Reproduce “MDER-DR: Multi-Hop Question Answering with Entity-Centric Summaries” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Riccardo Campi, Nicolò Oreste Pinciroli Vago, Mathyas Giudici, Marco Brambilla, Piero Fraternali

The MDER-DR framework for multi-hop question answering over Knowledge Graphs (KGs) pre-computes multi-hop evidence into entity-centric summaries during indexing to address contextual information loss. This system consistently outperformed existing RAG baselines across multiple QA benchmarks, including a 49% improvement in Soft Exact Match on WikiQA, while demonstrating enhanced cross-lingual robustness.

Key insights

  • Shifting the aggregation of multi-hop relational evidence from query time to the indexing phase through entity-centric summarization can fundamentally transform and improve multi-hop QA.
  • Leveraging Large Language Models (LLMs) at various stages of KG construction (chunking, summarization, triple extraction, enrichment) and retrieval (query decomposition, resolution, response generation) effectively mitigates contextual information loss.
  • Building a language-agnostic KG representation and integrating translation directly into the processing pipeline significantly enhances the cross-lingual robustness of QA systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.