Hierarchical Representation-based Dynamic Reasoning Network for Biomedical Question Answering

Proceedings of the 29th International Conference on Computational Linguistics, pages 1480–1489 October 12–17, 2022. 1480 Hierarchical Representation-based Dynamic Reasoning Network for Biomedical Question Answering Jianguo Mao1,2†‡ Jiyuan Zhang3† Zengfeng Zeng3, Weihua Peng3, Wenbin Jiang3∗ Xiangdong Wang1, Hong Liu1, Yajuan Lyu3 1Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2 University of Chinese Academy of Sciences, Beijing, China 3 Baidu Inc., Beijing, China {maojianguo20s,xdwang,hliu}@ict.ac.cn {zhangjiyuan01,zengzengfeng,jiangwenbin,lvyajuan}@baidu.com Abstract Recently, Biomedical Question Answering (BioQA) has attracted growing attention due to its application value and technical challenges. Most existing works treat it as a semantic matching task that predicts answers by computing confidence among questions, options and evidence paragraphs, which is insufficient for scenarios that require complex reasoning based on a deep understanding of biomedical evidences. We propose a novel model termed Hierarchical Representation-based Dynamic Reasoning Network (HDRN) to tackle this problem. It first constructs the hierarchical representations for biomedical evidences to learn semantics within and among evidences. It then performs dynamic reasoning based on the hierarchical representations of evidences to solve complex biomedical problems. Against the existing state-of-the-art model, the proposed model significantly improves more than 4.5%, 3% and 1.3% on three mainstream BioQA datasets, PubMedQA, MedQA-USMLE and NLPEC. The ablation study demonstrates the superiority of each improvement of our model. https://github.com/mikeblueskydl/HDRN 1 Introduction Machine reading comprehension (MRC) (Ra- jpurkar et al., 2016, 2018; Yang et al., 2018) tasks are often used to evaluate the intelligence degree of a system, and many recent large-scale pre-trained language models (Lan et al., 2019; Zaheer et al., 2020) have surpassed the human performance on open-domain MRC. In recent years, Biomedical Question Answering (BioQA) (Tsatsaronis et al., 2012; Wang et al., 2018; Tang et al., 2019; Li et al., *Corresponding author. †Equal contribution. ‡This work was done during internship at Baidu Inc. Question：A 35-year-old Caucasian female presents to the hospital alarmed by her recent truncal weight gain, facial hair growth, and thinning skin. During the physical exam, the physician finds that the patient is hypertensive. Serum analysis reveals hyperglycemia. The physician suspects a pituitary adenoma. Which dexamethasone test result would help confirm the physician's suspicions? Options： A: Low-dose, increased ACTH; high-dose, decreased ACTH B: Low-dose, decrease in ACTH; high-dose, no change in ACTH C: Low-dose, no change in ACTH; high-dose, no change in ACTH D: Low-dose, no change in ACTH; high-dose, decreased ACTH E1: If cortisol production is driven by an ACTH producing pituitary adenoma dexamethasone suppression is ineffective at low doses but usually induces suppression at high doses. Inappropriately low ACTH levels in the setting of low cortisol levels are characteristic of diminished ACTH reserve. E2: High dose dexamethasone suppresses ACTH production by a pituitary adenoma serum cortisol is lowered but does not suppress ectopic ACTH production serum cortisol remains high. Cortisol stimulates gluconeogenesis and insulin resistance resulting in hyperglycemia as well as muscle cell protein breakdown and lipolysis to provide sub strates for hepatic gluconeogenesis. E3: The mechanism of hypertension may be related to stimulation …

Content truncated to 500 words