Despite the great advances in generative dialogue systems, existing dialogue generation models are still unsatisfactory in maintaining persona consistency. In order to make the dialogue generation model generate more ...Despite the great advances in generative dialogue systems, existing dialogue generation models are still unsatisfactory in maintaining persona consistency. In order to make the dialogue generation model generate more persona-consistent responses, this paper proposes a model named BERT-HCM (Personalized Dialogue Generation Model Based on BERT and Hierarchical Copy Mechanism). The model uses an encoder based on BERT initialization to encode persona information and dialogue queries and subsequently uses a Transformer decoder incorporating a hierarchical copy mechanism to dynamically copy the input-side content to guide the model in generating responses. The experimental results show that the proposed model improves on both automatic and human evaluation metrics compared to the baseline model and is able to generate more fluent, relevant and persona-consistent responses.展开更多
文摘Despite the great advances in generative dialogue systems, existing dialogue generation models are still unsatisfactory in maintaining persona consistency. In order to make the dialogue generation model generate more persona-consistent responses, this paper proposes a model named BERT-HCM (Personalized Dialogue Generation Model Based on BERT and Hierarchical Copy Mechanism). The model uses an encoder based on BERT initialization to encode persona information and dialogue queries and subsequently uses a Transformer decoder incorporating a hierarchical copy mechanism to dynamically copy the input-side content to guide the model in generating responses. The experimental results show that the proposed model improves on both automatic and human evaluation metrics compared to the baseline model and is able to generate more fluent, relevant and persona-consistent responses.
