SIRE: Separate Intra- and Inter-sentential Reasoning for Document-level Relation Extraction

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 524–534 August 1–6, 2021. ©2021 Association for Computational Linguistics 524 SIRE: Separate Intra- and Inter-sentential Reasoning for Document-level Relation Extraction Shuang Zeng1,3, Yuting Wu1,2 and Baobao Chang1∗ 1The MOE Key Laboratory of Computational Linguistics, Peking University, China 2Wangxuan Institute of Computer Technology, Peking University, China 3School of Software and Microelectronics, Peking University, China {zengs,wyting,chbb}@pku.edu.cn Abstract Document-level relation extraction has at- tracted much attention in recent years. It is usu- ally formulated as a classification problem that predicts relations for all entity pairs in the doc- ument. However, previous works indiscrimi- nately represent intra- and inter-sentential re- lations in the same way, confounding the dif- ferent patterns for predicting them. Besides, they create a document graph and use paths between entities on the graph as clues for log- ical reasoning. However, not all entity pairs can be connected with a path and have the correct logical reasoning paths in their graph. Thus many cases of logical reasoning cannot be covered. This paper proposes an effec- tive architecture, SIRE, to represent intra- and inter-sentential relations in different ways. We design a new and straightforward form of logi- cal reasoning module that can cover more logi- cal reasoning chains. Experiments on the pub- lic datasets show SIRE outperforms the pre- vious state-of-the-art methods. Further analy- sis shows that our predictions are reliable and explainable. Our code is available at https: //github.com/PKUnlp-icler/SIRE. 1 Introduction Relation Extraction (RE) is an important way of obtaining knowledge facts from natural language text. Many recent advancements (Sahu et al., 2019; Christopoulou et al., 2019; Yao et al., 2019b; Nan et al., 2020; Zeng et al., 2020; Wang et al., 2020) manage to tackle the document-level relation ex- traction (doc-level RE) that extracts semantic re- lations among entities across multiple sentences. Due to its strong correlation with real-world sce- narios, doc-level RE has attracted much attention in the field of information extraction. The doc-level RE task is usually formulated as a classification problem that predicts possible rela- ∗Corresponding author. ABBA Live [1] ABBA Live is an album of live recordings by Swedish pop group ABBA, released by Polar Music in 1986. … [6] The tracks were mostly taken from ABBA’s concerts at Wembley Arena in London in November 1979. … [13] It was remastered … Head: Polar Music Tail: Swedish relation: country of origin evidence: [1] Head: Wembley Arena Tail: London relation: located in evidence: [6] IBM Research–Brazil [1] IBM Research–Brazil is one of twelve research laboratories comprising IBM Research, its first in South America. … [2] It was established in June 2010, with locations in São Paulo and Rio de Janeiro. … [5] In collaboration with Brazil’s government, it will help IBM … [6] … IBM has 4 priority areas in Brazil … South America São Paulo Brazil country [1, 2, 5] continent [1, 5] has part [1, 5] continent [1, 2, 5] Figure 1: Two examples from DocRED (Yao et al., 2019b) for illustration of intra- and …

Content truncated to 500 words