Bottom-Up Constituency Parsing and Nested Named Entity Recognition with Pointer Networks

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2403 - 2416 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Bottom-Up Constituency Parsing and Nested Named Entity Recognition with Pointer Networks Songlin Yang, Kewei Tu∗ School of Information Science and Technology, ShanghaiTech University Shanghai Engineering Research Center of Intelligent Vision and Imaging {yangsl,tukw}@shanghaitech.edu.cn Abstract Constituency parsing and nested named entity recognition (NER) are similar tasks since they both aim to predict a collection of nested and non-crossing spans. In this work, we cast nested NER to constituency parsing and pro- pose a novel pointing mechanism for bottom- up parsing to tackle both tasks. The key idea is based on the observation that if we traverse a constituency tree in post-order, i.e., visit- ing a parent after its children, then two con- secutively visited spans would share a bound- ary. Our model tracks the shared boundaries and predicts the next boundary at each step by leveraging a pointer network. As a result, it needs only linear steps to parse and thus is efficient. It also maintains a parsing con- figuration for structural consistency, i.e., al- ways outputting valid trees. Experimentally, our model achieves the state-of-the-art perfor- mance on PTB among all BERT-based mod- els (96.01 F1 score) and competitive perfor- mance on CTB7 in constituency parsing; and it also achieves strong performance on three benchmark datasets of nested NER: ACE2004, ACE2005, and GENIA 1. 1 Introduction Constituency parsing is an important task in natural language processing, having many applications in downstream tasks, such as semantic role labeling (Fei et al., 2021), opinion mining (Xia et al., 2021), among others. Named entity recognition (NER) is a fundamental task in information extraction and nested NER has been receiving increasing attention due to its broader applications (Byrne, 2007). Constituency parsing and nested NER are simi- lar tasks since they both aim to predict a collection of nested and non-crossing spans (i.e., if two spans overlap, one must be a subspan of the other). Fig.1 ∗Corresponding Author 1Our code is publicly available at https://github. com/sustcsonglin/pointer-net-for-nested Mary acquired those three businesses in May 1988 . NP Indonesia 's reformist-minded President Abdurrahman Wahid PER has blamed the army and police for the deaths . PER PER ORG (a) Span representation: {(0,1,NP),(2,5,NP),(6,8,NP),(5,8,PP),(1,8,VP),(0,9,S)} Span representation: {(0,1,PER),(0,4,PER),(0,6,PER),(8,12,ORG)} Pointing representation: NP VP PP S NP {(0->1,NP),(1->2,Ø),(2->5,NP),(5->6,Ø),(6->8,NP), (8->5,PP),(8->1,VP),(8->9,Ø),(9->0,S)} (b) Pointing representation: {(0->1,PER),(1->4,Ø),(4->0,PER),(4->6,Ø), (6->0,PER),(6->8,Ø),(8->12,ORG),(12->16,Ø),(16->0,Ø)} Figure 1: (a) an example non-binary constituency tree. (b) an example sentence with nested named entities. We show the span and pointing representations. shows example span representations of both tasks. The difference between the two tasks is that the collection of spans form a connected tree in con- stituency parsing, whereas they form several tree fragments in nested NER. However, we can add a node that spans the whole sentence to connect all tree fragments in nested NER to form a tree. Because of the similarity, there are some previous studies adapting methods from the constituency parsing literature to tackle nested …

Content truncated to 500 words