<![CDATA[Retrieval-augmented generation (RAG) reduces unsupported generation by grounding answers in source content, but retrieval alone does not guarantee that every output claim is attributable to evidence. This paper presents Evidence-Chain Reliable RAG, an empirical hallucination-detection and provenance framework that scores whether generated response sentences are supported by the corresponding RAG source record. The evaluation uses the complete RAGTruth JSONL data available for this study: 2,965 source records, 17,790 assistant responses, and 14,289 exact-offset hallucination spans across Data2Text, QA, and summarization. The experiment converts word-level spans into response-level, sentence-level, and character-span targets; extracts lexical, BM25, TF-IDF, unsupported-number, unsupported-entity, refusal, and Evidence-Chain Score features; and compares seven methods. On the official held-out test split of 2,700 responses, RandomForest achieved the best case-level F1 of 0.626 and PR-AUC of 0.553. The proposed ECS-Span model achieved case-level F1 of 0.614, ROC-AUC of 0.742, and PR-AUC of 0.536 while also producing deterministic provenance explanations. At sentence level, RandomForest again achieved the highest F1 of 0.321; the proposed method obtained F1 of 0.312, ROC-AUC of 0.777, and PR-AUC of 0.245. Exact character-span localization remained difficult, with character-level F1 of 0.197 because sentence-level predictions often include supported text around shorter hallucinated spans. The findings indicate that evidence-chain features are useful for interpretable RAG auditing, but precise span extraction requires token-level sequence labeling or a comparable fine-grained model.]]>
Copyrights © 2025
