TLDR: A new system called CER (Combining Evidence and Reasoning) has been developed to combat biomedical misinformation. It uses AI to detect claims in text, web pages, and videos, retrieves scientific evidence from PubMed, and then uses large language models to assess the veracity of these claims and provide clear, evidence-based justifications. CER has shown state-of-the-art performance in fact-checking and can even help detect deepfake medical content.
In an era where health information spreads rapidly across digital platforms, distinguishing fact from fiction has become a critical challenge. Biomedical misinformation, ranging from misleading social media posts to fabricated videos, poses significant risks to public health and informed clinical decisions. To address this growing concern, a new comprehensive fact-checking system named CER (Combining Evidence and Reasoning) has been developed by researchers Mariano Barone, Antonio Romano, Giuseppe Riccio, Marco Postiglione, and Vincenzo Moscato.
CER is designed specifically for biomedical content and stands out by integrating specialized components for claim detection, scientific evidence retrieval, and veracity assessment. Unlike many existing fact-checking systems that primarily focus on structured text, CER possesses the unique ability to analyze claims from diverse sources, including videos and general web content. It achieves this by automatically transcribing audio from videos and extracting text from web pages, making it highly adaptable to real-world misinformation scenarios.
The system’s workflow is divided into four main stages. First, the
Claim Detection
phase uses advanced Large Language Models (LLMs) to identify statements within various inputs—be it plain text, web pages, or video transcriptions—that require verification. For videos, it leverages robust speech recognition models like Whisper small-v3 to accurately transcribe spoken content, even with medical terminology.
Next, the
Scientific Evidence Retrieval
module interfaces with PubMed, a vast repository of peer-reviewed biomedical literature. It focuses on extracting relevant evidence from research abstracts using sophisticated retrieval methods, ensuring that claims are evaluated against credible scientific findings.
The
LLM Reasoning
phase then employs powerful LLMs, such as Meta-Llama-3.1-405B-Instruct, to act as assistants. These models assess the veracity of claims by combining them with the retrieved scientific evidence and generate detailed justifications. This step is crucial for mitigating the risk of AI-generated hallucinations and ensuring transparency in the verification process.
Finally, the Also Read:
- Evaluating AI’s Medical Accuracy: A New Benchmark for Chinese Healthcare Texts
- Automating Scientific Benchmarks: How Reasoning Traces Elevate Small Language Models
Veracity Prediction
stage integrates these outputs into a classification framework, assigning one of three labels: “true,” “false,” or “not enough information” (NEI). This stage can utilize either zero-shot classification for rapid deployment or fine-tuning with domain-specific data for enhanced accuracy.
Evaluations on standard benchmarks like HealthFC, BioASQ-7, and SciFact demonstrate that CER achieves state-of-the-art performance, showing significant improvements in accuracy compared to existing approaches. Furthermore, CER has shown promising preliminary results as a complementary safeguard in deepfake detection pipelines, an area where traditional detectors often struggle to keep pace with rapidly evolving synthetic media technologies.
The development of CER offers a robust and transparent tool for researchers, fact-checkers, healthcare professionals, and policymakers. By providing clear, evidence-based explanations alongside claim assessments, it significantly reduces the effort required to assess the credibility of biomedical claims and enhances accuracy through advanced AI techniques. For more technical details, you can refer to the original research paper: Combating Biomedical Misinformation through Multi-modal Claim Detection and Evidence-based Verification.
