Table-based Fact Verification with Self-labeled Keypoint Alignment

Proceedings of the 29th International Conference on Computational Linguistics, pages 1401–1411 October 12–17, 2022. 1401 Table-based fact verification with self-labeled keypoint alignment Guangzhen Zhao†, Peng Yang†∗ † School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China {zhaogz, pengyang}@seu.edu.cn Abstract Table-based fact verification aims to verify whether a statement sentence is trusted or fake. Most existing methods rely on graph feature or data augmentation but fail to inves- tigate evidence correlation between the state- ment and table effectively. In this paper, we propose a self-Labeled Keypoint Alignment model, named LKA, to explore the correla- tion between the two. Specifically, a dual- view alignment module based on the state- ment and table views is designed to discrimi- nate the salient words through multiple inter- actions, where one regular and one adversari- al alignment network cooperatively character the alignment discrepancy. Considering the interaction characteristic inherent in the align- ment module, we introduce a novel mixture-of- experts block to elaborately integrate the inter- acted information for supporting the alignmen- t and final classification. Furthermore, a con- trastive learning loss is utilized to learn the pre- cise representation of the structure-involved words, encouraging the words closer to words with the same table attribute and farther from the words with the unrelated attribute. Experi- mental results on three widely-studied datasets show that our model can outperform the state- of-the-art baselines and capture interpretable evidence words. 1 Introduction Table-based fact verification aims to uncover the factuality attribute of the sentence relying on the available structured (Chen et al., 2020b; Wang et al., 2021b; Gupta et al., 2020) textual evidence. Current methods can be divided into two group- s. The first one exploits the logical form of the statement with graph neural networks (Zhong et al., 2020; Shi et al., 2020, 2021). The other focuses on extending table-aware pre-trained language mod- els (PLMs) (Eisenschlos et al., 2020; Herzig et al., 2020). ∗Corresponding author. Label: Entailed czech republic nation 2 belgium switzerland italy denmark 1 0 0 0 silver 2 0 1 0 0 bronze 1 0 0 0 1 total 5 1 1 1 1 Statement:3 nation win no gold medal at the 1998 uci cyclo - cross world championship. Title:1998 uci cyclo - cross world championships 2018 Winter Olympics title 30 Goals scored Attendance 154 (5.13 per match) 138,327 (4,611 per match) Label: Contradition Statement: 38 matches were played. (a) one example in TABFACT (b) one example in INFOTABS gold Matches played Figure 1: Two examples of table-based fact verification task. Keypoints are highlighted in yellow. (a) is a rela- tional web table. (b) is an entity web table. In PLM-oriented fact verification, the majority of methods treat the statement-table pair as plain text and then further capture latent essential infor- mation relying on multiple Transformer (Vaswani et al., 2017) blocks. It is intuitive that only partial table cells are associated with the statement while other cells are redundant (Wang et al., 2021a; Yin et al., …

Content truncated to 500 words