Towards Comprehensive Description Generation from Factual Attribute-value Tables

Abstract

The comprehensive descriptions for factual attribute-value tables, which should be accurate, informative and loyal, can be very helpful for end users to understand the structured data in this form. However previous neural generators might suffer from key attributes missing, less informative and groundless information problems, which impede the generation of high-quality comprehensive descriptions for tables. To relieve these problems, we first propose force attention (FA) method to encourage the generator to pay more attention to the uncovered attributes to avoid potential key attributes missing. Furthermore, we propose reinforcement learning for information richness to generate more informative as well as more loyal descriptions for tables. In our experiments, we utilize the widely used WIKIBIO dataset as a benchmark. Additionally we create WB-filter based on WIKIBIO to test our model in the simulated user-oriented scenarios, in which the generated descriptions should accord with particular user interests. Experimental results show that our model outperforms the state-of-the-art baselines on both automatic and human evaluation.

Sketch and Refine: Towards Faithful and Informative Table-to-Text Generation

“…Human Evaluation Human ratings on the generated descriptions provide more reliable reflection of the model performance. Following Liu et al (2019b), we conduct comprehensive human evaluation between our model and the baselines. The annotators are asked to evaluate from three perspectives: fluency, coverage (how much table content is recovered) and correctness (how much generated content is faithful to the source table).…”

Towards Faithful Neural Table-to-Text Generation with Content-Matching Constraints

“…Other related works are (Perez-Beltrachini and Lapata, 2018;Liu et al, 2019b). For (Perez-Beltrachini and Lapata, 2018), the content selection mechanism training with multi-task learning and reinforcement learning is proposed.…”

“…For (Perez-Beltrachini and Lapata, 2018), the content selection mechanism training with multi-task learning and reinforcement learning is proposed. For (Liu et al, 2019b), they propose force attention and reinforcement learning based method. Their learning methods are completely different from our method that simultaneously incorporates optimaltransport matching loss and embedding similarity loss.…”

“…Their learning methods are completely different from our method that simultaneously incorporates optimaltransport matching loss and embedding similarity loss. Moreover, the REINFORCE algorithm (Williams, 1992) and policy gradient method used in (Perez-Beltrachini and Lapata, 2018;Liu et al, 2019b) exhibits high variance when training the model.…”

Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing

“…Among the baselines shown in Table A4, SQLova is the one that's strictly comparable to BRIDGE as both use BERT-large-uncased. Hydra-Net uses RoBERTa-Large (Liu et al, 2019a) and X-SQL uses MT-DNN (Liu et al, 2019b).…”