Enhancing Content Planning for Table-to-Text Generation with Data Understanding and Verification

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2905–2914 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2905 Enhancing Content Planning for Table-to-Text Generation with Data Understanding and Verification Heng Gong,1 Wei Bi, Xiaocheng Feng,1 Bing Qin,1 Xiaojiang Liu, Ting Liu1 1Department of Computer Science and Technology, Harbin Institute of Technology, Harbin, China {hgong,xcfeng,qinb,tliu}@ir.hit.edu.cn weibi@cse.ust.hk xiaojiangliu84@hotmail.com Abstract Neural table-to-text models, which select and order salient data, as well as verbalizing them fluently via surface realization, have achieved promising progress. Based on results from pre- vious work, the performance bottleneck of cur- rent models lies in the stage of content plan- ning (selecting and ordering salient content from the input). That is, performance drops drastically when an oracle content plan is re- placed by a model-inferred one during surface realization. In this paper, we propose to en- hance neural content planning by (1) under- standing data values with contextual numeri- cal value representations that bring the sense of value comparison into content planning; (2) verifying the importance and ordering of the selected sequence of records with policy gradi- ent. We evaluated our model on ROTOWIRE and MLB, two datasets on this task, and results show that our model outperforms existing sys- tems with respect to content planning metrics. 1 Introduction Table-to-text generation refers to the task of gener- ating text from structured data. Models for this task can be mainly categorized into two types: pipeline-style models, which decompose the gen- eration process into sequential stages, including content planning (Stage 1, selecting and ordering salient content from the input) and surface real- ization (Stage 2, converting the content plan to surface string) (Kukich, 1983; McKeown, 1985); and end-to-end models, which entangle aforemen- tioned stages and generate text directly from struc- tured data through a neural encoder-decoder frame- work (Wiseman et al., 2017; Nie et al., 2018). As in Fig. 1, this task provides tables with redundant records. Each record has three elements: table row header (entity, e.g. Conley), table column header (type, e.g. points) and table cell (value, e.g. 32). Models are expected to generate descriptive text Player PTS AST REB BLK STL Nets LeVert 14 1 3 0 0 Lopez 8 2 4 0 1 Foye 14 1 3 0 2 Lin 18 1 4 0 0 Booker 8 1 9 1 1 Kilpatrick 23 3 5 1 1 NCP+CC (NCP) Gold The Memphis Grizzlies defeated the Brooklyn Nets , 109 - 122 , at FedEx Forum on Friday . The Grizzlies ( 36 - 28 ) came into this game as a huge favorite and they left no doubt with this result . In fact , there were 10 ties and 11 lead changes in total , with no team leading by more than 12 points . Three-point shooting was the difference in the game , with the Grizzlies going 53 percent from the field and 47 percent from three - point range , while the Grizzlies went 53 percent from the field and 47 percent …

Content truncated to 500 words