DHA: Product Title Generation with Discriminative Hierarchical Attention for E-commerce

Product titles play an important role in E-Commerce sites. However, manually crafting product titles needs tremendous time and human effort. It is expected that product titles can be automatically generated, but existing generation methods usually require densely labeled data that are unavailable in the real world. We formulate a novel product title generation task that generates the title from the image and auxiliary information (e.g., category) to address the gap. To generate titles that are consistent with search queries, we construct the first large-scale dataset (AEPro) and propose a Discriminative Hierarchical Attention (DHA) model. The DHA model first identifies the image regions related to the product of interest (POI) with a POI attention module. Then, based on the title context, the identified image regions are further revised by a generation attention module. Finally, the titles are generated by dynamically attending to these image regions. Experiments on the AEPro dataset demonstrate the effectiveness of the DHA model. Besides, online A/B testing results show that 61.8% of the titles generated by the DHA model are accepted directly or with minor modifications. The exposure rate of the products with machine-generated titles is improved by 40%.