UniRPG: Unified Discrete Reasoning over Table and Text as Program Generation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7494 - 7507 December 7-11, 2022 ©2022 Association for Computational Linguistics UniRPG: Unified Discrete Reasoning over Table and Text as Program Generation Yongwei Zhou1, Junwei Bao2∗, Chaoqun Duan2, Youzheng Wu2, Xiaodong He2, Tiejun Zhao1 1Harbin Institute of Technology 2JD AI Research ywzhou@hit-mtlab.net baojunwei001@gmail.com tjzhao@hit.edu.cn Abstract Question answering requiring discrete reason- ing, e.g., arithmetic computing, comparison, and counting, over knowledge is a challeng- ing task. In this paper, we propose UniRPG, a semantic-parsing-based approach advanced in interpretability and scalability, to perform Unified discrete Reasoning over heterogeneous knowledge resources, i.e., table and text, as Program Generation. Concretely, UniRPG con- sists of a neural programmer and a symbolic program executor, where a program is the com- position of a set of pre-defined general atomic and higher-order operations and arguments ex- tracted from table and text. First, the program- mer parses a question into a program by gen- erating operations and copying arguments, and then, the executor derives answers from table and text based on the program. To alleviate the costly program annotation issue, we design a distant supervision approach for programmer learning, where pseudo programs are automati- cally constructed without annotated derivations. Extensive experiments on the TAT-QA dataset show that UniRPG achieves tremendous im- provements and enhances interpretability and scalability compared with previous state-of-the- art methods, even without derivation annota- tion. Moreover, it achieves promising perfor- mance on the textual dataset DROP without derivation annotation. 1 1 Introduction Question answering requiring discrete reasoning, e.g., arithmetic computing, sorting, comparison, and counting, over knowledge is a challenging task and has attracted wide attention. Advanced approaches to this task are roughly divided into mixture-of-experts (MoE)-based and semantic pars- ing (SP)-based methods. The former divides ques- tions into several limited categories according to ∗Work was done during the first author’s internship at JD AI mentored by Junwei Bao: baojunwei001@gmail.com. 1The code is released at https://github.com/ JD-AI-Research-NLP/UniRPG. P : (1) Note 14 – Earnings (loss) per share. (2) Basic earnings (loss) per share is computed by ... Q : What is the change in Basic earnings (loss) per share between 2018 and 2019? A : 0.06 D : {Solution: 0.53-0.47, Table: [(3,1), (3,2)]} 2019 2018 2017 Net income (loss) 18398 16617 7904 Basic weighted average number of common shares outstanding 34589 35250 35917 Basic earnings (loss) per share 0.53 0.47 0.22 Diluted 0.52 0.46 0.22 w/ derivation: G : DIFF(CV(131, 134), CV(135, 138)) w/o derivation: G : DIFF(CV(131, 134), CV(135, 138)) DIFF(CV(148, 151), CV(152, 155)) T : Figure 1: The upper block shows a training instance from the TAT-QA dataset. It consists of a table T , ques- tion Q, paragraphs P, answer A and derivation D. The bottom block displays the constructed programs G under w/ and w/o derivation settings. The gold program is in red, and the others are in green. their required reasoning type and designs multi- ple specialized predictors to deal with each cate- gory (Dua et al., 2019; Ran et …

Content truncated to 500 words