Nested Named Entity Recognition as Latent Lexicalized Constituency Parsing

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6183 - 6198 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Nested Named Entity Recognition as Latent Lexicalized Constituency Parsing Chao Lou, Songlin Yang, Kewei Tu∗ School of Information Science and Technology, ShanghaiTech University Shanghai Engineering Research Center of Intelligent Vision and Imaging {louchao,yangsl,tukw}@shanghaitech.edu.cn Abstract Nested named entity recognition (NER) has been receiving increasing attention. Recently, Fu et al. (2020) adapt a span-based con- stituency parser to tackle nested NER. They treat nested entities as partially-observed con- stituency trees and propose the masked inside algorithm for partial marginalization. How- ever, their method cannot leverage entity heads, which have been shown useful in entity mention detection and entity typing. In this work, we resort to more expressive structures, lexicalized constituency trees in which con- stituents are annotated by headwords, to model nested entities. We leverage the Eisner-Satta algorithm to perform partial marginalization and inference efficiently. In addition, we pro- pose to use (1) a two-stage strategy (2) a head regularization loss and (3) a head-aware label- ing loss in order to enhance the performance. We make a thorough ablation study to investi- gate the functionality of each component. Ex- perimentally, our method achieves the state-of- the-art performance on ACE2004, ACE2005 and NNE, and competitive performance on GENIA, and meanwhile has a fast inference speed. Our code will be publicly available at: github.com/LouChao98/nner_as_parsing. 1 Introduction Named Entity Recognition (NER) is a fundamen- tal task in information extraction, playing an es- sential role in many downstream tasks. Nested NER brings more flexibility than flat NER by al- lowing nested structures, thereby enabling more fine-grained meaning representations and broader applications (Byrne, 2007; Dai, 2018). Traditional sequence-labeling-based models have achieved re- markable performance on flat NER but fail to han- dle nested entities. To resolve this problem, there are many layer-based methods (Ju et al., 2018; Fisher and Vlachos, 2019; Shibuya and Hovy, ∗Corresponding Author 2020; Wang et al., 2020, 2021) proposed to rec- ognize entities layer-by-layer in bottom-up or top- down manners. However, they suffer from the error propagation issue due to the cascade decoding. Recently, Fu et al. (2020) adapt a span-based constituency parser to tackle nested NER, treating annotated entity spans as a partially-observed con- stituency tree and marginalizing latent spans out for training. Their parsing-based method, namely PO- TreeCRF, admits global exact inference thanks to the CYK algorithm (Cocke, 1969; Younger, 1967; Kasami, 1965), thereby eliminating the error prop- agation problem. However, their method does not consider entity heads, which provide important clues for entity mention detection (Lin et al., 2019; Zhang et al., 2020d) and entity typing (Katiyar and Cardie, 2018; Choi et al., 2018; Chen et al., 2021). For example, University and California are strong clues of the existence of ORGEDU and STATE en- tities in Fig.1. Motivated by this and inspired by head-driven phrase structures, Lin et al. (2019) pro- pose the Anchor-Region Network (ARN), which identifies all entity heads firstly …

Content truncated to 500 words