COPEN: Probing Conceptual Knowledge in Pre-trained Language Models

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5015 - 5035 December 7-11, 2022 ©2022 Association for Computational Linguistics COPEN: Probing Conceptual Knowledge in Pre-trained Language Models Hao Peng1,2∗, Xiaozhi Wang1,2∗, Shengding Hu1,2, Hailong Jin1,2, Lei Hou1,2† , Juanzi Li1,2, Zhiyuan Liu1,2, Qun Liu3 1Department of Computer Science and Technology, BNRist; 2KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing, 100084, China 3Huawei Noah’s Ark Lab {peng-h21, wangxz20}@mails.tsinghua.edu.cn Abstract Conceptual knowledge is fundamental to hu- man cognition and knowledge bases. However, existing knowledge probing works only focus on evaluating factual knowledge of pre-trained language models (PLMs) and ignore concep- tual knowledge. Since conceptual knowledge often appears as implicit commonsense behind texts, designing probes for conceptual knowl- edge is hard. Inspired by knowledge represen- tation schemata, we comprehensively evaluate conceptual knowledge of PLMs by designing three tasks to probe whether PLMs organize entities by conceptual similarities, learn con- ceptual properties, and conceptualize entities in contexts, respectively. For the tasks, we col- lect and annotate 24k data instances covering 393 concepts, which is COPEN, a COncep- tual knowledge Probing bENchmark. Exten- sive experiments on different sizes and types of PLMs show that existing PLMs systemat- ically lack conceptual knowledge and suffer from various spurious correlations. We be- lieve this is a critical bottleneck for realiz- ing human-like cognition in PLMs. COPEN and our codes are publicly released at https: //github.com/THU-KEG/COPEN. 1 Introduction Pre-trained language models (PLMs) have achieved superior performance on most NLP tasks requir- ing substantial world knowledge (Qiu et al., 2020; Han et al., 2021). It is interesting and meaning- ful to probe the extent and scope of world knowl- edge within PLMs. Existing knowledge probing works have evaluated PLMs’ knowledge about en- tities (Broscheit, 2019; Tenney et al., 2019a) and their relations (Petroni et al., 2019; Jiang et al., 2020; Roberts et al., 2020), i.e., factual knowledge, but ignore conceptual knowledge. ∗ Equal contribution † Corresponding author: L.Hou Animals can move. UK Grumpy Cat USA Mammal Animal Bird Country Dolly nationality nationality Relation Instance of Subclass of Most mammals have navels. Birds have feathers. … … … … Entity Graph Concept Taxonomy Property … Figure 1: An example knowledge graph. Entities are or- ganized by concepts through the Instance of relation and concepts are organized into a taxonomy through the Subclass of relation. Each concept has certain prop- erties. Existing work only probes factual knowledge in entity graphs, ignoring conceptual knowledge in the concept taxonomy and Instance of relation. Conceptual knowledge, especially the abstrac- tion ability, is fundamental to all kinds of hu- man cognition (Carey, 1991; Collins and Olson, 2014) including language processing (Waxman and Markow, 1995; Wellsby and Pexman, 2014). Just as the quote of psychologist Gregory Murphy, con- cepts are the glue that holds our mental world together (Murphy, 2004). Moreover, knowledge bases (Suchanek et al., 2007; Auer et al., 2007; Vrandeˇci´c, 2012) organize massive entities via con- cept taxonomies as illustrated in Figure 1, which enable broad applications (Lv et al., …

Content truncated to 500 words