Prediction, selection, and generation: a knowledge-driven conversation system

In conversational systems, we can use external knowledge to generate more diverse sentences and make these sentences contain actual knowledge. Leveraging knowledge for conversation system is important but challenging. Firstly, the conversation system needs to find the appropriate knowledge. Secondly, the knowledge needs to be coded effectively and generated into fluent utterances. In this paper, we propose a knowledge-driven conversation system to address the above challenges. This system consists of three modules, namely topic predictor, knowledge selector, and dialogue generator. The topic predictor uses a combination of non-deep learning (coarse-grained) and deep learning (fine-grained) to form a rough recall and fine sorting process, and uses them to predict conversation topics. The knowledge selector aims to find the most appropriate knowledge based on the filtered topics. We propose the Bert2Transformer model as our dialogue generator, which can generate rich and fluent utterances based on contextual and relevant knowledge. On the public corpus KdConv, our system outperforms a strong baseline and achieves state-of-the-art results. In the ablation study, we analyze the effectiveness of the proposed components in detail and investigate the performance factors that may affect the knowledge-driven conversation generation. Experimental results show the proposed system achieves a significant improvement compared with traditional baseline methods. The average BLEU scores of our system achieve 35.92 and 23.24, respectively, given appropriate knowledge and without appropriate knowledge, while the Distinct-2 scores of our system achieve 16.32 and 15.93, respectively. The training corpus is publicly available ( https://github.com/wulaoshi/dialogue_train_data ).