Question Classification via Multiclass Kernel-based Vector Machines

Question classification is an important component of most modern question answering systems. At the same time, the question taxonomies designed for modern question answering systems are more and more complex and fine-grained. Moreover, the prediction of question type is required to be more specific to generate better answers, especially for some ambiguous questions. All of demands above bring a new challenge to current question classifiers. Learning strategy and features involved in learning are vital factors to quality of question classifier. However, few researches have made serious investigations on learning strategy itself but the exploration of new type features. In this paper we develop a question classifier based on multiclass vector machines to learn a direct multiclass mapping from questions to question categories, with only lexical feature that makes our question classifier be independent of some specific language. The experimental results showed the performance of our question classifier is comparable to question classifiers based on other popular learning algorithms.