Cross-lingual Inference with A Chinese Entailment Graph

Findings of the Association for Computational Linguistics: ACL 2022, pages 85 - 96 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Extract-Select: A Span Selection Framework for Nested Named Entity Recognition with Generative Adversarial Training Peixin Huang1, Xiang Zhao1,2, Minghao Hu3∗, Yang Fang1, Xinyi Li1 and Weidong Xiao1,2 1 National University of Defense Technology, Changsha, China 2 Collaborative Innovation Center of Geospatial Technology, Wuhan, China 3 Information Research Center of Military Science, Beijing, China {huangpeixin15, xiangzhao, fangyang12, wdxiao}@nudt.edu.cn huminghao16@gmail.com Abstract Nested named entity recognition (Nested NER) aims to identify named entities which may overlap. Span-based approaches regard nested NER as a two-stage task—span extrac- tion and classification, thus having the innate ability to handle this task. However, they face the problems of error propagation, ig- norance of span boundary, difficulty in long entity recognition and requirement on large- scale annotated data. In this paper, we propose Extract-Select, a span selection framework for nested NER, to tackle these problems. Firstly, we introduce a span selection framework in which nested entities with different entity cat- egories would be separately extracted by the extractor, thus naturally avoiding error propa- gation in prior two-stage approaches. In the inference phase, the trained extractor selects final results specific to the given entity cate- gory. Secondly, we propose a hybrid selection strategy in the extractor, which not only makes full use of both span boundary and span con- tent, but also improves the ability of long en- tity recognition. Thirdly, we design a discrimi- nator to evaluate the extraction result, and train both extractor and discriminator with gener- ative adversarial training (GAT). The use of GAT greatly alleviates the stress on the dataset size. Experimental results on four benchmark datasets demonstrate that Extract-Select out- performs competitive nested NER models, ob- taining state-of-the-art results. The proposed model also performs well with less labeled data, proving the effectiveness of GAT. 1 Introduction Named entity recognition (NER) aims at detect- ing the spans and semantic categories of entities from the text. Previous studies usually treat NER as a sequential labeling problem (Ma and Hovy, 2016; Chiu and Nichols, 2016). These studies re- strict each token belonging to at most one entity mention, and hence it is unable to handle nested NER (Huang et al., 2015), where one token may ∗ Corresponding author Alpha B2 proteins bound the PEBP2 site within the mouse GM-CSF promoter The premier of the western Canadian province of British Columbia ... LOC GPE GPE PER PROTEIN： DNA： Alpha, PEBP2, GM-CSF PEBP2 site, mouse GM-CSF promoter Figure 1: Examples for nested entities from ACE2005 and GENIA corpora. belong to multiple mentions. For example in Fig- ure 1, a LOC (i.e., Location) entity “western Cana- dian” is nested in another GPE (i.e., Geo-Political Entity) entity “the western Canadian province of British Columbia”. Some studies seek to reconcile sequential la- beling with nested NER (Alex et al., 2007; Ju et al., 2018). However, sequential labeling is naturally unsuitable for assigning multiple labels to a single token. Considering that, some …

Content truncated to 500 words