Dual Memory Network for Medical Dialogue Generation

Task-oriented dialogue systems have achieved re-markable success recently. However, the generation quality of medical dialogue systems is not ideal since medical conversations tend to be more complex than other conversations, placing higher demands on text comprehension and knowledge-based reasoning. To address this challenge, we propose DM-Net, which consists of two major memory modules. The encoding of dialogue historical content is utilized as the query to query dialogue context memory and clinical experience memory to achieve a correct understanding of dialogue content and sensible reasoning of external knowledge, thereby generating accurate and reliable responses. Extensive experiments on two large-scale medical dialogue datasets demonstrate that DM-Net outperforms a large number of strong baselines in terms of objective and subjective evaluation metrics.