Leveraging Expert Guided Adversarial Augmentation For Improving Generalization in Named Entity Recognition

Findings of the Association for Computational Linguistics: ACL 2022, pages 1947 - 1955 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Leveraging Expert Guided Adversarial Augmentation For Improving Generalization in Named Entity Recognition Aaron Reich1,2, Jiaao Chen1, Aastha Agrawal1, Yanzhe Zhang1, Diyi Yang1 1Georgia Institute of Technology 2Pionetechs, Inc. {areich8, jchen896, aagrawal319, z_yanzhe, dyang888}@gatech.edu Abstract Named Entity Recognition (NER) systems often demonstrate great performance on in-distribution data, but perform poorly on examples drawn from a shifted distribution. One way to evaluate the generalization ability of NER models is to use adversarial examples, on which the specific variations associated with named entities are rarely considered. To this end, we propose leveraging expert-guided heuristics to change the entity tokens and their surrounding contexts thereby altering their entity types as adversarial attacks. Using expert-guided heuristics, we augmented the CoNLL 2003 test set and manually annotated it to construct a high-quality challenging set. We found that state-of-the-art NER systems trained on CoNLL 2003 training data drop per- formance dramatically on our challenging set. By training on adversarial augmented training examples and using mixup for regularization, we were able to significantly improve the performance on the challenging set as well as improve out-of-domain generalization which we evaluated by using OntoNotes data. We athave publicly released our dataset and code https://github.com/GT-SALT/ Guided-Adversarial-Augmentation. 1 Introduction Deep learning models have achieved great per- formance on many natural language processing (NLP) problems (Bahdanau et al., 2016; Devlin et al., 2019). However, many recent works have shown that these models often rely on spurious correlations which are not necessarily the causal artifacts. Thus, these models perform well on the in-distribution test set but are likely to exhibit a huge performance decline on out-of-distribution data (e.g. real world data) (Tu et al., 2020; Kaushik and Lipton, 2018; Poliak et al., 2018; Gururangan et al., 2018; Zhang et al., 2019; Glockner et al., 2018). Prior works have constructed adversar- ial examples for benchmarking the generalization ability of state-of-the-art NLP models on out-of- distribution examples (Kaushik et al., 2020; Zhang et al., 2019; Glockner et al., 2018). Proposed ap- proaches such as random word swapping (Jin et al., 2020) and the appending of a sentence to the end of text (Jia and Liang, 2017) do not take into consider- ation the unique linguistic properties and variations associated with named entities. As a key problem setting involving the classification of semantic cat- egories of entities (e.g., Organizations, Locations) (Nadeau and Sekine, 2007), NER is still in need of improved benchmarks of true generalization. Previous works (Bernier-Colborne and Langlais, 2020; Fu et al., 2020; Stanislawek et al., 2019) have shown that words which have different entity la- bels in different scenarios often lead to frequently occurring errors of NER models. This can be espe- cially problematic in specific domain applications where this challenging case is common. For exam- ple, when training an NER model for political text mining, it would be of great importance to differ- entiate between the categories of Clinton (Person) and the …

Content truncated to 500 words