SILVER: Generating Persuasive Chinese Product Pitch

Abstract

Building a silver-tongued salesbot is attractive and profitable. The first and pivotal step is to generate a product pitch, which is a short piece of persuasive text which both convey product information and deliver persuasive explanations related to customer demand. Recent advances in deep neural networks have empowered text generation systems to produce natural language descriptions of products. However, to produce persuasive product pitches, deep neural networks need to be fed with massive amounts of persuasive samples, which are not available due to huge labelling cost. This paper proposes SILVER, a persuasive Chinese product pitch generator, which addresses the issue of insufficient labeled data with data-level, knowledge-level and model-level solutions. At the data level, SILVER employs statistic analysis to automatically derive weak supervision rules that correlate with persuasive texts. At the model level, SILVER apply the weak supervision rules to re-rank outputs from an ensemble of models to enhance pitch generation performance. Finally, at the knowledge level, SILVER incorporates attribute hierarchy to embed product information in the pitch. Both automatic and human-involved evaluations on real data demonstrate that SILVER is able to produce more fluent, catchy and informative snippets than state-of-the-art text generation approaches.