MidMed: Towards Mixed-Type Dialogues for Medical Consultation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8145–8157 July 9-14, 2023 ©2023 Association for Computational Linguistics MidMed: Towards Mixed-Type Dialogues for Medical Consultation Xiaoming Shi1∗, Zeming Liu2∗, Chuan Wang3, Haitao Leng4, Kui Xue1, Xiaofan Zhang1, Shaoting Zhang1† 1 Shanghai Artificical Intelligence Laboratory, Shanghai, China 2 Research Center for Social Computing and Information Retrieval, HIT, Harbin, China 3 State Key Laboratory of Information Security, IIE, CAS, Beijing, China 4 MMU KuaiShou Inc., Hangzhou, China {shixiaoming, xuekui, zhangxiaofan, shaotingzhang}@pjlab.org.cn zmliu@ir.hit.edu.cn; wangchuan@iie.ac.cn; lenghaitao@kuaishou.com Abstract Most medical dialogue systems assume that patients have clear goals (medicine query- ing, surgical operation querying, etc.) be- fore medical consultation. However, in many real scenarios, due to the lack of medical knowledge, it is usually difficult for patients to determine clear goals with all necessary slots. In this paper, we identify this chal- lenge as how to construct medical consulta- tion dialogue systems to help patients clarify their goals. To mitigate this challenge, we propose a novel task and create a human-to- human mixed-type medical consultation dia- logue corpus, termed MidMed 1, covering five dialogue types: task-oriented dialogue for diag- nosis, recommendation, knowledge-grounded dialogue, QA, and chitchat. MidMed covers four departments (otorhinolaryngology, oph- thalmology, skin, and digestive system), with 8,175 dialogues. Furthermore, we build base- lines on MidMed and propose an instruction- guiding medical dialogue generation frame- work, termed InsMed, to address this task. Ex- perimental results show the effectiveness of InsMed. 1 Introduction Current medical dialogue systems (Xu et al., 2019; Liao et al., 2020; Zeng et al., 2020; Liu et al., 2022a) mainly focus on diagnosis by obtaining symptoms and then making diagnosis automati- cally. These dialogue systems have shown signifi- cant potential and alluring technological value to simplify diagnostic procedures (Semigran et al., 2015). Previous works assume that patients have explicit goals (medicine querying, surgical opera- tion querying, etc.), and perform in the way of task- oriented dialogue to accomplish patients’ goals. ∗ Equal contribution. † Corresponding author: Shaoting Zhang. 1MidMed is publicly available at https://github.com/xmshi- trio/MidMed However, explicit patient goals are usually un- available in real-world scenarios. For example, a patient wants to consult about his itchy skin but lacks medical knowledge. Thus, it is difficult for the patient to decide which slots (e.g. medicine or a surgical operation) are needed. To figure out ex- plicit patient goals, medical consultation services are needed, which provide advice of treatment, medicine, food, etc., as shown in Figure 1. How- ever, those medical consultation services are under explored in previous works. To facilitate the study of medical consul- tation, we construct a new human-to-human mixed-type dialogue dataset for medical consul- tation (MidMed), covering five dialogue types: task-oriented dialogue for diagnosis, knowledge- grounded dialogue, QA, recommendation, and chitchat. MidMed is constructed by revising di- alogues of MedDialog (a human-to-human medical diagnosis dialogue dataset) (Zeng et al., 2020). As shown in Figure 1, a patient queries about “sweaty hands”, and has no explicit goal for medicine or a surgical …

Content truncated to 500 words