FactMix: Using a Few Labeled In-domain Examples to Generalize to Cross-domain Named Entity Recognition

Proceedings of the 29th International Conference on Computational Linguistics, pages 5360–5371 October 12–17, 2022. 5360 FactMix: Using a Few Labeled In-domain Examples to Generalize to Cross-domain Named Entity Recognition Linyi Yang1,4∗, Lifan Yuan2∗†, Leyang Cui3, Wenyang Gao1, Yue Zhang1,4 1 School of Engineering, Westlake University 2 Huazhong University of Science and Technology 3 Tencent AI Lab 4 Institute of Advanced Technology, Westlake Institute for Advanced Study 1{yanglinyi, gaowenyang, zhangyue}@westlake.edu.cn 2lievanyuan173@gmail.com 3leyangcui@tencent.com Abstract Few-shot Named Entity Recognition (NER) is imperative for entity tagging in limited re- source domains and thus received proper at- tention in recent years. Existing approaches for few-shot NER are evaluated mainly un- der in-domain settings. In contrast, little is known about how these inherently faithful mod- els perform in cross-domain NER using a few labeled in-domain examples. This paper pro- poses a two-step rationale-centric data augmen- tation method to improve the model’s gener- alization ability. Results on several datasets show that our model-agnostic method signif- icantly improves the performance of cross- domain NER tasks compared to previous state- of-the-art methods, including the data aug- mentation and prompt-tuning methods. Our codes are available at https://github. com/lifan-yuan/FactMix. 1 Introduction Named Entity Recognition (NER) is a subtask of natural language processing, which detects the mentions of named entities in input text, such as location, organization, and person (Sang and De Meulder, 2003; Yang et al., 2017; Cui et al., 2021). It has attracted research from academia and industry due to its broadened usage in customer ser- vices and document parsing as a core task in natural language understanding (Nadeau and Sekine, 2007; Ma and Hovy, 2016; Cui and Zhang, 2019; Yamada et al., 2020). However, training data for NER is available only for limited domains. It has been shown that such labeled data introduces challenges for a model to generalize to new domains (Snell et al., 2017; Ma et al., 2021a; Lin et al., 2021). To address this problem, a line of research con- siders how to allow a model to effectively learn from a few labeled examples in a new target do- main (Zhang et al., 2021; Ma et al., 2021b; Das ∗Equal contribution. Random order of the authorship. †Work done at Westlake University as an intern. Input: Germany imported 47,600 sheep from Britain last year nearly half of total imports. London Manchester French Ireland Australia Korea Entity Set: Location Context-level Semi-fact Examples: Entity-level Replacement [Zeng et al., 2020] Examples: Germany imported 47,600 sheep from Britain last year nearly half of total imports. Query PLM [MASK] … Query PLM [MASK] Germany imported 3,000 beer from Britain last year nearly half of total imports. Germany exported 47,600 pies from Britain last year nearly half of total imports. … Figure 1: The demonstration of two components of Fact- Mix, namely context-level semi-fact and entity-level semi-fact examples. et al., 2021; Chen et al., 2022; Wang et al., 2022a,b). However, such methods still require manual label- ing for target domains, which makes them difficult to generalize to zero-shot diverse domain settings. A different …

Content truncated to 500 words