SetGNER: General Named Entity Recognition as Entity Set Generation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3074 - 3085 December 7-11, 2022 ©2022 Association for Computational Linguistics SetGNER: General Named Entity Recognition as Entity Set Generation Yuxin He1 and Buzhou Tang1,2,∗ 1Department of Computer Science, Harbin Institute of Technology, Shenzhen, China 2Peng Cheng Laboratory, Shenzhen, China 21S051047@stu.hit.edu.cn tangbuzhou@gmail.com Abstract Recently, joint recognition of flat, nested and discontinuous entities has received increasing attention. Motivated by the observation that the target output of NER is essentially a set of sequences, we propose a novel entity set gener- ation framework for general NER scenes in this paper. Different from sequence-to-sequence NER methods, our method does not force the entities to be generated in a predefined order and can get rid of the problem of error propa- gation and inefficient decoding. Distinguished from the set-prediction NER framework, our method treats each entity as a sequence and is capable of recognizing discontinuous mentions. Given an input sentence, the model first en- codes the sentence in word-level and detects po- tential entity mentions based on the encoder’s output, then reconstructs entity mentions from the detected entity heads in parallel. To let the encoder of our model capture better right- to-left semantic structure, we also propose an auxiliary Inverse Generation Training task. Ex- tensive experiments show that our model (w/o. Inverse Generation Training) outperforms state- of-the-art generative NER models by a large margin on two discontinuous NER datasets, two nested NER datasets and one flat NER dataset. Besides, the auxiliary Inverse Genera- tion Training task is found to further improve the model’s performance on the five datasets. 1 Introduction Named entity recognition (NER) is a fundamental task in the field of information extraction and has played an important role in the development of natural language processing. There exist three well- studied subtasks of NER, i.e. flat NER, nested NER and discontinuous NER, as illustrated in Figure 1. Recently, researchers have grown more inter- est in tackling the three subtasks jointly, which we refer to as general NER (Li et al., 2021; Dai *Corresponding Author. Figure 1: Examples of the discontinuous / nested / flat NER. It can be observed that entities may overlap with each other in general NER scenes and these overlapped entities are inherently unordered. Hence, instead of gen- erating entities sequentially in a predefined order, it is more suitable to generate the set of entities concurrently. et al., 2020; Yan et al., 2021). Existing frame- works for general NER fall into three categories: (1) span-based models; (2) sequence-to-sequence (seq2seq) models; (3) models based on other tech- niques like hyper-graphs and shift-reduce parsers. Among them, seq2seq models (Fei et al., 2021; Yan et al., 2021) have demonstrated SOTA performance. However, they organize target entities into a single sequence according to a predetermined order. This formulation violates the fact that the target enti- ties are inherently a unordered set, and introduces an incorrect bias (entity-order confounder) to the model (Tan et al., 2021; Zhang et al., 2022). In addition, …

Content truncated to 500 words