GeAR enhances document retrieval by adding generation capabilities to locate and explain relevant information, making search results more interpretable and fine-grained .
https://arxiv.org/abs/2501.02772
🔍 Original Problem:
Traditional bi-encoder retrieval systems compress complex query-document relationships into single similarity scores, making it hard to understand why documents match and locate specific relevant sections .
⚡ Solution in this Paper:
→ GeAR introduces a novel architecture combining bi-encoder retrieval with generation capabilities through a fusion encoder and text decoder
→ The system processes query-document-information triples using contrastive learning to optimize similarity matching
→ A text decoder generates relevant snippets from documents based on fused query-document representations
→ The model synthesizes high-quality training data using LLMs to support the enhanced capabilities
💡 Key Insights:
→ Generation and localization capabilities are synergistic - better generation leads to better information localization
→ Peak localization performance occurs in intermediate layers rather than the final layer
→ The approach maintains retrieval efficiency while adding interpretability
📊 Results:
→ Achieves 0.961 Recall@5 and 0.903 MAP@5 on information retrieval tasks
→ Demonstrates 0.885 Recall@1 and 0.965 MAP@1 for fine-grained localization
→ Generation quality reaches 87.4 ROUGE-1 and 87.1 ROUGE-L scores
Share this post