Multi-Aspect Dense Retrieval
Topics: AI (Deep Learning), Ranking, Scoring
Dense Retrieval models typically encode queries and documents using single-vector representations for retrieval in the embedding space, which can miss different aspects of relevance, especially in specific domains. This patent proposes a Multi-Aspect Dense Retrieval Model (MADRM) that uses multiple embeddings per aspect to capture various aspects of queries and documents, leading to improved relevance matching. Evaluations on an e-commerce dataset show significant improvements over existing models.