Selling products by machine- a user-sensitive adversarial training method for short title generation in mobile e-commerce

Tao Zhang

Abstract: In E-commerce portals, merchants don’t have the ability to write various titles for different customers, so they usually tend to write lengthy product titles to get most buyers’ attention. It is a crucial task to extract relevant keywords for displaying on the limited screen of mobile phones to fetch customers. Previous studies mainly focus on content-based methods, however, lack of user features may result in the generated titles containing rich product information while ignoring the user requirement. In this paper, we propose a Personalized Pointer Generative Adversarial Network (PPGAN) to generate personalized user-sensitive short titles. Even though in our sparse dataset the user-clicked data is limited due to the low Click-Through-Rate (CTR), our model could encourage the discriminator to identify the high-quality short titles from the user-unclicked data by employing an unsupervised information-theoretic assignment strategy. An extensive set of experiments on a large-scale E-commerce dataset indicates the advantage of the proposed method for generation qualities. Finally, we deploy our model into a real-word online E-commerce environment and boost the performance of CTR compared with state of art baseline methods.