A Unified Generative Framework for Various NER Subtasks

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5808–5822 August 1–6, 2021. ©2021 Association for Computational Linguistics 5808 A Unified Generative Framework for Various NER Subtasks Hang Yan1, Tao Gui2, Junqi Dai1, Qipeng Guo1, Zheng Zhang3, Xipeng Qiu1,4∗ 1Shanghai Key Laboratory of Intelligent Information Processing, Fudan University 1School of Computer Science, Fudan University 2Institute of Modern Languages and Linguistics, Fudan University 3New York University 4Pazhou Lab, Guangzhou, China {hyan19,tgui16,jqdai19,qpguo16,xpqiu}@fudan.edu.cn zz@nyu.edu Abstract Named Entity Recognition (NER) is the task of identifying spans that represent entities in sentences. Whether the entity spans are nested or discontinuous, the NER task can be cate- gorized into the flat NER, nested NER, and discontinuous NER subtasks. These subtasks have been mainly solved by the token-level sequence labelling or span-level classification. However, these solutions can hardly tackle the three kinds of NER subtasks concurrently. To that end, we propose to formulate the NER subtasks as an entity span sequence genera- tion task, which can be solved by a unified sequence-to-sequence (Seq2Seq) framework. Based on our unified framework, we can lever- age the pre-trained Seq2Seq model to solve all three kinds of NER subtasks without the special design of the tagging schema or ways to enumerate spans. We exploit three types of entity representations to linearize entities into a sequence. Our proposed framework is easy-to-implement and achieves state-of-the- art (SoTA) or near SoTA performance on eight English NER datasets, including two flat NER datasets, three nested NER datasets, and three discontinuous NER datasets 1. 1 Introduction Named entity recognition (NER) has been a funda- mental task of Natural Language Processing (NLP), and three kinds of NER subtasks have been recog- nized in previous work (Sang and Meulder, 2003; Pradhan et al., 2013a; Doddington et al., 2004; Kim et al., 2003; Karimi et al., 2015), including flat NER, nested NER, and discontinuous NER. As shown in Figure 1, the nested NER contains over- lapping entities, and the entity in the discontinuous NER may contain several nonadjacent spans. ∗Corresponding author. 1Code is available at https://github.com/yhcc/ BARTNER. (d) A unified generative solution for all NER tasks US The Lincoln Memorial the Lincoln Memorial the Lincoln Lincoln Memorial The Lincoln Memorial S2: S1: Barack Obama was born in the Person Location Person Location have much muscle pain and fatigue Disorder Disorder S3: B-Per I-Per O O O O B-Loc (a) Sequence labelling for flat NER (b) Span-based classification for nested NER Barack Obama <Person> US <Location> S1: The Lincoln Memorial <Location> Lincoln <Person> S2: muscle pain < Disorder > muscle fatigue <Disorder> S2: (c) Transition-based method for discontinuous NER OUT OUT SHIFT SHIFT LEFT-REDUCE COMPLETE … Actions: Figure 1: Examples of three kinds of NER subtasks. (a) - (c) illustrate flat NER, nested NER, discontinuous NER, and their corresponding mainstream solutions re- spectively. (d) Our proposed generative solution to solve all NER subtasks in a unified way. The sequence labelling formulation, which will assign a tag …

Content truncated to 500 words