LoFT: Enhancing Faithfulness and Diversity for Table-to-Text Generation via Logic Form Control

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 554–561 May 2-6, 2023 ©2023 Association for Computational Linguistics LOFT: Enhancing Faithfulness and Diversity for Table-to-Text Generation via Logic Form Control Yilun Zhao∗ 1 Zhenting Qi∗2 Linyong Nan1 Lorenzo Jaime Yu Flores1 Dragomir Radev1 1Yale University 2 Zhejiang University yilun.zhao@yale.edu zhenting.19@intl.zju.edu.cn Abstract Logical Table-to-Text (LT2T) generation is tasked with generating logically faithful sen- tences from tables. There currently exists two challenges in the field: 1) Faithfulness: how to generate sentences that are factually cor- rect given the table content; 2) Diversity: how to generate multiple sentences that offer dif- ferent perspectives on the table. This work proposes LOFT, which utilizes logic forms as fact verifiers and content planners to con- trol LT2T generation. Experimental results on the LOGICNLG dataset demonstrate that LOFT is the first model that addresses unfaith- fulness and lack of diversity issues simultane- ously. Our code is publicly available at https: //github.com/Yale-LILY/LoFT. 1 Introduction Table-to-Text (T2T) generation aims to produce natural language descriptions from structured ta- bles. A statement generated from tabular data can be inferred based on different levels of information (e.g., value of a specific cell, logical operation re- sult across multiple cells). Although current T2T models (Lebret et al., 2016; Wiseman et al., 2017; Puduppully et al., 2019; Parikh et al., 2020) have shown remarkable progress in fluency and coher- ence, they mainly focus on surface-level realiza- tions without much logical inference. Recently, Chen et al. (2020a) proposed LOG- ICNLG, which is tasked with generating textual descriptions that require logical reasoning over tab- ular data (i.e., LT2T generation). LT2T genera- tion is challenging as it requires a model to learn the logical inference knowledge from table-text pairs and generate multiple factually correct sen- tences. Another challenge for LT2T generation is the diversity of generated text. Natural Language Generation (NLG) encourages the diverse output of statements over a single input, as it provides ∗Equal Contributions. The average earnings of Hale Irwin and Dana Quigley is 19,663,467 Five of the players are from the same country, United States Dana Quigley had less wins than Larry Nelson Most of the players had earnings less than 18,964,040 Larry Nelson had the least number of earnings Hale Irwin and Gil Morgan represent the same country Hale Irwin had more wins than Gil Morgan Hale Irwin had more earnings than Gil Morgan Hale Irwin had the more wins than Bruce Fleisher Bruce Fleisher had the highest earnings of any player with 13,990,356 1 Hale Irwin United States 24,920,665 45 1 Hale Irwin 2 Gil Morgan United States 18,964,040 25 2 Gil Morgan 3 Dana Quigley United States 14,406,269 11 3 Dana Quig 4 Bruce Fleisher United States 13,990,356 18 4 Bruce Fleis 5 Larry Nelson United States 13,262,808 19 5 Larry Nelso 1 Hale Irwin United States 24,920,665 45 2 Gil Morgan United States 18,964,040 25 3 Dana Quigley United States 14,406,269 11 4 Bruce Fleisher United States 13,990,356 18 5 Larry Nelson United …

Content truncated to 500 words