Find Parent then Label Children: A Two-stage Taxonomy Completion Method with Pre-trained Language Model

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1032–1042 May 2-6, 2023 ©2023 Association for Computational Linguistics Find Parent then Label Children: A Two-stage Taxonomy Completion Method with Pre-trained Language Model Fei Xia1,2∗, Yixuan Weng1∗, Shizhu He1,2, Kang Liu1,2 and Jun Zhao1,2 1 National Laboratory of Pattern Recognition„ Institute of Automation, CAS 2 School of Artificial Intelligence, University of Chinese Academy of Sciences xiafei2020@ia.ac.cn,wengsyx@gmail.com {shizhu.he, kliu, jzhao}@nlpr.ia.ac.cn Abstract Taxonomies, which organize domain concepts into hierarchical structures, are crucial for building knowledge systems and downstream applications. As domain knowledge evolves, taxonomies need to be continuously updated to include new concepts. Previous approaches have mainly focused on adding concepts to the leaf nodes of the existing hierarchical tree, which does not fully utilize the taxonomy’s knowledge and is unable to update the original taxonomy structure (usually involving non- leaf nodes). In this paper, we propose a two- stage method called ATTEMPT for taxonomy completion. Our method inserts new con- cepts into the correct position by finding a par- ent node and labeling child nodes. Specifi- cally, by combining local nodes with prompts to generate natural sentences, we take advan- tage of pre-trained language models for hy- pernym/hyponymy recognition. Experimental results on two public datasets (including six domains) show that ATTEMPT performs best on both taxonomy completion and extension tasks, surpassing existing methods. 1 Introduction Taxonomies1 are an important form of domain knowledge that organize concepts into hierarchi- cal structures, representing “hypernym-hyponym” relationships among concepts in the form of trees or directed acyclic graphs (Shen et al., 2020). Taxonomies are essential components of knowl- edge systems such as ontologies and knowledge graph (Yu et al., 2020), and are widely used in various downstream applications, including search engineering (Yin and Shah, 2010), recommenda- tion systems (Huang et al., 2019; Zhang et al., ∗ These authors contribute this work equally. 1In this paper, we mainly focus on the taxonomy repre- sented as tree rather than directed acyclic graph, because trees are the mainstream form at present, such as the online catalog taxonomies of Amazon and Yelp. Existing Taxonomy Expanded Taxonomy Science Natural Science Social Science Chemistry Physics Anthropology Civics Mechanics Archeology Science Natural Science Chemistry Mechanics Archeology Physics Anthropology Civics New concept Social Science Figure 1: An example of taxonomy completion for a non-leaf node. The new concept “Social Science" needs to be inserted into the correct position in the ex- isting taxonomy. 2014), and information filtering (Demeester et al., 2016). As domain knowledge continues to evolve, es- pecially with the rapid growth of web content, new concepts are constantly emerging. In order to stay current, original taxonomies must incorpo- rate these new concepts and adapt their hierarchi- cal relationships. For example, as shown in Figure 1, with the advancement of sociology and science, the concept of "Social Science" should be added to the science knowledge system, and the original structure should be adjusted accordingly. However, existing taxonomies are primarily constructed by human experts (Shen et al., …

Content truncated to 500 words