Context-Aware Network for Multi-Turn Response Selection in Retrieval-Based Chatbots

Multi-turn response selection is a major challenge for chatbot dialogue systems. The existing methods either ignore the interactions among previous utterances for context modeling, or regard all the previous utterances of the same importance. In this paper, we propose a context-aware network (CAN) for the multi-turn response selection task. The main idea of CAN is to learn the unified representation vector of the utterance and response for response matching. First, CAN uses a hierarchical attention mechanism to extract important features of utterances for modeling the utterance representation . Then CAN separates the current message from the historical dialogues, and calculates the matching degree between the current message and each historical dialogue to get the utterances aggregation representation. Finally, CAN uses multi-layer perceptron (MLP) to predicate the score of each response. Experiments on three public datasets show that CAN outperforms the baseline models on most metrics and demonstrates compatibility across domains for multi-turn response selection.