DoSEA: A Domain-specific Entity-aware Framework for Cross-Domain Named Entity Recogition

Proceedings of the 29th International Conference on Computational Linguistics, pages 2147–2156 October 12–17, 2022. 2147 DoSEA: A Domain-specific Entity-aware Framework for Cross-Domain Named Entity Recogition Minghao Tang1,2, Peng Zhang2,3∗, Yongquan He4, Yongxiu Xu1,2, Chengpeng Chao1,2 and Hongbo Xu1,2 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, UCAS, Beijing, China 3School of Cyberspace Security, NJUST, Nanjing, China 4Meituan, Beijing, China {tangminghao,hbxu}@iie.ac.cn, heyongquan@meituan.com Abstract Cross-domain named entity recognition aims to improve performance in a target domain with shared knowledge from a well-studied source domain. The previous sequence-labeling based method focuses on promoting model parame- ter sharing among domains. However, such a paradigm essentially ignores the domain- specific information and suffers from entity type conflicts. To address these issues, we propose a novel machine reading comprehen- sion based framework, named DoSEA, which can identify domain-specific semantic differ- ences and mitigate the subtype conflicts be- tween domains. Concretely, we introduce an entity existence discrimination task and an entity-aware training setting, to recognize in- consistent entity annotations in the source do- main and bring additional reference to bet- ter share information across domains. Experi- ments on six datasets prove the effectiveness of our DoSEA. Our source code can be obtained from https://github.com/mhtang1995/DoSEA. 1 Introduction Named entity recognition(NER) is a fundamental task in natural language processing and has been extensively studied in various domains. However, acquiring a high performance NER model heav- ily relies on labor-intensive annotated data (Huang et al., 2015; Devlin et al., 2019). Thus, there is a growing interest in cross-domain NER, which aims to exploit the information on a well-studied source domain to improve the performance in a target do- main (Pan and Yang, 2010). Following Daumé III (2007), we focus on the supervised cross-domain NER setting, which utilizes abundant annotated samples from the source domain and small anno- tated samples from the target domain. Previous studies (Kim et al., 2015; Lin and Lu, 2018; Wang et al., 2018b; Jia and Zhang, 2020) typ- ically treat cross-domain NER task as a sequence ∗Peng Zhang is the corresponding author U.S. President Bill Cliton had to drop the resort of Ballybunion Examples from CoNLL-2003 Examples from CrossNER Politics Hagelin stood as a candidate for President of the U.S. for the by Tutsi forces in Rwanda’s civil war ... LOC PER LOC POL COU POLP Spanish Republicans to Mexico after the Civil War... MISC COU EVE MISC MISC PER: Person LOC: Location ORG: Organisation MISC: Miscellaneous ORG POL: Politician COU: Country EVE: Event POLP: Politicalparty from a whirlwind ... former Hutu fled to Zaire after being defeated Natural Law Party , a party found ... his role in admitting defeated Figure 1: Examples from CoNLL-2003 and CrossNER Politics dataset. labeling problem, classifying each word as a type of entity. However, cross-domain NER is chal- lenging due to the entity type difference between domains, since the target domain contains specific entity types. As Figure 1 shows, CrossNER Politics dataset (Liu et al., 2021) contains specific entity types(e.g., Event, Politician, Country …

Content truncated to 500 words