Combining Spans into Entities: A Neural Two-Stage Approach for Recognizing Discontiguous Entities

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6216–6224, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6216 Combining Spans into Entities: A Neural Two-Stage Approach for Recognizing Discontiguous Entities Bailin Wang ILCC, School of Informatics University of Edinburgh bailin.wang@ed.ac.uk Wei Lu StatNLP Research Group Singapore University of Technology and Design luwei@sutd.edu.sg Abstract In medical documents, it is possible that an entity of interest not only contains a discon- tiguous sequence of words but also overlaps with another entity. Entities of such structures are intrinsically hard to recognize due to the large space of possible entity combinations. In this work, we propose a neural two-stage ap- proach to recognizing discontiguous and over- lapping entities by decomposing this problem into two subtasks: 1) it first detects all the overlapping spans that either form entities on their own or present as segments of discon- tiguous entities, based on the representation of segmental hypergraph, 2) next it learns to com- bine these segments into discontiguous entities with a classifier, which filters out other incor- rect combinations of segments. Two neural components are designed for these subtasks re- spectively and they are learned jointly using a shared encoder for text. Our model achieves the state-of-the-art performance in a standard dataset, even in the absence of external fea- tures that previous methods used. 1 Introduction Named entity recognition (NER) aims at identi- fying shallow semantic elements in text and has been a crucial step towards natural language un- derstanding (Tjong Kim Sang and De Meulder, 2003). Extracted entities can facilitate various downstream tasks like question answering (Ab- ney et al., 2000), relation extraction (Mintz et al., 2009; Liu et al., 2017), event extraction (Riedel and McCallum, 2011; Lu and Roth, 2012; Li et al., 2013), and coreference resolution (Soon et al., 2001; Ng and Cardie, 2002; Chang et al., 2013). The underlying assumptions behind most NER systems are that an entity should contain a con- tiguous sequence of words and should not overlap with each other. However, such assumptions do Figure 1: Entities are highlighted with colored under- lines. “laceration ... esophagus” and “stomach ... lac” contain discontiguous sequence of words and the latter also overlaps with another entity “blood in stomach”. not always hold in practice. First, entities or men- tions1 with overlapping structures frequently exist in news (Doddington et al., 2004) and biomedi- cal documents (Kim et al., 2003). Second, entities can be discontiguous, especially in clinical texts (Pradhan et al., 2014b). For example, Figure 3 shows three entities where two of them are discon- tiguous (“laceration ...esophagus” and “stomach ...lac”), and the second discontiguous entity also overlaps with another entity (“blood in stomach”). Such discontiguous entities are intrinsically hard to recognize considering the large search space of possible combinations of entities that have discontiguous and overlapping structures. Muis and Lu (2016a) proposed a hypergraph- based representation to compactly encode discon- tiguous entities. However, this …

Content truncated to 500 words