TL;DR
Large Language Models (LLMs) often "hallucinate" by generating non-factual information, a major hurdle for their trustworthiness. This paper introduces MiKaNi, a novel multilingual system that tackles this by combining external knowledge retrieval with a BERT-based model to identify and flag these factual errors across multiple languages, achieving competitive results in a recent shared task.
What Problem Does This Paper Address?
The increasing prevalence of LLMs in various applications is hampered by their tendency to generate "hallucinations" – outputs that are either inconsistent with the provided source material (intrinsic) or factually incorrect based on world knowledge (extrinsic). While much research has focused on detecting these hallucinations in English, the multilingual nature of LLMs means that solutions need to work across many languages. The SemEval-2025 Task 3 (Mu-SHROOM) specifically addresses this gap by providing hallucination annotations for fourteen languages, challenging researchers to develop multilingual hallucination detectors. This paper tackles the problem of building such a system that is not only multilingual but also knowledge-aware, aiming to improve the reliability of LLM outputs.
Methods
The MiKaNi system employs a two-part pipeline designed for token-level hallucination identification:
- Retrieval-based Fact Verification Model (RFVM): This component focuses on detecting extrinsic hallucinations. It works by:
- Atomic Fact Extraction: Using GPT-4o, the LLM's answer is broken down into small, verifiable statements (atomic facts). These facts are translated into English for consistency.
- Search Term Generation: GPT-4o generates relevant search terms for each atomic fact.
- Wikipedia Fact Retrieval: These search terms are used to query Wikipedia, retrieving relevant passages. The system uses a three-step process: search, rank (using BM25), and select the most relevant evidence.
- Fact Verification and Hallucination Prediction: The retrieved Wikipedia evidence, along with the original question-answer pair, is fed back into GPT-4o to estimate the probability of each token in the answer being a hallucination.
- BERT-based Model (BM): This component is designed to capture structural and semantic characteristics of hallucinated content, potentially identifying intrinsic hallucinations as well.
- It uses a pre-trained BERT model, prompted with the instruction, question, and answer.
- The BERT output is enhanced with part-of-speech (POS) embeddings.
- This combined representation is processed through several layers to generate a "BM embedding" and a "BERT annotation."
Finally, the outputs from both the RFVM and BM, along with other linguistic features (like POS tags and question-answer entity matches), are fed into a Support Vector Regression model (SVRM). This ensemble model learns to predict the final, combined probability of each token being a hallucination.
Datasets
The primary dataset used is from the SemEval-2025 Task 3 – Mu-SHROOM dataset, which provides hallucination annotations in fourteen languages. The paper mentions using this dataset for training, validation, and testing. For the BERT-based model training, they specifically used the Mu-SHROOM validation datasets for English, German, French, Finnish, Swedish, Italian, and Spanish.
Key Insights and Ideas
- Multilingualism is Crucial: The paper strongly advocates for moving beyond English-centric hallucination detection, demonstrating that a unified approach can work across diverse languages.
- Ensemble Power: Combining retrieval-based fact-checking with a language model's internal understanding (BERT) and further refining with an SVR model significantly boosts performance. The SVRM often improves scores even when individual models perform similarly.
- Wikipedia as a Reliable Source: Leveraging Wikipedia as an external knowledge source for fact verification is a robust strategy for identifying extrinsic hallucinations.
- Token-Level Granularity: Identifying hallucinations at the token level provides much finer-grained information than sentence-level detection, which is crucial for precise error correction.
- Language-Specific Challenges: Performance varies across languages, potentially due to data characteristics, the quality of NLP tools (like SpaCy/Stanza) for certain languages, or challenges in cross-lingual translation for the retrieval component.
Key Results
The MiKaNi system achieved competitive results across multiple languages in the SemEval-2025 Task 3.
- It secured top-10 ranks in eight languages, including English.
- The system outperformed the provided baselines (neural baseline, mark all, mark none) in most languages, with the exception of Chinese where the "mark all" baseline performed slightly better.
- For instance, in English, they achieved an IoU score of 0.5249 and a rank of 8/44. In French, they ranked 2/33 with an IoU of 0.6314.
- The paper highlights that the SVRM ensemble model consistently improved upon the individual RFVM and BM scores, demonstrating the benefit of combining different approaches.
Limitations and Future Work
- Reliance on Closed-Source Models: The RFVM heavily relies on OpenAI's GPT-4o for atomic fact extraction and search term generation. This limits reusability for researchers without API access or financial resources.
- Inference Latency: The pipeline's modular and sequential design leads to high latency, as it involves multiple steps including retrieval and LLM calls.
- Future Work:
- Experiment with open-source and more lightweight models to reduce reliance on proprietary APIs and improve efficiency.
- Optimize the pipeline for better inference speed, especially for deployment in LLM interfaces.
- Further explore incorporating additional metrics into the SVRM to improve predictions, particularly for languages like Italian where performance ceilings are observed.
- Expand language support beyond the current capabilities, limited by the availability of NLP tools like SpaCy/Stanza and multilingual BERT.
Summary updated