PersuAIDE ! An Adaptive Persuasive Text Generation System for Fashion Domain

Abstract

Persuasiveness is a creative art which aims at inducing certain set of beliefs in the target audience. In an e-commerce setting, for a newly launched product, persuasive descriptions are often composed to motivate an online buyer towards a successful purchase. Such descriptions can be catchy taglines, product-summaries, style-tipsetc.. In this paper, we present PersuAIDE! - a persuasive system based on linguistic creativity to generate various forms of persuasive sentences from the input product specification. To demonstrate the effectiveness of the proposed system, we have applied the technology to fashion domain, where, for a given fashion product like"red collar shirt" we were able to generate descriptive sentences that not only explain the item but also garner positive attention, making it persuasive. PersuAIDE! identifies fashion related keywords from input specifications and intelligently expands the keywords to creative phrases. Once such compatible phrases are obtained, persuasive descriptions are synthesized from the set of phrases and input keywords with the help of a neural language model trained on a large domain-specific fashion corpus. We evaluate the system on a large fashion corpus collected from different sources using (a) automatic text generation metrics used for Machine Translation and Automatic Summarization evaluation and Readability measurement, and (b) human judgment scores evaluating the persuasiveness and fluency of the generated text. Experimental results and qualitative analysis show that an unsupervised system like ours can produce more creative and better constructed persuasive output than supervised generative counterparts based on neural sequence-to-sequence models and statistical machine translation.