Multi-Aspect Joint Retrieval for E-Commerce: Bridging Product Catalogs and Customer Reviews

Abstract

E-commerce platforms support diverse retrieval needs, from finding relevant products to identifying useful information in customer reviews. In these settings, relevance depends not only on textual content, but also on structured aspects such as brand, category, price, rating, and popularity. Recent multi-aspect dense retrievers show that these signals improve product search, but they are typically designed for a single task over a single corpus, leaving open whether aspect-aware representations can be shared across related e-commerce retrieval tasks. In this paper, we study joint multi-aspect dense retrieval for e-commerce, where a single model supports retrieval over both product catalogs and customer reviews. We propose J-MADRAL, a joint dense retrieval model that learns shared aspect-aware representations from textual content and structured metadata across product and review search. To evaluate this setting, we construct Search ESCI, a novel synthetic review-search benchmark derived from Amazon Reviews 2023 and ESCI-related product resources, and evaluate on Amazon ESCI, TREC Product Search 2023, and Search ESCI. Experiments against sparse, dense, and multi-aspect retrieval baselines show that joint training substantially improves review search while maintaining competitive product search effectiveness. These results demonstrate the value of shared aspect-aware representations for integrated e-commerce search. To support reproducibility, we release the source code, data, and model weights at https://github.com/alemarco96/J-MADRAL.