Shranjeno v:
| Main Authors: | , , , |
|---|---|
| Format: | Recurso digital |
| Jezik: | |
| Izdano: |
Zenodo
2026
|
| Teme: | |
| Online dostop: | https://doi.org/10.5281/zenodo.20025988 |
| Oznake: |
Označite
Brez oznak, prvi označite!
|
Kazalo:
- Large Language Models (LLMs) have shown strong performance across a variety of natural language processing tasks, including question answering, summarization, and conversational systems. However, these models often generate hallucinated outputs, where statements are incorrect or unsupported by reliable evidence. This issue limits their reliability in applications that require accurate and trustworthy information. To address this challenge, this work proposes a multi-signal framework for hallucination detection that combines retrieval-based grounding, natural language inference (NLI), and semantic similarity analysis. The system employs a hybrid retrieval strategy using FAISS-based dense retrieval and BM25-based sparse retrieval to gather supporting evidence. Generated responses are further processed through a fact atomization stage to extract individual claims, which are then verified against retrieved evidence using an NLI model. The proposed framework was evaluated on a subset of 1,000 samples from the HaluEval benchmark dataset. Experimental results show that the system achieves an accuracy of 92.4%, a precision of 93.1%, a recall of 91.6%, and an F1-score of 92.34. Compared to simpler single-signal approaches, the multi-signal framework demonstrates improved reliability in identifying hallucinated content. Overall, the proposed approach provides a scalable and interpretable solution for improving the factual grounding of LLM-based systems.