Two Level Question Classification Based on SVM and Question Semantic Similarity

Question classification is very important in question answering system. This paper presents our research about question classification in a real-world on-line interactive question answering system in computer service & support domain. In the domain, questions are divided into 15 cursory categories and 220 sub-categories. The difference of this system is that standard question sentences represent the subcategories rather than only classification criterion. For the special situation, the two level question classification method is present in the paper. Support vector machine method is adopted to train a classifier on coarse categories; question semantic similarity model is used to classify the question into sub-categories. The lexical feature and domain ontology concept hierarchy is constructed and exploited to enhance the expression capacity of the feature characteristic for both feature selection for SVM and question semantic similarity computing. When trained and tested on the 11000 question instances in the domain, our approach reaches an accuracy up to 91.5%, which outperforms the result of the baseline.