MALTO at SemEval-2024 Task 6: Leveraging Synthetic Data for LLM Hallucination Detection

TL;DR

This paper tackles semantic hallucination detection in NLG by combining a data-augmentation pipeline that uses LLM-assisted pseudo-labeling and sentence rephrasing with a three-model DeBERTa-based ensemble (Baseline, C-RLFT, Sequential). Pseudo labels (SOLAR) and rephrasings (GPT-4) expand the scarce labeled data, while a curriculum-like sequential fine-tuning and a learned ensemble fusion improve detection performance. On SHROOM, MNLI-finetuned backbones and the ensemble achieve top accuracy (~0.80) and strong recall, illustrating the benefits of synthetic data and task-aligned pretraining for faithfulness classification. The approach provides a practical pathway to mitigating hallucinations in NLG applications by leveraging diverse training signals and an interpretable voting mechanism.

Abstract

In Natural Language Generation (NLG), contemporary Large Language Models (LLMs) face several challenges, such as generating fluent yet inaccurate outputs and reliance on fluency-centric metrics. This often leads to neural networks exhibiting "hallucinations". The SHROOM challenge focuses on automatically identifying these hallucinations in the generated text. To tackle these issues, we introduce two key components, a data augmentation pipeline incorporating LLM-assisted pseudo-labelling and sentence rephrasing, and a voting ensemble from three models pre-trained on Natural Language Inference (NLI) tasks and fine-tuned on diverse datasets.

Figures (1)

• Figure 1: Pipeline architecture depicting data augmentation techniques and weighted ensemble of strategies