关键词: CLSM, 注意力机制, 检索式匹配问答

Abstract: The traditional Retrieval Matching Question and Answer model (RMQA) uses some old methods in the feature extraction of questions and candidate answers, which ignores the semantic information of sentences and affects the accuracy of question and answer results substantially. Recently, the Convolutional Latent Semantic Model (CLSM) has got good results in extracting sentence semantic features. As for RMQA task, we drop the word-n-gram layer and letter-n-gram layer of the CLSM to build a CLSM-based semantic feature extraction model. And we also propose an entity_attention layer to balance the matrix’s value of important entity. We design three groups of contrast experiments, the experimental results showed that the improved model had a 4%-10% improvement over the traditional translation model in NDCG. It is proved that the model can strengthen the information of the core word through the entity concern layer, and use the convolution neural network to capture the useful contextual structure information of semantic matching effectively so as to improve the accuracy of the retrieval matching question and answer.

Key words: CLSM, Attention, RMQA