A Deep Semantic Matching Network for Answer Selection

Answer selection is a critical task for a conversation system. The previous methods about the answer selection cannot incorporate the context dependency information into the representations of the sentences in the conversation, which ignores the effective part for the answer selection. In this paper, we design an algorithm that contains attention mechanism to compute the matching degree for question-answer pairs. In order to capture the context dependency information between the sentences, the question-answer pair is first processed by the attention module designed in our network. The semantic matrices at different levels of granularity are constructed to carry abundant and computer-understandable semantic information of the sentences. Next, the multiple matching matrices are generated to represent the matching between them. More abstract and sophisticated matching features are extracted from the matching matrices by our network. Finally, the matching features are used to compute the matching degree of the question-answer pairs. We evaluate our proposed architecture on a benchmark dataset about the answer selection. Experimental results show that our architecture, which contains the attention mechanism, is more effective and more superior than baseline models.