Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Can LLMs Detect Intrinsic Hallucinations in Paraphrasing and Machine Translation?

Evangelia Gogoulou, Shorouq Zahra, Liane Guillou, Luise Dürlich, Joakim Nivre

TL;DR

The paper investigates whether open-access LLMs can reliably detect intrinsic hallucinations in paraphrase and machine translation tasks under the HalluciGen framework. It compares a spectrum of model families, sizes, and prompting strategies against multilingual NLI baselines, using carefully constructed datasets with English and non-English instances. Results show that while some LLMs achieve strong detection performance, NLI models frequently provide competitive or superior accuracy, especially in paraphrase; English prompts generally yield the best results, though prompt choice only modestly affects outcomes. The study concludes that no single factor—model size, instruction tuning, or prompt language—consistently predicts success, and highlights limitations such as dataset English-centrism and potential category imbalance, suggesting directions for future detector-training data and cross-model evaluation.

Abstract

A frequently observed problem with LLMs is their tendency to generate output that is nonsensical, illogical, or factually incorrect, often referred to broadly as hallucination. Building on the recently proposed HalluciGen task for hallucination detection and generation, we evaluate a suite of open-access LLMs on their ability to detect intrinsic hallucinations in two conditional generation tasks: translation and paraphrasing. We study how model performance varies across tasks and language and we investigate the impact of model size, instruction tuning, and prompt choice. We find that performance varies across models but is consistent across prompts. Finally, we find that NLI models perform comparably well, suggesting that LLM-based detectors are not the only viable option for this specific task.

Can LLMs Detect Intrinsic Hallucinations in Paraphrasing and Machine Translation?

TL;DR

The paper investigates whether open-access LLMs can reliably detect intrinsic hallucinations in paraphrase and machine translation tasks under the HalluciGen framework. It compares a spectrum of model families, sizes, and prompting strategies against multilingual NLI baselines, using carefully constructed datasets with English and non-English instances. Results show that while some LLMs achieve strong detection performance, NLI models frequently provide competitive or superior accuracy, especially in paraphrase; English prompts generally yield the best results, though prompt choice only modestly affects outcomes. The study concludes that no single factor—model size, instruction tuning, or prompt language—consistently predicts success, and highlights limitations such as dataset English-centrism and potential category imbalance, suggesting directions for future detector-training data and cross-model evaluation.

Abstract

A frequently observed problem with LLMs is their tendency to generate output that is nonsensical, illogical, or factually incorrect, often referred to broadly as hallucination. Building on the recently proposed HalluciGen task for hallucination detection and generation, we evaluate a suite of open-access LLMs on their ability to detect intrinsic hallucinations in two conditional generation tasks: translation and paraphrasing. We study how model performance varies across tasks and language and we investigate the impact of model size, instruction tuning, and prompt choice. We find that performance varies across models but is consistent across prompts. Finally, we find that NLI models perform comparably well, suggesting that LLM-based detectors are not the only viable option for this specific task.
Paper Structure (30 sections, 2 figures, 7 tables)

This paper contains 30 sections, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Dataset Description
  4. Paraphrase Generation
  5. Machine Translation
  6. Experimental Setup
  7. Models
  8. Prompting
  9. Evaluation
  10. NLI Baseline
  11. Results
  12. Paraphrase
  13. Machine Translation
  14. Discussion
  15. Research Questions
  16. ...and 15 more sections

Figures (2)

  • Figure 1: The average proportion of incorrectly labeled source-hyp paraphrase pairs (averaged over all prompts and prompt and data language combinations) filtered by hallucination category. Here, hatch represents the proportion of outputs that were invalid (i.e. falling outside {$hyp1$, $hyp2$}).
  • Figure 2: The average proportion of incorrectly labeled source-hyp translation pairs (averaged over all prompts and prompt and data language combinations) filtered by hallucination category. Here, the hatch represents the proportion of outputs that were invalid (i.e. falling outside {$hyp1$, $hyp2$}).