Hallucination Detection in Natural Language Generation via Knowledge-Transfer Distillation
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Lecture Notes in Networks and Systems
Series Info
Lecture Notes in Networks and Systems; pp 138–149
Scientific Journal Rankings
Orcid
Abstract
In Natural Language Generation (NLG), hallucinations occur when the models’ output content is semantically unsupported or factually incorrect, creating real serious challenges and critical problems for many real-world applications. In this research work, a novel framework for hallucination detection is proposed. Our method integrates the principles of knowledge transfer and knowledge distillation by utilizing the hard labels from a pre-trained NLI transformer model with soft probabilistic labels from a large teacher model (GPT-OSS-120B). We introduce a novel hybrid dual-objective loss function, termed Focal–Distillation Loss, which combines a modified Cross-Entropy—Focal Cross-Entropy—for hard labels with Kullback–Leibler divergence for soft probabilistic labels. This loss function is used to fine-tune the DeBERTa-v3 base model as the student model. This dual supervision with the new dual-objective loss enables the student model to capture complex patterns in the generated text and learn them. Our novel framework is demonstrated by an experiment on the SHROOM 2024 hallucination detection dataset, which achieves state-of-the-art performance, reaching an accuracy of 70.0% and outperforming several existing systems, and there is still room for improvement. These findings demonstrate the effectiveness of combining model-based knowledge transfer with LLM-based distillation, supported by the proposed Focal–Distillation Loss, in advancing hallucination detection within Natural Language Generation.
Description
SJR 2025
0.165
Q4
H-Index
57
Subject Area and Category:
Computer Science
Computer Networks and Communications
Signal Processing
Engineering
Control and Systems Engineering
Citation
Radwan, A., Emad, M., & Hamdi, A. (2027). Hallucination detection in natural language generation via knowledge-transfer distillation. Lecture Notes in Networks and Systems, 138–149. https://doi.org/10.1007/978-3-032-34434-2_12
