Learn to Combine Linguistic and Symbolic Information for Table-based Fact Verification

Proceedings of the 28th International Conference on Computational Linguistics, pages 5335–5346 Barcelona, Spain (Online), December 8-13, 2020 5335 Learn to Combine Linguistic and Symbolic Information for Table-based Fact Verification Qi Shi, Yu Zhang, Qingyu Yin, Ting Liu Department of Computer Science and Technology, Harbin Institute of Technology, Harbin, China {qshi,zhangyu,qyyin,tliu}@ir.hit.edu.cn Abstract Table-based fact verification is expected to perform both linguistic reasoning and symbolic rea- soning. Existing methods lack attention to take advantage of the combination of linguistic infor- mation and symbolic information. In this work, we propose HeterTFV, a graph-based reasoning approach, that learns to combine linguistic information and symbolic information effectively. We first construct a program graph to encode programs, a kind of LISP-like logical form, to learn the semantic compositionality of the programs. Then we construct a heterogeneous graph to in- corporate both linguistic information and symbolic information by introducing program nodes into the heterogeneous graph. Finally, we propose a graph-based reasoning approach to reason over the multiple types of nodes to make an effective combination of both types of information. Experimental results on a large-scale benchmark dataset TABFACT illustrate the effect of our approach. 1 Introduction Fact verification aims to verify whether a fact is entailed or refuted by the given evidence, which has attracted increasing attention. Recent researches mainly focus on the unstructured text as the evidence and ignoring the evidence with the structured or semi-structured format. A recently proposed dataset TABFACT (Wenhu Chen and Wang, 2020) fills this gap, which is designed to deal with the table-based fact verification problem, namely, verifying whether a statement is correct by the given semi-structured table evidence. It is well accepted that symbolic information (such as count and only) plays a great role in understand- ing semi-structured evidence based statements (Wenhu Chen and Wang, 2020). However, most existing approaches for fact verification (Thorne et al., 2018; Nie et al., 2019; Zhou et al., 2019; Liu et al., 2020; Zhong et al., 2020b; Soleimani et al., 2020) focus on the understanding of natural language, namely, linguistic reasoning, but fail to consider symbolic information, which plays an important role in com- plex reasoning (Liang et al., 2017; Dua et al., 2019; Chen et al., 2019). Due to the diversity of natural language expressions, it is difficult to capture symbolic information effectively from natural language directly. Consequently, how to leverage symbolic information effectively becomes a crucial problem. To alleviate this problem, Zhong et al. (2020a) propose a graph module network that concatenates graph- enhanced linguistic-level representations and program-guided symbolic-level representations together to predict the labels. However, their method focuses on the representation of symbolic information, rather than take advantage of the combination of both types of information. More specifically, we believe that the concatenation operation between two types of representations is not effective enough to leverage the linguistic information and symbolic information to perform reasoning. In recent studies, graph neural networks show their powerful ability in dealing with semi-structured data (Bogin et al., 2019a; Bogin et al., 2019b). Under this consideration, we …

Content truncated to 500 words