BERT-KRS: A BERT-Based Model for Knowledge-Grounded Response Selection in Retrieval-Based Chatbots

Knowledge-grounded response selection in retrieval-based chatbots aims to select the proper response given the conversation context and background knowledge. Most of existing BERT-based approaches only consider the relations between the conversation context and response candidates, and often ignore the effect of BERT in domain adaptation with task-specific data and fail to consider the relevance between the context and knowledge. To address these issues, we propose a BERT-based model for Knowledge-grounded Response Selection (BERT-KRS). We first design a domain adaptation task, consistency detection, to yield a domain-adaptive BERT for capturing the consistency relation between conversations and knowledge. Then we take the domain-adaptive BERT as the encoder and propose an attention-based matching network to strengthen the relationship between the context and knowledge before matching with response candidates. Experimental results on two canonical benchmarks indicate that our BERT-KRS model outperforms the strong baseline methods.