Multi-grained Named Entity Recognition

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1430–1440 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1430 Multi-Grained Named Entity Recognition Congying Xia1,5, Chenwei Zhang1, Tao Yang2, Yaliang Li3∗, Nan Du2, Xian Wu2, Wei Fan2, Fenglong Ma4, Philip Yu1,5 1University of Illinois at Chicago, Chicago, IL, USA 2Tencent Medical AI Lab, Palo Alto, CA, USA; 3Alibaba Group, Bellevue, WA, USA 4University at Buffalo, Buffalo, NY, USA; 5Zhejiang Lab, Hangzhou, China {cxia8,czhang99,psyu}@uic.edu; yaliang.li@alibaba-inc.com {tytaoyang,kevinxwu,davidwfan}@tencent.com nandu2048@gmail.com; fenglong@buffalo.edu Abstract This paper presents a novel framework, MGNER, for Multi-Grained Named Entity Recognition where multiple entities or en- tity mentions in a sentence could be non- overlapping or totally nested. Different from traditional approaches regarding NER as a se- quential labeling task and annotate entities consecutively, MGNER detects and recog- nizes entities on multiple granularities: it is able to recognize named entities without ex- plicitly assuming non-overlapping or totally nested structures. MGNER consists of a De- tector that examines all possible word seg- ments and a Classifier that categorizes enti- ties. In addition, contextual information and a self-attention mechanism are utilized through- out the framework to improve the NER per- formance. Experimental results show that MGNER outperforms current state-of-the-art baselines up to 4.4% in terms of the F1 score among nested/non-overlapping NER tasks. 1 Introduction Effectively identifying meaningful entities or en- tity mentions from the raw text plays a crucial part in understanding the semantic meanings of natu- ral language. Such a process is usually known as Named Entity Recognition (NER) and it is one of the fundamental tasks in natural language pro- cessing (NLP). A typical NER system takes an utterance as the input and outputs identified enti- ties, such as person names, locations, and organi- zations. The extracted named entities can benefit various subsequent NLP tasks, including syntac- tic parsing (Koo and Collins, 2010), question an- swering (Krishnamurthy and Mitchell, 2015) and relation extraction (Lao and Cohen, 2010). How- ever, accurately recognizing representative entities in natural language remains challenging. ∗Work was done when the author Yaliang Li was at Ten- cent America. Previous works treat NER as a sequence label- ing problem. For example, Lample et al. (2016) achieve a decent performance on NER by incor- porating deep recurrent neural networks (RNNs) with conditional random field (CRF) (Lafferty et al., 2001). However, a critical problem that arises by treating NER as a sequence labeling task is that it only recognizes non-overlapping entities in a single, sequential scan on the raw text; it fails to detect nested named entities which are embed- ded in longer entity mentions, as illustrated in Fig- ure 1. Facility Last night , at the Chinese embassy in France , there was a holiday atmosphere . GPE GPE Figure 1: An example from the ACE-2004 dataset (Doddington et al., 2004) in which two GPEs (Geo- graphical Entities) are nested in a Facility Entity. Due to the semantic structures within natural language, nested entities can be ubiquitous: e.g. 47% of …

Content truncated to 500 words