Learning Aligned Cross-Modal and Cross-Product Embeddings for Generating the Topics of Shopping Needs

Abstract

The paper addresses the issue of generating keywords to describe the topic of a shopping need based on the titles and photos of products being browsed or compared. We extend to learn cross-modal and cross-product embeddings to capture the relationships between textual and visual semantics and the shared features between comparable products. Experiments conducted on 3 real-world datasets have shown that the keywords decoded from such embeddings gain significant improvement compared to state-of-the-art cross-modal embeddings.