End-to-end Case-Based Reasoning for Commonsense Knowledge Base Completion

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3509–3522 May 2-6, 2023 ©2023 Association for Computational Linguistics End-to-end Case-Based Reasoning for Commonsense Knowledge Base Completion Zonglin Yang♣∗ Xinya Du♠∗ Erik Cambria♣ Claire Cardie♦ ♣ Nanyang Technological University ♠ University of Texas at Dallas ♦ Cornell University {zonglin.yang, cambria}@ntu.edu.sg xinya.du@utdallas.edu cardie@cs.cornell.edu Abstract Pretrained language models have been shown to store knowledge in their parameters and have achieved reasonable performance in com- monsense knowledge base completion (CKBC) tasks. However, CKBC is knowledge-intensive and it is reported that pretrained language models’ performance in knowledge-intensive tasks are limited because of their incapa- bility of accessing and manipulating knowl- edge. As a result, we hypothesize that pro- viding retrieved passages that contain relevant knowledge as additional input to the CKBC task will improve performance. In particular, we draw insights from Case-Based Reason- ing (CBR) – which aims to solve a new prob- lem by reasoning with retrieved relevant cases, and investigate the direct application of it to CKBC. On two benchmark datasets, we demon- strate through automatic and human evaluations that our End-to-end Case-Based Reasoning Framework (ECBRF) generates more valid knowledge than the state-of-the-art COMET model for CKBC in both the fully supervised and few-shot settings. From the perspective of CBR, our framework addresses a fundamental question on whether CBR methodology can be utilized to improve deep learning models. 1 Introduction Commonsense knowledge helps humans navigate everyday situations seamlessly (Apperly, 2010) and is required for many intelligent scenarios (Davis and Marcus, 2015). To automatically enlarge the scale of commonsense knowledge base for the ben- efit of reducing labeling labor and expense, Knowl- edge Graph Completion (KGC) has become a hot research topic (Ji et al., 2022). The general KGC task is to expand existing knowledge graphs by us- ing well-trained classifiers—they are trained with existing annotated samples and predict whether or not there is a relationship between two existing en- tities in a knowledge graph (Wang et al., 2017). ∗Contribution starts from their stay at Cornell. Subject Relation Object hardware shop at location mall world map has property draw with grid-lines PersonX receives its reward wants to keep the prize PersonX wins the big Jackpot wants to get its money Table 1: Commonsense knowledge base tuples. Exam- ples are from ConceptNet and Atomic. Although KGC methods can automatically find un- labeled relationships, they are always classification or ranking tasks and are limited to existing entities in a knowledge graph and can’t extend to new en- tities (Ji et al., 2022). To extend to new entities, COMET (Bosselut et al., 2019) proposes to use text generation for exploring and discovering new entities, which is called commonsense knowledge base completion (CKBC) task, utilizing the knowl- edge within the pretrained language models (PLM), which has been with process in recent years (De- vlin et al., 2019; Radford et al., 2019). Specifically, COMET uses subject and relation as direct input to PLM and aims to generate objects, most of which are …

Content truncated to 500 words