GraphIE: A Graph-Based Framework for Information Extraction

Proceedings of NAACL-HLT 2019, pages 751–761 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 751 GraphIE: A Graph-Based Framework for Information Extraction Yujie Qian1, Enrico Santus1, Zhijing Jin2, Jiang Guo1, and Regina Barzilay1 1Computer Science and Artificial Intelligence Laboratory, MIT 2Department of Computer Science, The University of Hong Kong {yujieq, jiang guo, regina}@csail.mit.edu, {esantus, zhijing}@mit.edu Abstract Most modern Information Extraction (IE) sys- tems are implemented as sequential taggers and only model local dependencies. Non-local and non-sequential context is, however, a valu- able source of information to improve predic- tions. In this paper, we introduce GraphIE, a framework that operates over a graph repre- senting a broad set of dependencies between textual units (i.e. words or sentences). The al- gorithm propagates information between con- nected nodes through graph convolutions, gen- erating a richer representation that can be exploited to improve word-level predictions. Evaluation on three different tasks — namely textual, social media and visual information extraction — shows that GraphIE consistently outperforms the state-of-the-art sequence tag- ging model by a significant margin.1 1 Introduction Most modern Information Extraction (IE) systems are implemented as sequential taggers. While such models effectively capture relations in the local context, they have limited capability of exploit- ing non-local and non-sequential dependencies. In many applications, however, such dependencies can greatly reduce tagging ambiguity, thereby im- proving overall extraction performance. For in- stance, when extracting entities from a document, various types of non-local contextual information such as co-references and identical mentions may provide valuable cues. See for example Figure 1, in which the non-local relations are crucial to dis- criminate the entity type of the second mention of Washington (i.e. PERSON or LOCATION). Most of the prior work looking at the non-local dependencies incorporates them by constraining 1Our code and data are available at https://github. com/thomas0809/GraphIE. … his father . Washington came from a prosperous family of planters … Jefferson was invited by the president to serve as Secretary of State … Text Graph He was involved in a wide range of duties for Washington ’s request … … his father. Washington came from a prosperous family of planters … Jefferson was invited by Washington to serve as Secretary of State … He was involved in a wide range of duties for Washington’s request … local Figure 1: Example of the entity extraction task with an ambiguous entity mention (i.e. “...for Washing- ton’s request...”). Aside from the sentential forward and backward edges (green, solid) which aggregate lo- cal contextual information, non-local relations — such as the co-referent edges (red, dashed) and the identical- mention edges (blue, dotted) — provide additional valuable information to reduce tagging ambiguity (i.e. PERSON or LOCATION). Best viewed in color. the output space in a structured prediction frame- work (Finkel et al., 2005; Reichart and Barzilay, 2012; Hu et al., 2016). Such approaches, how- ever, mostly overlook the richer set of structural relations in the input space. With reference to the example in Figure 1, the co-referent depen- dencies …

Content truncated to 500 words