糖尿病健康管理对话数据集构建

Dialogue Dataset Construction for Diabetes Health Management

我国是全球糖尿病患病人数最多的国家,患病人数仍在持续快速增长,糖尿病已成为我国重大公共卫生问题。该文关注的糖尿病健康管理对话系统服务于糖尿病患者,为患者解答日常生活中糖尿病相关问题,而目前缺乏用于训练对话系统模型的糖尿病相关数据。基于此,该文构建了首个标注体系完整的糖尿病健康管理中文对话数据集“Diachat”,以支持健康管理对话系统研究。Diachat收集了来自线上聊天平台糖尿病患者与医生的693段对话(Dialogue),共4686句语料(Sentence),完成了6594条对话动作(Dialogue act)标注。Diachat数据集采用基于对话动作的表示方式进行意图表示并定义了15个对话动作标签(Act label)。同时,Diachat定义了6个领域(Domain)涵盖语料涉及的领域,分别为:问题(Problem)、饮食(Diet)、行为(Behavior)、运动(Sport)、治疗(Treatment)、基本信息(Profile)。为了支持构建完整的对话系统,Diachat为用户端和系统端分别构造了对话状态,并为每段对话构造了对话目标。基于Diachat数据集,该课题进行了管道(Pipeline)体系的对话系统四个模块的基本实现。实验结果显示,Diachat数据集能够支持糖尿病健康管理对话系统构建,各模块仍有较大提升空间。

As the country with the largest number of diabetes cases in the world,diabetes has become a major public health problem in China.Onto develope the diabetes health management dialogue system,there is currently a lack of diabetes-related data for training the dialogue model.In this paper,the first Chinese diabetics-doctors dialogue dataset"Diachat"with complete annotation schemas is presented.Diachat consists of 693 conversations between diabetics and doctors from an online chat platform,with a total of 4686 sentences annotated with 6594 annotations The Diachat dataset employs a dialogue act-based representation for intent representation and defines 15 act labels.Meanwhile,Diachat defines 6 domains,including Problem,Diet,Behavior,Sport,Treatment,and user Profile,covering the major fields of the dataset.In order to support the construction of a complete dialogue system,Diachat generates dialogue states for the user side and system side respectively,together with conversation goal for each dialogue.Based on the Diachat dataset,this paper performs a preliminary implementation of the four modules under the pipeline framework.The experimental results showed that the Diachat dataset can support the construction of the diabetes health management dialogue system.