Step-by-Step Fact Verification System for Medical Claims with Explainable Reasoning

TL;DR

The paper addresses the challenge of verifying medical claims in open-domain settings using an iterative, step-by-step FV pipeline that generates up to five follow-up questions $q_1, \dots, q_5$, retrieves evidence via $R(q,s)$, and reasons with $M_r$ after summarizing with $M_s$, producing a final veracity label $v$ and explanation $e$. It compares this approach to a traditional three-part pipeline and demonstrates improvements across SciFact, HealthFC, and CoVERT datasets. Key contributions include a configurable system with predicate-logic augmentation, evaluation across multiple LLMs and knowledge sources, and insights into how source, predicates, and model choice affect performance. The results suggest strong potential for domain-specific, explainable FV in medical misinformation, with implications for researchers and public health applications, while acknowledging limitations such as reliance on external APIs and the need to handle Not Enough Information in future work.

Abstract

Fact verification (FV) aims to assess the veracity of a claim based on relevant evidence. The traditional approach for automated FV includes a three-part pipeline relying on short evidence snippets and encoder-only inference models. More recent approaches leverage the multi-turn nature of LLMs to address FV as a step-by-step problem where questions inquiring additional context are generated and answered until there is enough information to make a decision. This iterative method makes the verification process rational and explainable. While these methods have been tested for encyclopedic claims, exploration on domain-specific and realistic claims is missing. In this work, we apply an iterative FV system on three medical fact-checking datasets and evaluate it with multiple settings, including different LLMs, external web search, and structured reasoning using logic predicates. We demonstrate improvements in the final performance over traditional approaches and the high potential of step-by-step FV systems for domain-specific claims.

Figures (7)

• Figure 1: The step-by-step fact verification system used in our study iteratively collects additional knowledge and evidence until it can predict a veracity verdict.
• Figure 2: Two out of ten few-shot examples used in the prompt for generating the first verification question.
• Figure 3: Two out of ten few-shot examples used in the prompt for generating the follow-up questions (after the first one had been generated).
• Figure 4: Two out of ten few-shot examples for the verifier module. In this step, the LLM decides if there is enough evidence to make the final veracity prediction or if question generation shall continue.
• Figure 5: Two out of ten few-shot examples for question generation in the predicate pipeline. Each generated question is accompanied by a predicate defining the question and a simple instruction on what to verify.