Attention to describe products with attributes
Shuo Li, Kota Yamaguchi · 2017
In e-commerce environment, shop owners and advertisers give descriptive details of the product to attract potential customers. Can a computer vision technique recognize and describe the details of a product in the same way? In this paper, we study how the attention mechanism benefits in product phrase generation with attributes. We present a phrase generation model consisting of convolutional neural networks, recurrent neural networks, and the attention mechanism to look into the detail of the image. We construct attribute-rich phrases from metadata in Easy dataset that consist of an adjective, a material tag, and product category, and learn the model to describe products. Our empirical results suggest that our model improves the description quality in both machine-translation metric and human evaluation.