Towards Medical Machine Reading Comprehension with Structural Knowledge and Plain Text

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1427–1438, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1427 Towards Medical Machine Reading Comprehension with Structural Knowledge and Plain Text Dongfang Li1, Baotian Hu1, Qingcai Chen1,2,⇤⇤,Weihua Peng3,⇤, Anqi Wang1 1Harbin Institute of Technology (Shenzhen), Shenzhen, China 2Peng Cheng Laboratory, Shenzhen, China 3Baidu, International Technology (Shenzhen) Co., Ltd. crazyofapple@gmail.com, {hubaotian, qingcai.chen}@hit.edu.cn pengweihua@baidu.com, 19s051040@stu.hit.edu.cn Abstract Machine reading comprehension (MRC) has achieved significant progress on the open do- main in recent years, mainly due to large-scale pre-trained language models. However, it per- forms much worse in specific domains such as the medical field due to the lack of exten- sive training data and professional structural knowledge neglect. As an effort, we first col- lect a large scale medical multi-choice ques- tion dataset (more than 21k instances) for the National Licensed Pharmacist Examination in China. It is a challenging medical examina- tion with a passing rate of less than 14.2% in 2018. Then we propose a novel reading comprehension model KMQA, which can fully exploit the structural medical knowledge (i.e., medical knowledge graph) and the reference medical plain text (i.e., text snippets retrieved from reference books). The experimental re- sults indicate that the KMQA outperforms ex- isting competitive models with a large margin and passes the exam with 61.8% accuracy rate on the test set. 1 Introduction With the advent of large scale datasets such as SQuAD (Rajpurkar et al., 2016, 2018), RACE (Lai et al., 2017), and Natural Questions (Kwiatkowski et al., 2019; Lee et al., 2019) on the open domain, machine reading comprehension (MRC) has be- come a hot topic in the natural language process- ing field. In the past few years, the MRC has ob- tained substantial progress, and many recent mod- els have surpassed the human performance on sev- eral datasets. The superiority of these models is mainly attributed to two significant aspects: 1) the powerful representations ability of large pre-trained language models (PLMs), which can cover or re- member most of the language variations implicitly. ⇤ Co-corresponding authors Question: £⇧，s，27Å，n b b b' ' 'Y Y Yã ã ãù ù ùé é é3t， —Â化å”ú⇢HBV-DNA 2 ⇥ 105 copies/mL, ALT 122 U/L⇥flàÂó ó ó≈ ≈ ≈“ “ “ª ª ªó ó ó，ñ Ñoi/Í*？ A female patient, aged 27 years old, has been diagnosed with chronic hepatitis B for 3 years. Recent results show: HBV- DNA 2 ⇥ 105 copies/mL, ALT 122 U/L. The initial diagnosis is to take antiviral treatment for her. Which is the preferred one among the following drugs? Options: A. ?÷z˜ Ara adenosine. B. iˇaÊ Entecavir. X C. €洛Ê Famciclovir. D. )ÙÊó Ribavirin. E. ¶2x† Sodium foscarnet. Option B retrieved text snippets: 4ä(éóYãùé≈“Ñoi …s+ö, ?∑èÊ, rp -↵, )ÙÊó, iˇaÊI... Drugs used clinically against hepatitis B virus include lamivu- dine, adefovir, interferon-↵, ribavirin, entecavir, ... Option B knowledge facts: (iˇaÊ, ⇥î«, b'Yãùé) (entecavir, indication, chronic hepatitis B) (iˇaÊ, åß⌃{, ó≈“o) (entecavir, second class, antiviral drugs) Table 1: An example from our multiple-choice QA task …

Content truncated to 500 words