Improving Encoder by Auxiliary Supervision Tasks for Table-to-Text Generation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5979–5989 August 1–6, 2021. ©2021 Association for Computational Linguistics 5979 Improving Encoder by Auxiliary Supervision Tasks for Table-to-Text Generation Liang Li1,2, Can Ma1∗, Yinliang Yue1∗ and Dayong Hu3 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3Heilongjiang Network Space Research Center, Harbin 150010, China {liliang, macan, yueyinliang}@iie.ac.cn superhudayong@163.com Abstract Table-to-text generation aims at automatically generating natural text to help people conve- niently obtain salient information in tables. Al- though neural models for table-to-text have achieved remarkable progress, some problems are still overlooked. Previous methods can- not deduce the factual results from the entity’s (player or team) performance and the relations between entities. To solve this issue, we first build an entity graph from the input tables and introduce a reasoning module to perform rea- soning on the graph. Moreover, there are dif- ferent relations (e.g., the numeric size relation and the importance relation) between records in different dimensions. And these relations may contribute to the data-to-text generation. However, it is hard for a vanilla encoder to capture these. Consequently, we propose to utilize two auxiliary tasks, Number Ranking (NR) and Importance Ranking (IR), to super- vise the encoder to capture the different rela- tions. Experimental results on ROTOWIRE and RW-FG show that our method not only has a good generalization but also outperforms pre- vious methods on several metrics: BLEU, Con- tent Selection, Content Ordering. 1 Introduction Table-to-text generation is an essential task for text generation from structured data. It aims at auto- matically producing descriptive natural language text to help people obtain the salient information from the tables. Over the past several years, neu- ral text generation methods have made significant progress on this task. Lebret et al. (2016); Wiseman et al. (2017); Bao et al. (2018) view the input table as a record sequence and model it as a machine translation task. To generate text containing more salient and well-organized facts, Sha et al. (2018); Moryossef et al. (2019); Trisedya et al. (2020); ∗Corresponding authors: Can Ma, Yinliang Yue The Boston Celtics dominated the visiting New York Knicks, 115 - 87, on Friday night at TD Garden … Isaiah Thomas was huge for Boston ( 4 - 4 ) as he led the way offensively with 29 points on 9-of-17 shooting, in only 28 minutes... Avery Bradley and Marcus Smart both filled the stat sheet. Smart finished with 12 points, 10 assists, six rebounds and three steals, while Bradley notched 15 points, 10 rebounds, two assists and four steals… Kristaps Porzingis was the high-point man with 14 points, along with six rebounds and two blocks , in 23 minutes. Derrick Rose added 11 points , six assists and four rebounds , while the only other player to tally double-digits for the Knicks was Justin Holiday… … Team Player … VS PlayFor …

Content truncated to 500 words