[Paper Review] Med-MMHL: A Multi-Modal Dataset for Detecting Human- and LLM-Generated Misinformation in the Medical Domain

Yanshen Sun, Jianfeng He+3 more|arXiv (Cornell University)|Jun 15, 2023

Misinformation and Its ImpactsSocial Sciences11 citations

TL;DR

Med-MMHL is a multi-modal misinformation dataset for medicine that includes human- and LLM-generated fake content across multiple diseases, with benchmarks at sentence, document, and multimodal levels.

ABSTRACT

The pervasive influence of misinformation has far-reaching and detrimental effects on both individuals and society. The COVID-19 pandemic has witnessed an alarming surge in the dissemination of medical misinformation. However, existing datasets pertaining to misinformation predominantly focus on textual information, neglecting the inclusion of visual elements, and tend to center solely on COVID-19-related misinformation, overlooking misinformation surrounding other diseases. Furthermore, the potential of Large Language Models (LLMs), such as the ChatGPT developed in late 2022, in generating misinformation has been overlooked in previous works. To overcome these limitations, we present Med-MMHL, a novel multi-modal misinformation detection dataset in a general medical domain encompassing multiple diseases. Med-MMHL not only incorporates human-generated misinformation but also includes misinformation generated by LLMs like ChatGPT. Our dataset aims to facilitate comprehensive research and development of methodologies for detecting misinformation across diverse diseases and various scenarios, including human and LLM-generated misinformation detection at the sentence, document, and multi-modal levels. To access our dataset and code, visit our GitHub repository: \url{https://github.com/styxsys0927/Med-MMHL}.

Motivation & Objective

• Address limitations of prior medical misinformation datasets by including multiple diseases beyond COVID-19.
• Incorporate visual data (images) to enable multimodal misinformation detection.
• Include LLM-generated misinformation (e.g., ChatGPT) alongside human-generated content.
• Provide benchmarks for sentence-, document-, and multimodal-level misinformation detection.
• Crawl data from news, tweets, and images to create a diverse dataset suitable for generalization across diseases.

Proposed method

• Crawl real and fake medical news from authoritative sources and fact-checking sites to build real, human-fake, and LLM-fake texts with associated images.
• Simulate LLM-generated misinformation by perturbing real articles with ChatGPT-3.5 style oppositional sentences and refine language.
• Collect linked tweets corresponding to real and fake news to form tweet-based tasks.
• Create multimodal subsets by aligning text with associated images for fake news and fake claims.
• Define five benchmark tasks (fake news, LLM-fake sentence, multimodal fake news, fake tweets, multimodal tweets) spanning sentences, documents, and multimodal data.

Experimental results

Research questions

• RQ1 How well can models distinguish real medical news from human- and LLM-generated misinformation across multiple diseases?
• RQ2 What is the impact of multimodal information (images) on misinformation detection in the medical domain?
• RQ3 Are LLM-generated fake sentences as detectable as LLM-generated fake news in this domain?
• RQ4 Can models generalize misinformation detection beyond COVID-19 to other diseases within a medical corpus?
• RQ5 How do different baselines perform on sentence-, document-, and multimodal-level misinformation tasks in Med-MMHL?

Key findings

• Transformer-based baselines outperform non-transformer methods on fake news detection, but class imbalance yields higher precision and lower recall.
• FN-BERT achieves the best document-level fake news and tweet detection among the evaluated baselines.
• LLM-generated fake sentences are harder to detect than LLM-generated fake news, indicating a need for improved sentence-level detection approaches.
• Multimodal models (CLIP, VisualBERT) show strong performance on multimodal fake news detection tasks but are limited when only real vs. fake in text.
• The dataset includes 3,455 real news articles, 469 human-generated fake news items, 2,095 LLM-generated fake news items, 2,283 real claims, and 3,567 fake claims, with extensive tweet and multimodal pairings (as detailed in Table 2).