MDER-DR: Multi-Hop Question Answering with Entity-Centric Summaries
New AI research tackles a core RAG weakness, improving answers to complex, multi-step questions by up to 66%.
A research team from Politecnico di Milano has unveiled MDER-DR, a novel framework designed to solve a persistent problem in Retrieval-Augmented Generation (RAG) systems. Standard RAG often struggles with multi-hop question answering—questions that require connecting information from multiple facts or entities—because converting text into simple knowledge graph triples strips away vital contextual nuance. MDER-DR tackles this with a two-part LLM-driven pipeline: the Map-Disambiguate-Enrich-Reduce (MDER) method for indexing, which generates rich, context-derived descriptions for triples and creates entity-level summaries, and the Decompose-Resolve (DR) mechanism for retrieval, which breaks down user queries and grounds them in the knowledge graph through iterative reasoning.
This approach eliminates the need for explicit graph traversal during question answering, making the system more efficient and robust against incomplete or complex relational data. In experiments across standard and domain-specific benchmarks, MDER-DR demonstrated substantial performance gains, achieving up to a 66% improvement over standard RAG baselines while maintaining cross-lingual robustness. The framework is domain-agnostic, meaning it can be applied to various fields from scientific literature to enterprise knowledge bases, offering a significant upgrade for applications that rely on accurate reasoning over interconnected information.
- MDER-DR improves multi-hop QA accuracy by up to 66% over standard RAG baselines by preserving contextual nuance lost in typical knowledge graph indexing.
- The framework uses a novel two-stage pipeline: MDER for creating entity-centric summaries during indexing and DR for decomposing and resolving complex queries.
- It is robust to sparse and incomplete data, works cross-lingually, and avoids the computational cost of explicit graph traversal during retrieval.
Why It Matters
This represents a major step forward for enterprise search, research assistants, and any AI tool that needs to answer complex, interconnected questions accurately.