📄 GROUNDEDKG-RAG: Grounded Knowledge Graph Index for Long-document Question Answering

Summary

Long-document question answering with RAG often depends on graph or tree indexes that LLMs build by summarizing text chunks, entities and communities. Systems like GraphRAG and RAPTOR work this way. They are expensive and slow, repeat content across hierarchy levels, and can hallucinate because their summaries are not tied to the source. Zhang and Marfurt propose GroundedKG-RAG, which builds its index with linguistic parsers instead of LLM summaries. Sentences are parsed with SRL or AMR into a directed graph. Entities and actions become nodes. PropBank-style semantic roles and the temporal order of actions become edges. Every node and edge keeps links to the sentences it was extracted from. Nodes are embedded together with their graph neighbourhood. At query time, the question goes through the same parsing step, matching nodes are found, and their grounded sentences are filtered and passed to an interchangeable answer-generating LLM. On NarrativeQA examples, the system matches a state-of-the-art proprietary long-context model at lower cost and beats GraphRAG on every metric. Because the index is human-readable, auditing and error analysis are easier.

Key points

• Graph structure: two node types (entity, action) and two edge types (action–entity semantic roles such as A0–A3, plus temporal "next" links between actions). Each element stores its grounding sentences.
• Graph construction: two variants, one built from SRL parses and one from AMR parses.
• Node embeddings: three variants are compared: basic, averaged neighbours, and attention-weighted neighbours.
• Evaluation: NarrativeQA, scored with Exact Match, Sequence Match, BERTScore and ROUGE-L F1, against a GraphRAG baseline using the same LLM. The appendix includes a sentence-count ablation on Robinson Crusoe.
• Error analysis: run on The Tale of Peter Rabbit, it identifies three error types: failed node matching, distracting retrieved sentences, and reasoning over long, similar contexts. Source code is available on request.