Bi-directional attention comparison for semantic sentence matching

Abstract

Semantic sentence matching, also known as calculation of text similarity, is one of the most important problems in natural language processing. Existing deep models mostly focus on the neural networks with attention mechanism. In this paper, we present a deep architecture to match two Chinese sentences, which only relies on alignment instead of long short-term memory network after attention mechanism is employed to get interaction information between sentence-pairs, the model becomes more lightweight and simple. Meanwhile, in order to capture semantic features enough, in addition to using max pooling and average pooling operation, we also employ a pooling operation named attention-pooling to aggregate information from the whole sentence, the final matching score is obtained after a multilayer perceptron classifier. Experiments are carried out on ATEC-NLP dataset and outline the effectiveness of our approach.