UMUTeam at SemEval-2025 Task 3: Detecting Hallucinations in Multilingual Texts Using Encoder-only Models Guided by Large Language Models

Ronghao Pan, Tomás Bernal-Beltrán, José Antonio García-Díaz, Rafael Valencia-GarcíaACL 2025

TL;DR

This paper tackles the critical challenge of detecting "hallucinations" – fabricated information – in multilingual texts generated by Large Language Models (LLMs). The authors propose a novel approach that fine-tunes a multilingual encoder model (XLM-ROBERTa) by incorporating contextual information from a powerful LLM (Llama-3.1-70B) to improve accuracy in identifying these factual errors.

What Problem Does This Paper Address?

Large Language Models (LLMs) like GPT-4, LLaMa, and Mistral have revolutionized Natural Language Processing, demonstrating remarkable capabilities in understanding and generating human language. However, a significant drawback is their tendency to "hallucinate," producing false or fabricated information with high confidence. This poses a serious risk, especially in critical applications such as medicine, legal advice, or news reporting, where factual accuracy is paramount. Detecting these hallucinations, particularly in multilingual contexts where language nuances and data availability vary, is a challenging but crucial task. SemEval-2025 Task 3 (Mu-SHROOM) specifically aims to address this by requiring participants to identify hallucinated text segments and estimate their probability at a character level across multiple languages.

Methods

The core of the proposed solution lies in a token classification approach. The authors leverage the pre-trained multilingual encoder model, XLM-ROBERTa-large, as their foundation. This model is chosen for its strong multilingual capabilities, having been trained on a vast corpus spanning 100 languages.

The key innovation is the integration of contextual information from Llama-3.1-70B, a more recent and knowledge-rich LLM. This is achieved by feeding the model's output (the text to be classified) and a "hallucination-free" reference generated by Llama-3.1-70B as a text_pair input. This allows XLM-ROBERTa to compare the potentially hallucinated output against a high-quality, contextually relevant reference, thereby enhancing its ability to distinguish between factual and fabricated content. The model is fine-tuned to predict a binary label (1 for hallucination, 0 otherwise) for each token.

Datasets

The paper utilizes the labeled training set provided for SemEval-2025 Task 3 (Mu-SHROOM). This dataset is designed for multilingual hallucination detection. The authors mention that for training, 80% of the dataset was used for training and 20% for validation. Specific details about the size and composition of the dataset beyond the number of examples per language are not extensively detailed, but it's noted that SV had 49 examples while others had 50.

Key Insights and Ideas

  • LLM Context as a Guiding Light: The most significant insight is that providing a high-quality, hallucination-free reference from a powerful LLM (Llama-3.1-70B) significantly aids a more specialized encoder model (XLM-ROBERTa) in detecting hallucinations. This suggests a synergistic approach where different LLM architectures can complement each other.
  • Token-Level Precision: The approach's ability to generate hallucination probabilities at the character level is crucial for fine-grained analysis, pinpointing the exact location of fabricated information.
  • Multilingual Robustness: The use of XLM-ROBERTa-large demonstrates the effectiveness of a single multilingual model in handling diverse linguistic landscapes, outperforming baselines across many languages.
  • Length as a Potential Indicator: The paper observes that longer responses from LLMs might correlate with a higher probability of generating hallucinations, possibly due to increased chances of producing unrelated or incorrect information. The teacher model's more controlled response length is highlighted as a contrast.

Key Results

The proposed approach consistently outperformed baseline methods (mark all, mark none, and mark neutral) across multiple languages in terms of both IoU (Intersection over Union) and Cor (Correction) metrics. For instance, in English (EN), the model achieved an IoU of 0.3667 and a Cor of 0.4966, and in Spanish (ES), an IoU of 0.2980 and a Cor of 0.4152. These results indicate a considerable improvement, particularly in detecting token-level hallucinations. The model also ranked competitively in the overall task evaluation for various languages.

Limitations and Future Work

The authors acknowledge that performance is slightly lower in languages with more complex grammatical structures or those less represented in the model's pre-training data (e.g., AR, FA, FI). This is attributed to the lower availability of data in these languages, which impacts the model's generalization capabilities.

Future work directions include:

  • Exploring more advanced fusion techniques for combining multilingual models and LLM context.
  • Investigating cultural and linguistic factors that influence hallucination generation.
  • Expanding the dataset and exploring language-specific strategies to improve generalization in low-resource languages.
  • Addressing the potential deployment challenges of large models like Llama-3.1-70B by exploring lightweight alternatives, such as distilled models or hybrid architectures, for real-time applications.
  • Conducting ablation studies to isolate the specific contribution of contextual input.

Summary updated

BibTeX
@inproceedings{pan2025umuteam,
  author = {Ronghao Pan and Tomás Bernal-Beltrán and José Antonio García-Díaz and Rafael Valencia-García},
  title = {{UMUTeam at SemEval-2025 Task 3: Detecting Hallucinations in Multilingual Texts Using Encoder-only Models Guided by Large Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure