Integrating Knowledge Graphs with Adaptive Schema into Wikipedia-based Question Answering Systems

Knowledge-graph-based retrieval can improve multi-hop question answering by linking evidence that is weakly connected to the query in embedding space. This paper presents an adaptive-schema, graph-enhanced retrieval framework for Wikipedia-based question answering that replaces open-ended LLM-based graph construction with local GLiNER2-based schema-guided extraction. The framework discovers missing corpus-specific relations from uncovered entity pairs and injects the resulting triples post hoc, avoiding repeated corpus-wide extraction. Candidate evidence is expanded with Personalized PageRank, reranked by a local cross-encoder, and selected using Graph-Path, a bounded topology-aware passage selection method. On 1,000 questions from MuSiQue, 2WikiMultihopQA, and HotpotQA, the framework achieves F1 scores of 0.4577, 0.6637, and 0.7154, improving over a BGE-based HippoRAG2 baseline. On MuSiQue, adaptive expansion adds 93 relation types and 1,691 triples, raising F1 by 0.0225 over the seed-only graph and achieving Recall@10 of 0.6423 versus 0.5955 for RAPTOR. These results show that controlled schema expansion and topology-aware evidence selection improve evidence coverage while reducing dependence on LLM-based graph construction; only final answer generation remains API-based. Keywords: GraphRAG, retrieval-augmented generation, multi-hop question answering, knowledge graphs, adaptive schema discovery, post-hoc triple injection, cross-encoder reranking.