Data-to-Text Generation with Style Imitation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1589–1598 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1589 Data-to-Text Generation with Style Imitation Shuai Lin1,2, Wentao Wang2, Zichao Yang2, Xiaodan Liang1∗, Frank F. Xu2 Eric P. Xing2,3, Zhiting Hu2,4∗ 1Sun Yat-sen University, 2Carnegie Mellon University, 3Petuum Inc., 4UC San Diego {shuailin97,xdliang328,zhitinghu}@gmail.com, {zichaoy,fangzhex,epxing}@cs.cmu.edu Abstract Recent neural approaches to data-to-text gen- eration have mostly focused on improving content fidelity while lacking explicit control over writing styles (e.g., word choices, sen- tence structures). More traditional systems use templates to determine the realization of text. Yet manual or automatic construction of high- quality templates is difficult, and a template acting as hard constraints could harm content fidelity when it does not match the record per- fectly. We study a new way of stylistic con- trol by using existing sentences as “soft” tem- plates. That is, the model learns to imitate the writing style of any given exemplar sentence, with automatic adaptions to faithfully describe the content record. The problem is challeng- ing due to the lack of parallel data. We de- velop a neural approach that includes a hybrid attention-copy mechanism, learns with weak supervisions, and is enhanced with a new con- tent coverage constraint. We conduct experi- ments in restaurants and sports domains. Re- sults show our approach achieves stronger per- formance than a range of comparison methods. Our approach balances well between content fidelity and style control given exemplars that match the records to varying degrees.1 1 Introduction Recent years have seen remarkable progress in neural natural language generation to produce well-formed coherent text (Sutskever et al., 2014; Vaswani et al., 2017). Yet, controllability over var- ious text properties, as an essential demand to en- sure the utility of generations in real-world appli- cations, has not attained the same level of advance- ment. Data-to-text generation is one of such ap- plications with ubiquitous practical use, in which ∗corresponding authors 1Data and code are publicly available at https:// github.com/ha-lins/DTG-SI natural language text is generated to describe a given data record such as a box score of a sports player or an infobox table of a restaurant. Though current data-to-text neural approaches with encoder-decoder models could produce fluent text with high fidelity to content (“what to say”), they largely lack control over the writing style, such as sentence structures and word choices (“how to say”). Many efforts have been made to promote the overall diversity in data-to-text generation through, e.g., latent variables (Ye et al., 2020) or customized model architectures (Jagfeld et al., 2018; Deriu and Cieliebak, 2018). Yet fine-grained style manipu- lation is not permitted. This contrasts with the traditional text generation systems which separate content planning and surface realization (Reiter and Dale, 1997), and usually determine the realiza- tion with explicit templates (Kukich, 1983; McRoy et al., 2000) or based on syntactic grammars (Robin and McKeown, 1996; Power et al., 2003). Controlling writing style with “hard” templates could suffer from unscalable template creation and lack of generation flexibility. …

Content truncated to 500 words