RI-Match: Integrating Both Representations and Interactions for Deep Semantic Matching

Existing deep matching methods can be mainly categorized into two kinds, i.e. representation focused methods and interaction focused methods. Representation focused methods usually focus on learning the representation of each sentence, while interaction focused methods typically aim to obtain the representations of different interaction signals. However, both sentence level representations and interaction signals are important for the complex semantic matching tasks. Therefore, in this paper, we propose a new deep learning architecture to combine the merits of both deep matching approaches. Firstly, two kinds of word level matching matrices are constructed based on word identities and word embeddings, to capture both exact and semantic matching signals. Secondly, a sentence level matching matrix is constructed, with each element stands for the interaction between two sentence representations at corresponding positions, generated by a bidirectional long short term memory (Bi-LSTM). In this way, sentence level representations are well captured in the matching process. The above matrices are then fed into a spatial recurrent neural network (RNN), to generate the high level interaction representations. Finally, the matching score is produced by a k-Max pooling and a multilayer perceptron (MLP). Experiments on paraphrasing identification shows that our model outperforms traditional state-of-the art baselines significantly.