Multi-layer attention neural network for sentence semantic matching

Sentence semantic matching, which used to determine the semantic relationship between two sentences, plays an important role in many tasks such as Natural Language inference (NLI), Machine Comprehension and Paraphrase Identification. In this paper, we propose the multi-layer attention neural network which employs multi-level attention between two sentences. To be specific, we employ two layers of soft-alignment attention-one after embedding layer and the other after encode layer. Then we aggregate the matching information from each attention layer and use both max-pooling and attention-pooling to obtain the final representation and feed it into a MLP classifier. We use four different datasets to evaluate the model and the results shows that our model has a significant performance.