. All above bring a big challenge to current question classifiers. However, previous research seldom pays attention to these situations above. In general, the labeled question dataset is usually small, so a feasible solution to these issues is to integrate new feedbacks and certain domain-specific knowledge into current model continuously. In this paper we explore the application of an online learning algorithm to question classification. The experimental results show that the performance of our approach is comparable to other popular learning algorithms: SVMs and SNoW. Furthermore, we evaluate our approach on ambiguous questions and the results prove its feasibility and efficiency.