ECNU_ICA_1 SemEval-2021 Task 4: Leveraging Knowledge-enhanced Graph Attention Networks for Reading Comprehension of Abstract Meaning

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 183–188 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 183 ECNU ICA 1 at SemEval-2021 Task 4: Knowledge-Enhanced Graph Attention Networks for Reading Comprehension of Abstract Meaning Pingsheng Liu, Linlin Wang∗, Qian Zhao, Hao Chen, Yuxi Feng, Xin Lin∗, liang he School of Computer Science and Technology, East China Normal University, Shanghai 200062, China {51205901014,51205901129,10175102236}@stu.ecnu.edu.cn im0qianqian@ica.stc.sh.cn {llwang,xlin,lhe}@cs.ecnu.edu.cn Abstract This paper describes our system ECNU ICA 1 for SemEval-2021 Task 4: Reading Compre- hension of Abstract Meaning. For this task, we utilize knowledge-enhanced Graph Atten- tion Networks with a novel semantic space transformation strategy. It leverages heteroge- neous knowledge to learn adequate evidences, and seeks for an effective semantic space of abstract concepts to better improve the ability of a machine in understanding abstract mean- ings of natural language. Experimental results show that our system achieves strong perfor- mance on this task in terms of both impercep- tibility and nonspecificity. 1 Introduction Recent years have witnessed the remarkable suc- cess of pre-trained language models in machine reading comprehension (MRC). Nevertheless, new research points out that these dominant approaches rely heavily on superficial text pattern-matching heuristics to achieve shallow comprehension on natural language (Zhang et al., 2020). For humans, the basic ability to represent abstract concepts guar- antees an in-depth understanding of natural lan- guage. Consequently, teaching machines to better comprehend abstract meaning is a significant and urgent step to push the frontier technique of MRC forward. If computers can understand passages as human do, we expect them to accurately predict abstract words that people can use in summaries of the given passages. Thus, researchers have recently proposed a reading comprehension of abstract meaning (Re- CAM) task in SemEval 2021. Unlike some previ- ous datasets such as CNN/Daily Mail (Hermann et al., 2015) that request computers to predict con- crete concepts, e.g., named entities, ReCAM re- quires machines to fill out abstract words removed *Equal corresponding authors. from human written summaries. In ReCAM, sub- task 1 and subtask 2 respectively evaluate the per- formance of machines towards imperceptibility and nonspecificity, two formal definitions of abstract- ness in natural language understanding(Spreen and Schulz, 1966; Changizi, 2008). Specifically, con- crete words refer to things, events, and proper- ties that we can perceive directly with our senses (Spreen and Schulz, 1966; Coltheart, 1981; Turney et al., 2011), e.g., donut, trees, and red. In contrast, abstract words refer to the ideas and concepts that are distant from immediate perception. Examples for abstract words include objective, culture, and economy. Subtask 1 requires machines to perform reading comprehension of abstract meaning for im- perceptible concepts, while subtask 2 concentrates on hypernyms, which is more abstract and different from the concrete concepts (Changizi, 2008). To better understand the abstract meaning, we utilize the Knowledge-Enhanced Graph Attention Network (KEGAT) architecture with a novel se- mantic space transformation strategy for ReCAM. It well incorporates structured knowledge base such as ConceptNet (Speer et al., 2017) and exploits a …

Content truncated to 500 words