Nested Named Entity Recognition with Span-level Graphs

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 892 - 903 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Nested Named Entity Recognition with Span-level Graphs Juncheng Wan, Dongyu Ru, Weinan Zhang∗, Yong Yu Shanghai Jiao Tong University {junchengwan,maxru,wnzhang,yyu}@apex.sjtu.edu.cn Abstract Span-based methods with the neural net- works backbone have great potential for the nested named entity recognition (NER) prob- lem. However, they face problems such as degenerating when positive instances and neg- ative instances largely overlap. Besides, the generalization ability matters a lot in nested NER, as a large proportion of entities in the test set hardly appear in the training set. In this work, we try to improve the span represen- tation by utilizing retrieval-based span-level graphs, connecting spans and entities in the training data based on n-gram features. Specif- ically, we build the entity-entity graph and span-entity graph globally based on n-gram similarity to integrate the information of sim- ilar neighbor entities into the span represen- tation. To evaluate our method, we conduct experiments on three common nested NER datasets, ACE2004, ACE2005, and GENIA datasets. Experimental results show that our method achieves general improvements on all three benchmarks (+0.30 ∼ 0.85 micro-F1), and obtains special superiority on low fre- quency entities (+0.56 ∼ 2.08 recall). 1 Introduction Named entity recognition is one of the major sub- tasks of information extraction for extracting cat- egorized named entities from unstructured text. Recently, neural-based NER architectures have shown remarkable performance with minimal fea- ture engineering, such as CNN-CRF (Collobert et al., 2011), BiLSTM+CRF (Lample et al., 2016), LSTM-CNN-CRF (Ma and Hovy, 2016) and Lat- tice LSTM (Zhang and Yang, 2018a). Despite their great success, nested NER raises new chal- lenges due to the deeply overlapping or nested en- tities (Finkel and Manning, 2009). In nested NER, a token may be included in multiple entities (Wang ∗Corresponding author Another tornado hit Geneva , near the Alabama - Florida line , said Mayor Warren Beck . GPE GPE GPE PER PER LOC Figure 1: An example of nested NER in ACE2005 dataset. and Lu, 2018) instead of a single one in the conven- tional setting, making the problem more difficult to solve. Previous exploration on nested NER can be mainly divided into three categories, using vari- ous architectures or different formulation for adap- tation to the nested scenario. Hypergraph-based methods use explicit hypergraphs to represent the possible nested structure or investigate the graph-shaped lexical/syntactic features (Lu and Roth, 2015; Katiyar and Cardie, 2018; Wang and Lu, 2018). Layered-based methods construct the nested structure through an action sequence (Wang et al., 2018; Fisher and Vlachos, 2019; Shibuya and Hovy, 2020) or layered-models (Ju et al., 2018; Wang et al., 2020). Span-based methods directly enumerate spans in a sentence, and perform cate- gorical prediction on each span (Lin et al., 2019; Eberts and Ulges, 2020; Luan et al., 2019; Tan et al., 2020). Span-based methods adopt the most simple and straightforward formulation as span classifica- tion, …

Content truncated to 500 words