Document-level Biomedical Relation Extraction Based on Multi-Dimensional Fusion Information and Multi-Granularity Logical Reasoning

Proceedings of the 29th International Conference on Computational Linguistics, pages 2098–2107 October 12–17, 2022. 2098 Document-level Biomedical Relation Extraction Based on Multi-Dimensional Fusion Information and Multi-Granularity Logical Reasoning Lishuang Li ∗and Ruiyuan Lian † and Hongbin Lu and Jingyao Tang School of Computer Science and Technology Dalian University of Technology Dalian, China lils@dlut.edu.cn Abstract Document-level biomedical relation extrac- tion (Bio-DocuRE) is an important branch of biomedical text mining that aims to automati- cally extract all relation facts from the biomedi- cal text. Since there are a considerable number of relations in biomedical documents that need to be judged by other existing relations, logi- cal reasoning has become a research hotspot in the past two years. However, current mod- els with reasoning are single-granularity only based on one element information, ignoring the complementary fact of different granular- ity reasoning information. In addition, obtain- ing rich document information is a prerequisite for logical reasoning, but most of the previ- ous models cannot sufficiently utilize document information, which limits the reasoning abil- ity of the model. In this paper, we propose a novel Bio-DocuRE model called FILR, based on Multi-Dimensional Fusion Information and Multi-Granularity Logical Reasoning. Specif- ically, FILR presents a multi-dimensional in- formation fusion module MDIF to extract suffi- cient global document information. Then FILR proposes a multi-granularity reasoning mod- ule MGLR to obtain rich inference information through the reasoning of both entity-pairs and mention-pairs. We evaluate our FILR model on two widely used biomedical corpora CDR and GDA 1. Experimental results show that FILR achieves state-of-the-art performance. 1 Introduction Biomedical relation extraction(Bio-RE) is an im- portant branch of biomedical text mining, which always draws researchers’ attention. According to the different length of processing text, Bio-RE is generally divided into sentence-level RE and document-level RE. In sentence-level RE, all in- volved entities and relations are within a sentence. ∗Corresponding author work.†The author contributes equally to the first author in this 1https://github.com/Luguo-ry/FILR [1] Acute vocal fold palsy after acute disulfiram intoxication . [2] Acute peripheral neuropathy caused by a disulfiram overdose … of it leading to vocal fold palsy . [3] A woman was … because of quadriparesis , lancinating pain , sensory loss, and paresthesia of the limbs . [4] … high dose of disulfiram in a suicide attempt . …. [6] … had mild to moderate ataxia and giddiness . …. [12] … caused by high - dose disulfiram intoxication . intra-sentence relations: <disulfiram, vocal fold palsy>; <disulfiram, peripheral neuropathy> inter-sentence relations: <disulfiram, quadriparesis>; <disulfiram, pain>; <disulfiram, paresthesia>; <disulfiram, ataxia> Figure 1: An example of Bio-DocuRE from the CDR dataset. We use the same color to denote the mentions of the same entity. Document-level RE is more complex than its sen- tence counterpart. As shown in Fig. 1, in document- level biomedical relation extraction(Bio-DocuRE), the processed text is a whole document containing multiple sentences. “disulfiram” is a chemical and other entities are diseases. Each entity always has more than one mention, for example, “disulfiram” has four mentions in the document. In …

Content truncated to 500 words