Supporting Medical Relation Extraction via Causality-Pruned Semantic Dependency Forest

Proceedings of the 29th International Conference on Computational Linguistics, pages 2450–2460 October 12–17, 2022. 2450 Supporting Medical Relation Extraction via Causality-Pruned Semantic Dependency Forest Yifan Jin1,2,∗, Jiangmeng Li1,2,∗,†, Zheng Lian1,2, Chengbo Jiao3, Xiaohui Hu1,2 1University of Chinese Academy of Sciences 2 Institute of Software Chinese Academy of Sciences 3University of Electronic Science and Technology of China {yifan2020,jiangmeng2019,lianzheng2017,hxh}@iscas.ac.cn chengbojiao@hotmail.com Abstract Medical Relation Extraction (MRE) task aims to extract relations between entities in medi- cal texts. Traditional relation extraction meth- ods achieve impressive success by exploring the syntactic information, e.g., dependency tree. However, the quality of the 1-best de- pendency tree for medical texts produced by an out-of-domain parser is relatively limited so that the performance of medical relation extrac- tion method may degenerate. To this end, we propose a method to jointly model semantic and syntactic information from medical texts based on causal explanation theory. We generate de- pendency forests consisting of the semantic- embedded 1-best dependency tree. Then, a task- specific causal explainer is adopted to prune the dependency forests, which are further fed into a designed graph convolutional network to learn the corresponding representation for downstream task. Empirically, the various comparisons on benchmark medical datasets demonstrate the effectiveness of our model. 1 Introduction Medical relation extraction (MRE) refers to iden- tifying relations among entities from medical lit- erature and reports. It plays a very important role in downstream tasks such as medical knowledge graph construction (Li et al., 2020; Rotmensch et al., 2017) and biomedical knowledge discov- ery (Quirk and Poon, 2016). On the other hand, as the number of medical literature increases, it becomes increasingly important to automatically discover the relation among entities in the litera- ture (Peng et al., 2017). The addition of syntactic structure has been demonstrated to be beneficial for various natural language processing tasks (Zaremoodi and Haf- fari, 2017; Zhou et al., 2005; Le and Zuidema, ∗Equal contribution. †Corresponding author. 2015). As a type of syntactic structure, the de- pendency tree capturing long-distance connections between words can indeed improve benchmark re- lation extraction methods (Tian et al., 2021; Chen et al., 2021; Zhang et al., 2018; Sun et al., 2020). We demonstrate an example in Figure 1. Specifi- cally, the 1-best dependency tree of the sentence “Aminopropylindenes derived from Grundmann’s ketone as a novel chemotype of oxidosqualene cyclase inhibitors” in the CPR dataset. Amino- propylindenes and oxidosqualene cyclase are the entities, and the relation between them is “down regulator”, denoted as “CPR:4”. However, in the medical field, the quality of the 1-best dependency tree generated by the out-of- domain parsers, e.g., parsers for the news domain, is relatively deficient. Generally, the main verb in a sentence is treated as the root node in the de- pendency tree, while, as the example shown in Figure 1, the entity Aminopropylindenes, appar- ently a noun, is treated as the root node. To solve this problem, multiple methods with dependency forests have been proposed (Song et al., 2019; Jin et al., 2020; Guo et al., 2021). Such approaches …

Content truncated to 500 words