The article presents ConstraintChecker, a plugin designed to enhance the reasoning capabilities of Large Language Models (LLMs) over Commonsense Knowledge Bases (CSKBs). Despite the significant advances in LLMs, they still struggle with CSKB reasoning due to their inability to acquire explicit relational constraints. ConstraintChecker addresses this issue by providing and checking explicit constraints. It uses a rule-based module to generate constraints and a zero-shot learning module to verify if a knowledge instance meets all the constraints. The final output is produced by combining the constraint-checking result with the output of the main prompting technique. Experimental results demonstrate the effectiveness of ConstraintChecker in improving the performance of all prompting methods.
Publication date: 26 Jan 2024
Project Page: https://github.com/HKUST-KnowComp/ConstraintChecker
Paper: https://arxiv.org/pdf/2401.14003
