Tell Me Why: Explainable Public Health Fact-Checking with Large Language Models

Majid Zarharan, Pascal Wullschleger, Babak Behkam Kia, Mohammad Taher Pilehvar, Jennifer FosterNAACL 2024

TL;DR

This paper explores how well Large Language Models (LLMs) can fact-check public health claims and, crucially, explain their reasoning. They found that while top-tier models like GPT-4 excel in zero-shot settings, open-source models become competitive with fine-tuning, and human evaluation reveals nuances in explanation quality that automatic metrics miss.

What Problem Does This Paper Address?

The COVID-19 pandemic underscored the critical need for accurate public health information. With information spreading rapidly online, manual fact-checking is overwhelmed. Automated fact-checking is essential, but for these systems to be trusted, they need to provide clear, convincing explanations for their veracity assessments. Traditional machine learning models offer some explainability, but modern LLMs, while powerful, are often black boxes. This paper tackles the challenge of making LLM-based fact-checking explainable, specifically in the public health domain, by evaluating their ability to not only predict claim veracity but also generate natural language explanations for those predictions.

Methods

The researchers investigated three core tasks: 1) Veracity Prediction: determining if a claim is true, false, mixture, or unproven. 2) Explanation Generation: providing a textual justification for the veracity label. 3) Joint Task: performing both prediction and explanation generation simultaneously. They experimented with various prompting strategies, including zero-shot (no examples) and few-shot (a few examples), and also employed Parameter-Efficient Fine-Tuning (PEFT) for open-source models. A key methodological contribution is their dual evaluation approach, combining established automatic metrics (like ROUGE and NLI-based coherence measures) with a novel, detailed human evaluation framework. This human evaluation assessed explanations across seven criteria, including claim repetition, internal consistency, external consistency, and suggested class accuracy.

Datasets

The primary dataset used in this study is PUBHEALTH (Kotonya and Toni, 2020b). This dataset contains over 12.2K public health claims, each with journalist-crafted gold-standard explanations. The claims are categorized into four veracity labels: true, false, mixture, and unproven.

Key Insights and Ideas

  • Open-source LLMs are catching up: While closed-source models like GPT-4 initially lead in zero-shot veracity prediction, open-source models (like Mistral-7B and Vicuna-13B) demonstrate significant potential, especially when fine-tuned.
  • Joint tasks yield better explanations: Explanations generated as part of a joint prediction-explanation task tend to be of higher quality than those generated solely for explanation. This suggests that the model benefits from the grounding of predicting veracity.
  • Human evaluation is crucial: Automatic metrics like ROUGE, which rely on exact matching, struggle with abstractive explanations. NLI-based metrics offer an improvement, but human evaluation is indispensable for capturing the nuances of explanation quality, revealing issues that automated metrics miss.
  • Gold explanations aren't always perfect: The study found that LLM-generated explanations sometimes outperformed the "gold standard" explanations from the PUBHEALTH dataset, suggesting that human annotators might also introduce ambiguities or biases.
  • Simplicity can be a pitfall: Some models, especially in zero-shot settings, resort to simply repeating the claim or context rather than providing a genuine explanation, which can inflate scores on certain automatic metrics but is undesirable.

Key Results

  • Veracity Prediction: In the zero-shot setting, GPT-4 achieved the highest macro-F1 score. However, with few-shot prompting and PEFT, open-source models like Mistral-7B (with PEFT) achieved a macro-F1 of 72.0, surpassing previous reported scores on the dataset.
  • Explanation Quality (Human Evaluation): The few-shot GPT-4 model was identified as the overall optimal choice for the explanation task based on the S7 metric (which considers all seven human evaluation criteria). For the joint task, few-shot GPT-3.5-D and Vicuna-13B showed promising results.
  • Model Performance: PEFT significantly boosted the performance of Vicuna-13B and Mistral-7B, making them competitive with larger models.

Limitations and Future Work

The authors acknowledge several limitations:

  • Computational Budget: Fine-tuning for Llama-70B and Falcon-180B was not possible due to computational constraints.
  • Domain Specificity: Experiments were limited to the English language and the public health domain.
  • Human Evaluation Scale: While extensive, the human evaluation involved 5 annotators, 10 models, and 52 samples per model (520 instances total), and more qualitative analysis is always possible.

Future work could explore other domains, languages, and more advanced fine-tuning techniques. The study also highlights the need for better automatic metrics that can reliably assess abstractive explanations.

Summary updated

BibTeX
@inproceedings{zarharan2024tell,
  author = {Majid Zarharan and Pascal Wullschleger and Babak Behkam Kia and Mohammad Taher Pilehvar and Jennifer Foster},
  title = {{Tell Me Why: Explainable Public Health Fact-Checking with Large Language Models}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure