开放域对话系统中的角色化对话生成技术研究

开放域对话系统不限定人机对话的话题范围，更加接近现实生活中的人人对话形态，能够应用在众多场景中，扮演不同的角色，如景区智能讲解员、商场智能导购员、情感陪护伴侣、儿童读书伴侣等，因此，开放域对话系统的角色化对话能力是衡量其智能化程度的重要依据。然而，开放域对话系统在角色化对话生成方面仍然面临着角色信息选择和表示能力较弱及生成的角色化回复多样性较差、生成的多轮对话中角色信息不一致以及角色迁移依赖大规模人工标注的角色对话数据等问题。因此，本项目主要关注开放域角色化对话生成模型对角色信息文本的建模、开放域多轮对话生成过程中的角色信息一致性以及角色对话数据资源受限条件下的开放域对话系统的角色迁移这三个关键科学问题，相应地，分别提出三个研究方案，以提升开放域对话系统的角色化对话能力，从而推进人类与机器的有效沟通和自由交互，为建设人机共融社会提供基础的技术保障。

Open domain dialogue systems have unlimited scope of topics in human-machine conversations, which are more approximated to human-human conversations. They can be applied to many scenarios and play different roles, such as an intelligent interpreter in scenic spot, an intelligent shopping guide in shopping mall, an emotional companion, a reading partner for children, etc. Therefore, the ability of persona-based conversing is an important basis for measuring the intelligence of open domain dialogue systems. However, there are still challenges in persona-based dialogue generation including the weakness on persona selection and representation and the lack of diversity of generated responses, the inconsistency of personas in generating multi-turn dialogues, and the dependence on large-scale manually labeled data of persona-based dialogues for persona transfer. Therefore, in this project, we mainly focus on three key scientific problems. First is the persona modeling on open domain persona-based dialogue generation. Second is the persona consistency on open domain multi-turn dialogue generation. Third is the persona transfer on low-resource persona-based dialogue data. Accordingly, we propose three research plans to improve the performance and ability of persona-based dialogue generation for open domain dialogue systems so that to advance the efficient communication and free interaction between human and machines, and provide a basic technical support for the construction of human-machine integration society.