Robust and resource-efficient table-based fact verification through multi-aspect adversarial contrastive learning

Abstract

Table-based fact verification focuses on determining the truthfulness of statements by cross-referencing data in tables. This task is challenging due to the complex interactions inherent in table structures. To address this challenge, existing methods have devised a range of specialized models. Although these models demonstrate good performance, they still exhibit limited sensitivity to essential variations in information relevant to reasoning within both statements and tables, thus learning spurious patterns and leading to potentially unreliable outcomes. In this work, we propose a novel approach named Multi-Aspect Adversarial Contrastive Learning (Macol), aimed at enhancing the accuracy and robustness of table-based fact verification systems under the premise of resource efficiency. Specifically, we first extract pivotal logical reasoning clues to construct positive and adversarial negative instances for contrastive learning. We then propose a new training paradigm that introduces a contrastive learning objective, encouraging the model to recognize noise invariance between original and positive instances while also distinguishing logical differences between original and negative instances. Extensive experimental results on three widely studied datasets TABFACT, INFOTABS and SEM-TAB-FACTS demonstrate that Macol achieves state-of-the-art performance on benchmarks across various backbone architectures, with accuracy improvements reaching up to 5.4%. Furthermore, Macol exhibits significant advantages in robustness and low-data resource scenarios, with improvements of up to 8.2% and 9.1%. It is worth noting that our method achieves comparable or even superior performance while being more resource-efficient compared to approaches that employ specific additional pre-training or utilize large language models (LLMs).