Boundary Smoothing for Named Entity Recognition

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7096 - 7108 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Boundary Smoothing for Named Entity Recognition Enwei Zhu1,2 and Jinpeng Li1,2,∗ 1HwaMei Hospital, University of Chinese Academy of Sciences 2Ningbo Institute of Life and Health Industry, University of Chinese Academy of Sciences {zhuenwei,lijinpeng}@ucas.ac.cn Abstract Neural named entity recognition (NER) mod- els may easily encounter the over-confidence issue, which degrades the performance and calibration. Inspired by label smoothing and driven by the ambiguity of boundary annota- tion in NER engineering, we propose bound- ary smoothing as a regularization technique for span-based neural NER models. It re-assigns entity probabilities from annotated spans to the surrounding ones. Built on a simple but strong baseline, our model achieves results bet- ter than or competitive with previous state- of-the-art systems on eight well-known NER benchmarks.1 Further empirical analysis sug- gests that boundary smoothing effectively mit- igates over-confidence, improves model cali- bration, and brings flatter neural minima and more smoothed loss landscapes. 1 Introduction Named entity recognition (NER) is one of the fun- damental natural language processing (NLP) tasks with extensive investigations. As a common setting, an entity is regarded as correctly recognized only if its type and two boundaries exactly match the ground truth. The annotation of boundaries is more ambigu- ous, error-prone, and raises more inconsistencies than entity types. For example, the CoNLL 2003 task contains four entity types (i.e., person, loca- tion, organization, miscellaneous), which are easy to distinguish between. However, the boundaries of a entity mention could be ambiguous, because of the “boundary words” (e.g., articles or modi- fiers). Considerable efforts are required to specify the “gold standard practice” case by case. Table 1 presents some examples from CoNLL 2003 An- ∗Corresponding author. 1Our code is available at https://github.com/ syuoni/eznlp. Text Boundary words [The [White House]ORG]ORG Article [The [Godfather]PER]PER Article [[Clinton]PER government]ORG Modifier [Mr. [Harry Schearer]PER]PER Person title [[John Doe]PER, Jr.]PER Name appositive Table 1: Examples of CoNLL 2003 Annotation Guide- lines and potential alternatives. The gold annotations are marked in blue [*], whereas the alternative annota- tions are in red [*]. notation Guidelines.2 In addition, some studies have also reported that incorrect boundary is a ma- jor source of entity recognition error (Wang et al., 2019; Eberts and Ulges, 2020). Recently, span-based models have gained much popularity in NER studies, and achieved state-of- the-art (SOTA) results (Eberts and Ulges, 2020; Yu et al., 2020; Li et al., 2021). This approach typically enumerates all candidate spans and classi- fies them into entity types (including a “non-entity” type); the annotated spans are scarce and assigned with full probability to be an entity, whereas all other spans are assigned with zero probability. This creates noticeable sharpness between the classifica- tion targets of adjacent spans, and may thus plague the trainability of neural networks. In addition, empirical evidence shows that these models easily encounter the over-confidence issue, i.e., the confi- dence of a predicted entity is much …

Content truncated to 500 words