Modularized Interaction Network for Named Entity Recognition

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 200–209 August 1–6, 2021. ©2021 Association for Computational Linguistics 200 Modularized Interaction Network for Named Entity Recognition Fei Li1, Zheng Wang2∗, Siu Cheung Hui2 , Lejian Liao1 , Dandan Song1∗, Jing Xu1 , Guoxiu He3 , Meihuizi Jia1 1Beijing Institute of Technology, China 2Nanyang Technological University, Singapore 3Wuhan University, China {lifei926,liaolj,sdd,xujing,jmhuizi24}@bit.edu.cn {wang zheng,asschui}@ntu.edu.sg, guoxiu.he@whu.edu.cn Abstract Although the existing Named Entity Recogni- tion (NER) models have achieved promising performance, they suffer from certain draw- backs. The sequence labeling-based NER models do not perform well in recognizing long entities as they focus only on word-level information, while the segment-based NER models which focus on processing segment in- stead of single word are unable to capture the word-level dependencies within the segment. Moreover, as boundary detection and type pre- diction may cooperate with each other for the NER task, it is also important for the two sub- tasks to mutually reinforce each other by shar- ing their information. In this paper, we pro- pose a novel Modularized Interaction Network (MIN) model which utilizes both segment- level information and word-level dependen- cies, and incorporates an interaction mecha- nism to support information sharing between boundary detection and type prediction to en- hance the performance for the NER task. We have conducted extensive experiments based on three NER benchmark datasets. The per- formance results have shown that the proposed MIN model has outperformed the current state- of-the-art models. 1 Introduction Named Entity Recognition (NER) is one of the fundamental tasks in natural language processing (NLP) that intends to find and classify the type of a named entity in text such as person (PER), location (LOC) or organization (ORG). It has been widely used for many downstream applications such as relation extraction (Xiong et al., 2018), entity link- ing (Gupta et al., 2017), question generation (Zhou et al., 2017) and coreference resolution (Barhom et al., 2019). ∗Corresponding authors. Currently, there are two types of methods for the NER task. The first one is sequence labeling-based methods (Lample et al., 2016; Chiu and Nichols, 2016; Luo et al., 2020), in which each word in a sen- tence is assigned a special label (e.g., B-PER or I- PER). Such methods can capture the dependencies between adjacent word-level labels and maximize the probability of predicted labels over the whole sentence. It has achieved the state-of-the-art perfor- mance in various datasets over the years. However, NER is a segment-level recognition task. As such, the sequence labeling-based models which focus only on word-level information do not perform well especially in recognizing long entities (Ye and Ling, 2018). Recently, segment-based methods (Kong et al., 2016; Li et al., 2020b; Yu et al., 2020b; Li et al., 2021) have gained popularity for the NER task. They process segment (i.e., a span of words) instead of single word as the basic unit and assign a special label (e.g., PER, ORG or LOC) …

Content truncated to 500 words