Research

Embedding-based Retrieval (EBR) is a powerful search retrieval technique in e-commerce to address semantic matches between search queries and products. However, commerce search engines like Facebook Marketplace Search are complex multi-stage systems with each stage optimized for different business objectives. Search retrieval system usually focuses on query-product semantic relevance, while search ranking puts more emphasis on up-ranking products for high quality engagement. As a result, the end-to-end search experience is a combined result of relevance, engagement, and the interaction between different stages of the system. This presents challenges to EBR systems in optimizing overall search experiences. In this paper we present Que2Engage, a search EBR system designed to bridge the gap between retrieval and ranking for better end-to-end optimization. Que2Engage takes a multimodal & multitask approach to infuse contextual information into the retrieval stage and balance different business objectives. We show the effectiveness of our approach via a multitask evaluation framework with thorough baseline comparisons and ablation studies. Que2Engage has been deployed into Facebook Marketplace Search engine and shows significant improvements in user engagement in two weeks of A/B testing.