SCL-RAI: Span-based Contrastive Learning with Retrieval Augmented Inference for Unlabeled Entity Problem in NER

Proceedings of the 29th International Conference on Computational Linguistics, pages 2313–2318 October 12–17, 2022. 2313 SCL-RAI: Span-based Contrastive Learning with Retrieval Augmented Inference for Unlabeled Entity Problem in NER Shuzheng Si1,2∗, Shuang Zeng1,2∗, Jiaxing Lin1,2, Baobao Chang1† 1Key Laboratory of Computational Linguistics, Peking University, MOE, China 2School of Software and Microelectronics, Peking University, China {sishuzheng,jxlin}@stu.pku.edu.cn {zengs,chbb}@pku.edu.cn Abstract Named Entity Recognition is the task to locate and classify the entities in the text. However, Unlabeled Entity Problem in NER datasets se- riously hinders the improvement of NER per- formance. This paper proposes SCL-RAI to cope with this problem. Firstly, we decrease the distance of span representations with the same label while increasing it for different ones via span-based contrastive learning, which relieves the ambiguity among entities and improves the robustness of the model over unlabeled entities. Then we propose retrieval augmented inference to mitigate the decision boundary shifting prob- lem. Our method significantly outperforms the previous SOTA method by 4.21% and 8.64% F1-score on two real-world datasets. 1 Introduction As a fundamental task in NLP, Named Entity Recognition aims to locate and classify named entities in the text. Due to the large-scale well- annotated datasets, deep-learning based methods (Li et al., 2022b; Devlin et al., 2019) have achieved great success. However, in real-world datasets, such as Ling and Weld (2012) with 112 fine-grained named entity tags, a large set of entity classes may cause inevitable missing annotations. Moreover, to obtain large NER datasets in practical scenar- ios, the distant supervision approach (Ren et al., 2015; Fries et al., 2017) may make this problem even worse, since the entity dictionary cannot cover all entities. Previous work (Li et al., 2021; Shang et al., 2018) find that this problem seriously hinders the performance of the NER model and name this problem as Unlabeled Entity Problem. As shown in Figure 1, the unlabeled second “NBA” may con- fuse model and introduce unnecessary noise. To cope with this problem, several attempts from different perspectives have been proposed. Inspired *Equal contribution. †Corresponding author. Lebron James and Steph Curry are both top players in the NBA, but it was Giannis Antetokounmpo who took home the 2021 NBA Finals. PERSON ORG PERSON PERSON ORG Figure 1: A toy case to show Unlabeled Entity Problem. The labeled entities are underlined with red lines while the unlabeled entities are underlined with dashed lines. by positive-unlabeled (PU) learning (Li and Liu, 2005), Peng et al. (2019) use a weighted loss to assign low weights to false negative words and build distinct binary classifiers for different entity types. However, they require prior information or heuristics (Li et al., 2022c) and the unlabeled entities still misguide the classifiers, bringing ambi- guity among neighboring entities (Li et al., 2021). Yang et al. (2018); Jie et al. (2019) introduce the Partial CRF (Lafferty et al., 2001) to marginalize the instances that are consistent with the incom- plete annotation. However, they require additional well-annotated corpus to get ground truth negative instances, which are usually unavailable in …

Content truncated to 500 words