Span-Level Model for Relation Extraction

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5308–5314 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5308 Span-Level Model for Relation Extraction Kalpit Dixit Amazon AWS AI, USA kddixit@amazon.com Yaser Al-Onaizan Amazon AWS AI, USA onaizan@amazon.com Abstract Relation Extraction is the task of identifying entity mention spans in raw text and then iden- tifying relations between pairs of the entity mentions. Recent approaches for this span- level task have been token-level models which have inherent limitations. They cannot eas- ily define and implement span-level features, cannot model overlapping entity mentions and have cascading errors due to the use of sequen- tial decoding. To address these concerns, we present a model which directly models all pos- sible spans and performs joint entity mention detection and relation extraction. We report a new state-of-the-art performance of 62.83 F1 (prev best was 60.49) on the ACE2005 dataset. 1 Introduction Many NLP tasks follow the pattern of taking raw text as input and then: detecting relevant spans and classifying the relations between those spans. Examples of this include Relation Extraction (Li and Ji, 2014), Coreference Resolution (Ng, 2010) and Semantic Role Labeling (Gildea and Juraf- sky, 2002). This class of NLP problems are in- herently span-level tasks. This paper focuses on Relation Extraction (RE), which is the task of en- tity mention detection and classifying the relations between each pair of those mentions. We report a new state-of-the-art performance of 62.83 F1 (prev best was 60.49) on the ACE2005 dataset. Here is a simple example of Relation Extrac- tion for the sentence, ”Washington, D.C. is the capital of the USA”. Step 1, Entity Mention De- tection will detect the spans ”Washington, D.C.” and ”USA” as LOCATIONS. Step 2, Relation Extraction will classify all directed pairs of de- tected entity mentions. It will classify the directed pair (”Washington, D.C.”, ”USA”) as having the relation IS CAPITAL OF. But the directed pair (”USA”, ”Washington, D.C.”) will be classified as having no relation (NONE). In more complex cases, each entity could participate in multiple dif- ferent relations. Since (Li and Ji, 2014), work on RE has re- volved around end-to-end systems: single models which first perform entity mention detection and then relation extraction. These recent works (Bek- oulis et al., 2018; Katiyar and Cardie, 2017; Miwa and Bansal, 2016; Li and Ji, 2014) have used se- quential token-level methods for both the steps. Token-level models are primarily constrained by the fact that each token has a single fixed repre- sentation while each token is a part of many dif- ferent spans. To model and extract spans, these token-level models have to resort to approximate span-level features which are increasingly indirect and expensive: Tree-LSTMs (Miwa and Bansal, 2016), CRFs (Bekoulis et al., 2018), Beam Search (Li and Ji, 2014) and Pointer Networks (Katiyar and Cardie, 2017). Their usage of the BILOU (Ratinov and Roth, 2009; Florian et al., 2006) token-tagging scheme makes modelling overlap- ping entities impossible. In …

Content truncated to 500 words