Beyond Topical Similarity: Contrastive Evidence Retrieval with Interpretable Attention Alignment in RAG
Researchers have developed a new retrieval framework called CERA, designed to improve the factuality and interpretability of Retrieval-Augmented Generation (RAG) systems. CERA utilizes a novel approach that combines contrastive learning with an attention alignment loss, specifically focusing on human-annotated factual rationales. This method moves beyond simple topical similarity to identify specific tokens that constitute supporting evidence, thereby enhancing the faithfulness and interpretability of evidence selection in RAG. AI
IMPACT Enhances RAG factuality and interpretability by identifying specific supporting evidence tokens.