Few-shot Named Entity Recognition with Self-describing Networks

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5711 - 5722 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Few-shot Named Entity Recognition with Self-describing Networks Jiawei Chen1,3,∗, Qing Liu1,3,∗, Hongyu Lin1,†, Xianpei Han1,2,4,†, Le Sun1,2 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China 4Beijing Academy of Artificial Intelligence, Beijing, China {jiawei2020,liuqing2020,hongyu,xianpei,sunle}@iscas.ac.cn Abstract Few-shot NER needs to effectively capture in- formation from limited instances and transfer useful knowledge from external resources. In this paper, we propose a self-describing mech- anism for few-shot NER, which can effectively leverage illustrative instances and precisely transfer knowledge from external resources by describing both entity types and mentions us- ing a universal concept set. Specifically, we design Self-describing Networks (SDNet), a Seq2Seq generation model which can univer- sally describe mentions using concepts, au- tomatically map novel entity types to con- cepts, and adaptively recognize entities on- demand. We pre-train SDNet with large-scale corpus, and conduct experiments on 8 bench- marks from different domains. Experiments show that SDNet achieves competitive perfor- mances on all benchmarks and achieves the new state-of-the-art on 6 benchmarks, which demonstrates its effectiveness and robustness. 1 Introduction Few-shot named entity recognition (FS-NER) aims to identify entity mentions corresponding to new entity types (i.e., novel types) with only a few il- lustrative examples. FS-NER is a promising tech- nique for open-domain NER which contains var- ious unforeseen types and very limited examples and therefore has attached great attention in recent years (Huang et al., 2020; Wang et al., 2021). The main challenge of FS-NER is how to ac- curately model the semantics of unforeseen entity types using only a few illustrative examples. To achieve this, FS-NER needs to effectively capture information in few-shot examples, meanwhile ex- ploiting and transferring useful knowledge from external resources. Unfortunately, information en- tailed in illustrative examples is very limited, i.e., the limited information challenge. And external ∗Equally Contribution. †Corresponding authors. Travelling to America, … Central Park is traversed … …… geographic entity Concepts Mentions types Wikipedia … to leave for America. Dr. Kohl came to Beijing … …… GPE Buck mountain tomorrow … … going to America instead. …… location OntoNotes WNUT17 ������ �� ������ ����� ���� �������� ��� �� ������ � ���� ����� ����� Figure 1: Examples of concept description. Wikipedia, OntoNotes, WNUT17 can transfer knowledge to each other by describing mentions and types using a univer- sal concept set. knowledge usually doesn’t directly match with the new task because it may contain irrelevant, hetero- geneous, or even conflicting knowledge (Beryozkin et al., 2019; Yu and Yang, 2020) – which we re- fer as knowledge mismatch challenge. For exam- ple, the schemas in Wikipedia, OntoNotes (Ralph et al., 2013) and WNUT17 (Derczynski et al., 2017) are conflicting, where “America” is geographic entity in Wikipedia, GPE in OntoNotes, and location in WNUT17. Such a knowledge mis- match problem makes it …

Content truncated to 500 words