Research and implementation of question classification model in Q&A system

Abstract

Question classification is the core of the question-and-answer (Q&A) sys-tem. This paper intends to use the method of deep learning to explore the question classification model in Q&A systems, the aim of which is to improve the accuracy of question classification. The characteristics of natural language questions, such as the use of short texts and basic grammar, were well considered. Subsequently, we want to fully extract the features of questions by using the following methods: multi-channel inputs, multi-granularity convolution kernels, and direct connection with high-speed channels. By combining the three methods, this paper proposes the multi-channel and Bidirectional long-and short-term memory and multi- granularity convolution neural net-work (MC–BLSTM–MGCNN) model to fully extract the features from interrogative sentences, both in time and spatial domains. To verify the validity of the model, this paper experimented with the TREC [1] classification standard dataset. Results achieved 96.6% accuracy, which is superior to the highest existing industry benchmark (96.1%). In addition, this paper used the complete TREC dataset to innovate further, and results obtained 98% accuracy, which greatly improved the classification.