Wikipedia multi-hop QA builds its graph locally: adaptive relation discovery lifts MuSiQue F1 from 0.4352 to 0.4577

A knowledge graph for Wikipedia multi-hop question answering can be built entirely locally by extraction under a preset relation schema, without a large model performing open information extraction over the whole corpus. Earlier HippoRAG-style systems built graphs through LLM-based open information extraction, which captured diverse relations but brought inference cost and dependence on commercial APIs at corpus scale. On the MuSiQue corpus, GPT-4o-mini open extraction produced 30,210 relation labels, 101,641 entity nodes and 127,640 triples; the local schema-guided pipeline produced 148 relation types, 41,084 entity nodes and 83,680 triples, of which 55 relation types came from the seed schema and 93 from adaptive discovery.

Corpus-specific relations missing from the preset schema can be recovered automatically from co-occurring entity pairs that the graph has not yet connected, and merged in by post-hoc triple injection without rerunning extraction over the corpus. A fixed relation schema improves reproducibility and deployment practicality, but relations present in the corpus and absent from the schema cannot be extracted, so related entities stay disconnected and multi-hop reasoning paths are cut. On MuSiQue, adaptive expansion added 93 relation types and injected 1,691 triples, raising F1 from 0.4352 with the seed-only graph to 0.4577, a gain of 0.0225; 2WikiMultihopQA added 109 relation types and 1,874 triples for a gain of 0.0076; HotpotQA added 43 relation types and 874 triples for a gain of only 0.0002.

On the retrieval side, Personalized PageRank first expands the query-entity neighborhood, a local cross-encoder reranks candidate facts, and a bounded breadth-first search selects passages by topological reachability. Conventional GraphRAG pipelines often use a generative large model to judge retrieval relevance, and structured-output parsing failures can silently degrade retrieval quality while adding external API cost. On the same adaptive MuSiQue graph, Graph-Path at 4 hops reaches F1 0.4577 and EM 0.3220, above RRF plus cross-encoder at 0.4280, bridging expansion plus cross-encoder at 0.4366 and raw BFS at 0.3932; as the hop limit rises from 2 to 5, F1 is 0.4125, 0.4455, 0.4577 and 0.4528.

Evaluation is limited to 1,000 questions each from the MuSiQue, 2WikiMultihopQA and HotpotQA Wikipedia multi-hop benchmarks, with graph construction, retrieval and reranking all run locally on a single consumer GPU with 6 GB of VRAM. The 4-hop Graph-Path radius was chosen from a sensitivity analysis on MuSiQue and reused across all datasets.

The benefit of adaptive expansion varies widely by dataset: F1 rises 0.0225 on MuSiQue, 0.0076 on 2WikiMultihopQA and only 0.0002 on HotpotQA, indicating the seed schema already covers most useful relations on HotpotQA. The Graph-Path result is framed as a diagnostic probe of graph quality rather than a conclusion that breadth-first routing generally beats relevance-oriented selection; F1 falls to 0.4528 at 5 hops, consistent with semantic drift from expanding to more distant nodes.