A Speaker-Aware Co-Attention Framework for Medical Dialogue Information Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4777 - 4786 December 7-11, 2022 ©2022 Association for Computational Linguistics A Speaker-Aware Co-Attention Framework for Medical Dialogue Information Extraction Yuan Xia1†, Zhenhui Shi1†, Jingbo Zhou1∗, Jiayu Xu1, Chao Lu1, Yehui Yang1, Lei Wang1, Haifeng Huang1, Xia Zhang2, Junwei Liu1 1Baidu Inc., China. 2Neusoft Corporation, China. 1{xiayuan,shizhenhui,zhoujingbo,xujiayu03,luchao,yangyehui01, wanglei15, huanghaifeng, liujunwei}@baidu.com, 2zhangx@neusoft.com Abstract With the development of medical digitization, the extraction and structuring of Electronic Medical Records (EMRs) have become chal- lenging but fundamental tasks. How to ac- curately and automatically extract structured information from medical dialogues is espe- cially difficult because the information needs to be inferred from complex interactions be- tween the doctor and the patient. To this end, in this paper, we propose a speaker-aware co- attention framework for medical dialogue in- formation extraction. To better utilize the pre- trained language representation model to per- ceive the semantics of the utterance and the candidate item, we develop a speaker-aware di- alogue encoder with multi-task learning, which considers the speaker’s identity into account. To deal with complex interactions between dif- ferent utterances and the correlations between utterances and candidate items, we propose a co-attention fusion network to aggregate the utterance information. We evaluate our frame- work on the public medical dialogue extraction datasets to demonstrate the superiority of our method, which can outperform the state-of-the- art methods by a large margin. 1 Introduction In the past decade, the collection and usage of Elec- tronic Medical Records (EMRs) have been proved as one of the most important applications in the pro- cess of medical digitization. However, the record- ing and writing of the EMRs may bring a signif- icant burden to doctors. Given the breakthrough advance of speech recognition technology, conver- sations between doctors and patients can be accu- rately recorded as text. However, such unstructured medical dialogue data cannot be easily utilized for medical research. How to automatically extract the structured information from these unstructured †Equal contribution. ∗Corresponding author. Patient:Doctor,-could-you-please-tell-me-is-it-a- coronary-disease? Doctor:Did-you-feel-angina? Patient:No,-I-felt-there-is-a-sense-of-suppression- in-the-chest. Doctor:The-result-of-echocardiography-is-normal.- Don’t-worry. Patient:OK,-thanks,-doctor. Dialogue-Window- Coronary-disease Patient-unk Annotation-Labels- Utterance Candidate-Item Stauts Angina Patient-neg Chest-tightness Patient-pos Echocardiography Doctor-pos Figure 1: An example of a patient-doctor dialogue and the corresponding annotated labels. textual medical dialogue data is an essential step to accelerate medical digitization. Compared with the general medical information extraction, the crucial challenge of the medical dia- logue extraction is that it has to take the speaker’s identity and utterance interactions into consider- ation. In conventional information extraction, a relation can largely be inferred by a sentence or a paragraph. However, in the medical dialogue extraction task, the candidate item and status infor- mation need to be detected and then verified by the conversations between the doctor and the patient. An example of a patient-doctor dialogue and the corresponding annotated labels is shown in Figure 1. For instance, the doctor asks the patient, “Did 4777 you feel angina?”, the patient responds, “No, I felt there is a sense of suppression in the chest.”, the …

Content truncated to 500 words