Learning to Generate Programs for Table Fact Verification via Structure-Aware Semantic Parsing

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7624 - 7638 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learning to Generate Programs for Table Fact Verification via Structure-Aware Semantic Parsing Suixin Ou, Yongmei Liu∗ Dept. of Computer Science, Sun Yat-sen University, Guangzhou 510006, China ousx@mail2.sysu.edu.cn, ymliu@mail.sysu.edu.cn Abstract Table fact verification aims to check the cor- rectness of textual statements based on given semi-structured data. Most existing methods are devoted to better comprehending logical op- erations and tables, but they hardly study gen- erating latent programs from statements, with which we can not only retrieve evidences ef- ficiently but also explain reasons behind veri- fications naturally. However, it is challenging to get correct programs with existing weakly supervised semantic parsers due to the huge search space with lots of spurious programs. In this paper, we address the challenge by leverag- ing both lexical features and structure features for program generation. Through analyzing the connection between the program tree and the dependency tree, we define a unified concept, operation-oriented tree, to mine structure fea- tures, and introduce Structure-Aware Seman- tic Parsing to integrate structure features into program generation. Moreover, we design a refined objective function with lexical features and violation punishments to further avoid spu- rious programs. Experimental results show that our proposed method generates programs more accurately than existing semantic parsers, and achieves comparable performance to the SOTA on the large-scale benchmark TABFACT. 1 Introduction With the rise of misleading information on the Inter- net, such as fake news, rumors and political deceit, fact-checking has been developed as a means of detecting and filtering false information. Table fact verification (TFV) is a specific fact-checking task that requires performing logical operations such as comparison, superlative and aggregation over given tables to verify textual statements. Programs play an important role in TFV. On one hand, correct programs can provide rationales for model decisions, which make reasoning analysis *Corresponding author Season Podiums … 1980 9 … 1981 8 … 1981 0 … … … ... Table Statement In 1981 season, the highest number of podiums was 8 and the lowest was 0. and v2:eq v4:eq v1:max 8 v3:min 0 v0:filter_eq podiums all_rows season 1981 Program Program generation V0 The table where column season equal to 1981 is [row1, row2]. V1 The max value of podiums in [row1, row2] is 8. V2 8 is equal to 8. V3 The min value of podiums in [row1, row2] is 0. V4 0 is equal to 0. Verbalized Evidence Evidence Accumulation Final Prediction ENTAILED REFUTED and(eq(max(filter_eq(all_rows,season,1981),podiums),8), eq(min(filter_eq(all_rows,season,1981),podiums),0)) Figure 1: The pipeline of ProgVGAT (Yang et al., 2020) on TFV. Here the task is, given a table and a statement, to predict whether the table entails the statement or refutes it. Verbalized evidences are verbal descriptions of the program execution procedure. and failure diagnosis feasible (Zhou et al., 2018). On the other hand, they can be used to fetch the key evidences for verification. Figure 1 gives an exam- …

Content truncated to 500 words