Multi-level Spatio-temporal Matching Network for Multi-turn Response Selection in Retrieval-based Dialogue Systems

The important task of multi-turn response selection in conversation systems must consider sufficient semantic information and spatio-temporal information when building retrieval-based chatbots. However, existing studies do not pay enough attention to both factors. In this study, a scheme of multi-turn response selection that combines a primary temporal matching module, an advanced temporal matching module, and a spatial matching module is proposed to extract matching information from context and response. The temporal matching modules progressively construct representations of the context and candidate responses at different granularities. Similarity matrices of the context and candidate responses are calculated and stacked using the spatial matching module. Convolutional neural network is then utilized to extract the spatial matching information. Finally, matching vectors of the three modules are fused to calculate the final matching score. Experimental results on two public datasets verify that our model can outperform state-of-the-art methods.