A convolutional neural network model for non-factoid Chinese answer selection

Previous research on answer selections was mainly concentrated on factoid questions. Nevertheless, non-factoid questions are equally important because they are common in community question answer (CQA). In this paper, we propose a semantic calculation based convolutional neural network (SCCNN) to address this problem. The main difference of SCCNN from previous models is the use of correlation information between question and answer (QA) sentences. This method calculates correlation features for all the words in sentences based on the similarity between a question and its corresponding answer. It will also be the direct input of the convolutional neural network. These information represents correlation degrees of all the tokens and allows for focusing on the important part of QA sentences. Experimental results show that our algorithm performs better than previous methods. In addition, we compare methods modeling Chinese words and characters individually, which indicates that for CNN-based methods, modeling the single character usually have better performance.