Thai Nested Named Entity Recognition Corpus

Findings of the Association for Computational Linguistics: ACL 2022, pages 1473 - 1486 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Thai Nested Named Entity Recognition Corpus Weerayut Buaphet†, Can Udomcharoenchaikit†,∗, Peerat Limkonchotiwat†, Attapol T. Rutherford‡, Sarana Nutanong† †School of Information Science and Technology, VISTEC, Thailand ‡Department of Linguistic, Chulalongkorn University, Thailand {Weerayut.b_s20, canu_pro, peerat.l_s19,snutanon}@vistec.ac.th, attapol.t@chula.ac.th Abstract This paper presents the first Thai Nested Named Entity Recognition (N-NER) dataset. Thai N-NER consists of 264,798 mentions, 104 classes, and a maximum depth of 8 layers obtained from news articles and restaurant re- views, a total of 4894 documents. Our work, to the best of our knowledge, presents the largest non-English N-NER dataset and the first non- English one with fine-grained classes. To understand the new challenges our proposed dataset brings to the field, we conduct an ex- perimental study on (i) cutting edge N-NER models with the state-of-the-art accuracy in En- glish and (ii) baseline methods based on well- known language model architectures. From the experimental results, we obtain two key findings. First, all models produce poor F1 scores in the tail region of the class distribu- tion. There is little or no performance improve- ment provided by these models with respect to the baseline methods with our Thai dataset. These findings suggest that further investiga- tion is required to make a multilingual N-NER solution that works well across different lan- guages. The dataset and code are available at: github.com/vistec-AI/Thai-NNER.git 1 Introduction Named Entity Recognition (NER) is a task of ex- tracting named entities from given text. It identi- fies the span of each entity and categorizes the iden- tified span into an entity category. NER is essen- tial in many downstream tasks, e.g., entity linking, question answering, and knowledge graph. In addi- tion, Yamada et al. (2020) show that the contextual- ized representations that include entity information can improve many downstream tasks. The conventional NER paradigm can only label one entity type for each entity span. For example, the entity “Chiang Mai University” will be con- sidered as a single span ignoring the nested struc- ture of the term “Chiang Mai,” which is the name ∗* - Corresponding author: canu_pro@vistec.ac.th ����� ��� ������ ��� ��� ���� ���� ���� ���� ��� ��� ���� ��� ���� ��� ���� ���� ��� ��� ��� ������� ������������ ��������������� ����������� ��� ������� ������ �������� �������������� ���� ������� ������� ������� ���� ��������� ������� �������� ������ �������� ���������� ��������� �������� ����� �������� �������� ���������� ����� �������� ��� ����� ����� ����������� ������������ ������� ����� ������� ���������� ������������ ���� ��� �������� ��� ������������ ������������ ������� ������� ������ ����������� ����������� ������������� �������� �������� �������� ������ ���� ����� ���� ������� �������� ����� ����� ���� ������ ���� ���� ���� ������� �������� ����� ����� ������ ����� ������ ������� ���������� ������������ ���������� ������� ����������� ���������� �������� ���� ���������� ������� �������� ������������� ������ ���������� ���� ����� ���� ��� �������� ����� �������� ��� ������ ���� ���� ���� ��� ���� ������� ��� ����� ������ ����� ����� Figure 1: An overview of named-entity classes in Thai N-NER corpus. Our …

Content truncated to 500 words