Robust (Controlled) Table-to-Text Generation with Structure-Aware Equivariance Learning

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5037 - 5048 July 10-15, 2022 ©2022 Association for Computational Linguistics Robust (Controlled) Table-to-Text Generation with Structure-Aware Equivariance Learning Fei Wang, Zhewei Xu, Pedro Szekely and Muhao Chen Department of Computer Science & Information Sciences Institute University of Southern California {fwang598,zheweixu,szekely,muhaoche}@usc.edu Abstract Controlled table-to-text generation seeks to generate natural language descriptions for high- lighted subparts of a table. Previous SOTA systems still employ a sequence-to-sequence generation method, which merely captures the table as a linear structure and is brittle when table layouts change. We seek to go beyond this paradigm by (1) effectively expressing the relations of content pieces in the table, and (2) making our model robust to content-invariant structural transformations. Accordingly, we propose an equivariance learning framework, LATTICE ( ), which encodes tables with a structure-aware self-attention mechanism. This prunes the full self-attention structure into an order-invariant graph attention that captures the connected graph structure of cells belonging to the same row or column, and it differentiates between relevant cells and irrelevant cells from the structural perspective. Our framework also modifies the positional encoding mechanism to preserve the relative position of tokens in the same cell but enforce position invariance among different cells. Our technology is free to be plugged into existing table-to-text genera- tion models, and has improved T5-based mod- els to offer better performance on ToTTo and HiTab. Moreover, on a harder version of ToTTo, we preserve promising performance, while pre- vious SOTA systems, even with transformation- based data augmentation, have seen significant performance drops.1 1 Introduction Table-to-text generation seeks to generate natu- ral language descriptions for content and entailed conclusions in tables. It is an important task that not only makes ubiquitous tabular data more dis- coverable and accessible, but also supports down- stream tasks of tabular semantic retrieval (Wang et al., 2021a), reasoning (Gupta et al., 2020), fact 1Our code is available at https://github.com/ luka-group/Lattice. Year 1992 Role Sung Sai Kit So Chan Wai Siu-bo Title Justice, My Foot! King of Beggars Royal Tramp Year Title Role 1992 Justice, My Foot! Sung Sai Kit Royal Tramp Wai Siu-bo King of Beggars So Chan In 1992, Stephen Chow played Sung Sai Kit in Justice, My Foot!, Wai Siu-bo, King of Beggars and So Chan in Royal Tramp. In 1992, Stephen Chow appeared as Sung Sai Kit in Justice, My Foot!, Wai Siu-bo in Royal Tramp and King of Beggars as So Chan. In 1992, Stephen Chow appeared in Royal Tramp, King of Beggars, Justice, My Foot!, Wai Siu-bo, Wai Siu-bo, and as Sung Sai Kit in So Chan. In 1992, Stephen Chow appeared as Sung Sai Kit in Justice, My Foot!, Wai Siu-bo in Royal Tramp and King of Beggars as So Chan. LATTICE (87.6 BLEU) T5 (60.7 BLEU) LATTICE (87.6 BLEU) T5 (1.3 BLEU) Stephen Chow Stephen Chow Figure 1: Description generation on content-equivalent tables with different layouts by T5 and LATTICE2. Cor- rect …

Content truncated to 500 words