Generating Formality-Tuned Summaries Using Input-Dependent Rewards

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 833–842 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 833 Generating Formality-tuned Summaries Using Input-dependent Rewards Kushal Chawla∗ University of Southern California Los Angeles, CA, USA kchawla@usc.edu Balaji Vasan Srinivasan, Niyati Chhaya Adobe Research Bangalore, India balsrini,nchhaya@adobe.com Abstract Abstractive text summarization aims at gen- erating human-like summaries by understand- ing and paraphrasing the given input content. Recent efforts based on sequence-to-sequence networks only allow the generation of a single summary. However, it is often desirable to ac- commodate the psycho-linguistic preferences of the intended audience while generating the summaries. In this work, we present a rein- forcement learning based approach to generate formality-tailored summaries for an input arti- cle. Our novel input-dependent reward func- tion aids in training the model with stylistic feedback on sampled and ground-truth sum- maries together. Once trained, the same model can generate formal and informal summary variants. Our automated and qualitative eval- uations show the viability of the proposed framework. 1 Introduction Efficient content consumption is not only driven by the contained information but also by the tone/style of the presentation. The style in text is its non-informational or non-factual aspect and usually drives the quality of response from its au- dience. A persuasive snippet or a teaser might drive up sales for a marketing message and a piece of formal text will appeal better to corporate ex- ecutives as against informal communication. Sim- ilarly, not all long form content is easy to read. A succinct representation of the content, i.e. the summary, plays an important role for its quick and efficient consumption. While, text summariza- tion and to some extent style-understanding have been independently studied, approaches to gener- ate style-tailored summaries are limited. ∗ Work done when author was a full time employee at Adobe Research Models for style predictions (Pavlick and Tetreault, 2016; Brooke et al., 2010; Danescu- Niculescu-Mizil et al., 2013) are limited to mea- suring style in text. There have been multi- ple attempts towards transfer of style (Artetxe et al., 2017; Han et al., 2017; Shen et al., 2017; Tikhonov and Yamshchikov, 2018). Rao and Tetreault (2018) introduced a parallel corpus for formality style transfer with neural machine translation benchmarks. Niu et al. (2017) gener- ate automatic translations tailored towards formal- ity. However, none of these approaches account for the length/succinctness of the created content, and hence do not address the stylized summariza- tion task. In this work, we propose an approach to generate summaries while simultaneously tai- loring towards formality preferences (Table 1). Tunable or controlled summary generation has picked up pace in recent times. Algorithms al- low for controlling various dimensions of the out- put summary such as the length or entities (Fan et al., 2017) and topics (Krishna and Srinivasan, 2018). Since these approaches primarily rely on the diversity in the given dataset, extending these approaches for formality tailored summa- rization would require a diverse summarization corpus that captures subtleties in various formal …

Content truncated to 500 words