How MDER‑DR Boosts Multi‑Hop KG QA with Entity‑Centric Summaries

MDER‑DR Two‑Stage Framework

The authors propose a domain‑agnostic KG‑QA framework that separates indexing and retrieval into two coordinated components, addressing the core issue that triple indexing discards contextual semantics.

MDER: Intelligent Indexing Strategy

Traditional methods store raw triples; MDER creates context‑aware entity summaries through a four‑step process:

Map : Identify entities and relations in the text.

Disambiguate : Resolve entity reference ambiguities.

Enrich : Generate natural‑language descriptions of triples based on context.

Reduce : Fuse entity‑level summaries while preserving key semantics.

Key advantage : avoids explicit graph traversal during retrieval, greatly improving efficiency.

DR: Iterative Retrieval Mechanism

For a user query, DR applies a decompose‑parse iterative reasoning strategy driven by a large language model:

Decompose : Break a complex query into multiple triple sub‑problems.

Parse : Anchor these triples in the knowledge graph and iteratively narrow the answer space.

LLM‑driven : The whole pipeline is powered by LLMs, offering robustness to sparse, incomplete, and complex relational data.

Experimental Results

On standard benchmarks and domain‑specific datasets, MDER‑DR achieves up to 66 % performance improvement over traditional RAG baselines while maintaining cross‑language robustness, demonstrating good generalization.

Advantages Summary

Semantic Preservation : Entity‑level summaries retain contextual details, solving the semantic loss of triple indexing.

Efficiency : Eliminates explicit graph traversal, making retrieval faster.

Robustness : Adapts well to sparsity and incompleteness in knowledge graphs.

Domain‑agnostic : Design is generic and can be quickly adapted to new domains.

End‑to‑End LLM‑driven : Leverages LLM reasoning without cumbersome rule engineering.

Conclusion

MDER‑DR proposes a “re‑index, light‑retrieve” paradigm: instead of traversing complex graph structures at query time, it generates semantically rich entity summaries during indexing, combined with an iterative query‑decomposition strategy to bridge semantic gaps in multi‑hop QA.

Multi-Hop Question Answering with Entity-Centric Summaries
https://arxiv.org/pdf/2603.11223