GraphRel: Modeling Text as Relational Graphs for Joint Entity and Relation Extraction

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1409–1418 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1409 GraphRel: Modeling Text as Relational Graphs for Joint Entity and Relation Extraction Tsu-Jui Fu Academia Sinica s103062110@m103.nthu.edu.tw Peng-Hsuan Li Academia Sinica jacobvsdanniel@gmail.com Wei-Yun Ma Academia Sinica ma@iis.sinica.edu.tw Abstract In this paper, we present GraphRel, an end-to-end relation extraction model which uses graph convolutional networks (GCNs) to jointly learn named entities and relations. In contrast to previous baselines, we consider the interaction between named entities and rela- tions via a relation-weighted GCN to better ex- tract relations. Linear and dependency struc- tures are both used to extract both sequential and regional features of the text, and a com- plete word graph is further utilized to extract implicit features among all word pairs of the text. With the graph-based approach, the pre- diction for overlapping relations is substan- tially improved over previous sequential ap- proaches. We evaluate GraphRel on two pub- lic datasets: NYT and WebNLG. Results show that GraphRel maintains high precision while increasing recall substantially. Also, GraphRel outperforms previous work by 3.2% and 5.8% (F1 score), achieving a new state-of-the-art for relation extraction. 1 Introduction Extracting pairs of entity mentions with seman- tic relations, i.e., triplets such as (BarackObama, PresidentOf, UnitedStates), is a central task in in- formation extraction and allows automatic knowl- edge construction from unstructured text. Though important and well-studied, three key aspects are yet to be fully handled in an unified framework: • End-to-end joint modeling of entity recognition and relation extraction; • Prediction of overlapping relations, i.e., rela- tions that share a common mention; • Consideration of the interaction between rela- tions, especially overlapping relations. Traditionally, a pipelined approach is used to first extract entity mentions using a named entity recognizer and then predict the relations between every pair of extracted entity mentions (Zelenko et al., 2003; Zhou et al., 2005; Chan and Roth, 2011). Joint entity recognition and relation ex- traction models (Yu and Lam, 2010; Li and Ji, 2014; Miwa and Sasaki, 2014; Ren et al., 2017) have been built to take advantage of the close in- teraction between these two tasks. While showing the benefits of joint modeling, these complicated methods are feature-based structured learning sys- tems and hence rely heavily on feature engineer- ing. With the success of deep neural networks, NN- based automatic feature learning methods have been applied to relation extraction. These methods use CNN, LSTM, or Tree-LSTM on the word se- quence between two entity mentions (Zeng et al., 2014; dos Santos et al., 2015), the shortest de- pendency paths between two entity mentions (Yan et al., 2015; Li et al., 2015), or the minimal constituency sub-tree spanning two entity men- tions (Socher et al., 2012) to encode relevant infor- mation for each pair of entity mentions. However, these methods are not end-to-end joint modeling of entities and relations. They assume entity men- tions are given and are expected to degrade …

Content truncated to 500 words