Neural Data-to-Text Generation via Jointly Learning the Segmentation and Correspondence

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7155–7165 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7155 Neural Data-to-Text Generation via Jointly Learning the Segmentation and Correspondence Xiaoyu Shen1,2 , Ernie Chang3, Hui Su4, Cheng Niu4 and Dietrich Klakow1 1Spoken Language Systems (LSV) 2Max Planck Institute for Informatics 3Department of Language Science and Technology, Saarland Informatics Campus 4Pattern Recognition Center, Wechat AI, Tencent Inc, China xshen@mpi-inf.mpg.de,cychang@coli.uni-saarland.de {aaronsu,niucheng}@tencent.com,dklakow@lsv.uni-saarland.de Abstract The neural attention model has achieved great success in data-to-text generation tasks. Though usually excelling at producing fluent text, it suffers from the problem of information missing, repetition and “hallucination”. Due to the black-box nature of the neural attention architecture, avoiding these problems in a sys- tematic way is non-trivial. To address this con- cern, we propose to explicitly segment target text into fragment units and align them with their data correspondences. The segmentation and correspondence are jointly learned as la- tent variables without any human annotations. We further impose a soft statistical constraint to regularize the segmental granularity. The re- sulting architecture maintains the same expres- sive power as neural attention models, while being able to generate fully interpretable out- puts with several times less computational cost. On both E2E and WebNLG benchmarks, we show the proposed model consistently outper- forms its neural attention counterparts. 1 Introduction Data-to-text generation aims at automatically pro- ducing natural language descriptions of structured database (Reiter and Dale, 1997). Traditional sta- tistical methods usually tackle this problem by breaking the generation process into a set of local decisions that are learned separately (Belz, 2008; Angeli et al., 2010; Kim and Mooney, 2010; Oya et al., 2014). Recently, neural attention models con- flate all steps into a single end-to-end system and largely simplify the training process (Mei et al., 2016; Lebret et al., 2016; Shen et al., 2017; Su et al., 2018, 2019; Chang et al., 2020). However, the black-box conflation also renders the genera- tion uninterpretable and hard to control (Wiseman et al., 2018; Shen et al., 2019a). Verifying the generation correctness in a principled way is non- trivial. In practice, it often suffers from the problem of information missing, repetition and “hallucina- tion” (Duˇsek et al., 2018, 2020). Source data: Name[Clowns], PriceRange[more than £30], EatType[pub], FamilyFriendly[no] Generation: 1⃝Name → 2⃝(Clowns) 3⃝FamilyFriendly → 4⃝(is a child-free) 5⃝PriceRange → 6⃝(, expensive) 7⃝EatType → 8⃝(pub.) Figure 1: Generation from our model on the E2E dataset. Decoding is performed segment-by-segment. Each segment realizes one data record. 1⃝˜ 8⃝ mark the decision order in the generation process. In this work, we propose to explicitly exploit the segmental structure of text. Specifically, we assume the target text is formed from a sequence of segments. Every segment is the result of a two- stage decision: (1) Select a proper data record to be described and (2) Generate corresponding text by paying attention only to the selected data record. This decision is repeated until all desired records have been realized. Figure 1 illustrates this process. …

Content truncated to 500 words