Harvest shopping advice: Neural Question Generation from multiple information sources in E-commerce

Abstract

The success of recent efforts in Question Generation (QG) has amazed scientists from academia and industry. In this paper, we explore to harvest shopping advice through a novel QG engine for e-commerce platforms. Unlike traditional QG methods conditioned on factual data, generating purchase-oriented questions depends on open-ended product properties and customer reviews. Besides, these questions should follow not only natural expressions but also user-interested aspects simultaneously. For this challenging task, an innovative generative adversarial net-based QG model is proposed – a generator featuring multi-source attention mechanism is employed to yield questions from multiple information sources; a discriminator featuring quality control is applied to fine-tune generated questions in terms of both language performance and aspect compatibility. We conduct extensive experiments on a new dataset comprised of Question-Review-Aspect-Property (Q-RAP) tuples from a real e-commerce site. Our experimental results demonstrate that the proposed approach achieves a significant superiority over seven state-of-the-art QG solutions. Meanwhile, this study indicates that customer reviews play a critical role in generating purchase-oriented questions, which confirms the validity of previous practices using buyer feedback to address natural language generation in e-commerce.