Focus-Driven Contrastive Learning for Medical Question Summarization

Proceedings of the 29th International Conference on Computational Linguistics, pages 6176–6186 October 12–17, 2022. 6176 Focus-Driven Contrastive Learning for Medical Question Summarization Ming Zhang1,2 Shuai Dou3 Ziyang Wang1,2 Yunfang Wu1,3∗ 1MOE, Key Laboratory of Computational Linguistics, Peking University 2School of Software and Microelectronics, Peking University 3School of Computer Science, Peking University 1{zhangming,wzy232303}@stu.pku.edu.cn 2{monkdou,wuyf}@pku.edu.cn Abstract Automatic medical question summarization can significantly help the system to understand consumer health questions and retrieve correct answers. The Seq2Seq model based on max- imum likelihood estimation (MLE) has been applied in this task, which faces two general problems: the model can not capture well ques- tion focus and and the traditional MLE strategy lacks the ability to understand sentence-level semantics. To alleviate these problems, we pro- pose a novel question focus-driven contrastive learning framework (QFCL). Specially, we pro- pose an easy and effective approach to generate hard negative samples based on the question focus, and exploit contrastive learning at both encoder and decoder to obtain better sentence- level representations. On three medical bench- mark datasets, our proposed model achieves new state-of-the-art results, and obtains a per- formance gain of 5.33, 12.85 and 3.81 points over the baseline BART model on three datasets respectively. Further human judgement and detailed analysis prove that our QFCL model learns better sentence representations with the ability to distinguish different sentence mean- ings, and generates high-quality summaries by capturing question focus. 1 Introduction A growing number of health questions are raised by consumers on websites nowadays, which are usu- ally written in natural language and including de- tailed and peripheral information not related to the answers. Summaries of such questions can greatly improve the performance in retrieving relevant an- swers (Ben Abacha and Demner-Fushman, 2019). Accordingly, the medical question summarization task is defined as summarizing the consumer health questions (CHQ) into frequently asked questions (FAQ), which are shorter but remain essential in- formation of the original question to get correct ∗*Corresponding author. Input question: consumer health question (CHQ) subject: gender dysphoria message: no health care on my son suffering from gender dysphoria what can we do to help him he worked out of high school no problems now not working and about shutting himself in his room 24/7 theres nothing this condition in our area we live in [location].no help in area what can we do he has had bad thoughts already please help us with some sort of info thank yuo [name] [location] Golden summary: frequently asked question (FAQ): Where can I find information on treatment and resources for gender dysphoria? Summary by BART (baseline): What are the treatments for weight loss? Summary by our model: What are the treatments for gender dysphoria? Table 1: An example of medical question summariza- tion in MeqSum dataset, where the question focus is highlighted in green. Summaries generated by BART and our model are also listed. answers. An example of medical question summa- rization is shown in Table 1. The Seq2Seq neural models have been widely used in abstractive summarization (Nallapati et al., …

Content truncated to 500 words