Claim Check-Worthiness Detection: How Well do LLMs Grasp Annotation Guidelines?

Laura Majer, Jan Snajder

Abstract

The increasing threat of disinformation calls for automating parts of the fact-checking pipeline. Identifying text segments requiring fact-checking is known as claim detection (CD) and claim check-worthiness detection (CW), the latter incorporating complex domain-specific criteria of worthiness and often framed as a ranking task. Zero- and few-shot LLM prompting is an attractive option for both tasks, as it bypasses the need for labeled datasets and allows verbalized claim and worthiness criteria to be directly used for prompting. We evaluate the LLMs' predictive and calibration accuracy on five CD/CW datasets from diverse domains, each utilizing a different worthiness criterion. We investigate two key aspects: (1) how best to distill factuality and worthiness criteria into a prompt and (2) what amount of context to provide for each claim. To this end, we experiment with varying the level of prompt verbosity and the amount of contextual information provided to the model. Our results show that optimal prompt verbosity is domain-dependent, adding context does not improve performance, and confidence scores can be directly used to produce reliable check-worthiness rankings.

1. Introduction

Fact-checking is a high-stakes, labor-intensive process essential for maintaining information integrity. To assist human experts, researchers focus on two preliminary automated steps: Claim Detection (CD)—identifying if a statement is a factual claim—and Claim Check-Worthiness (CW)—determining if that claim warrants professional investigation.

The primary challenge lies in the subjectivity of "check-worthiness," which varies significantly across domains, such as political debates versus social media health claims. Traditionally, these tasks required massive, manually labeled datasets. This study investigates whether Large Language Models (LLMs) can bypass expensive labeling by utilizing the original human annotation guidelines as prompts.

2. Experimental Framework

We evaluated three models—GPT-3.5-turbo, GPT-4-turbo, and Llama 3 8B—across five diverse English-language datasets: ClaimBuster (CB), CLEF CheckThat! 2022 (CLEF), EnvironmentalClaims (ENV), NewsClaims (NEWS), and PoliClaim (POLI).

Overview of the experimental framework varying verbosity and context levels.

As illustrated in Figure 1, the experimental design systematically varied two dimensions:

  1. Verbosity Level (V0–V3): Ranging from a naive baseline (V0) to the full human guidelines including task definitions, annotation principles, and specific examples (V3).
  2. Context Level (C0–C3): Ranging from the claim in isolation (C0) to the inclusion of co-text (C1), speaker/date metadata (C2), or both (C3).

3. Performance Analysis

Our results indicate that LLMs are highly competitive with traditional supervised models like BERT, often trailing by less than 0.05 F1 points without requiring any training data.

Predictive performance (F1) across models, datasets, and prompt verbosity levels.

As shown in Figure 4, GPT-4-turbo was the top performer, achieving F1 scores of 0.919 for CD and 0.927 for CW on the ClaimBuster dataset. Interestingly, the smaller Llama 3 8B model performed comparably to GPT models on several datasets, though it struggled with the specific nuances of the ClaimBuster set.

4. Key Insights

The "Goldilocks" Prompt

More information in a prompt does not always yield better results. While detailed guidelines (V3) improved performance on ClaimBuster, they significantly hindered performance on the EnvironmentalClaims dataset. Highly verbose prompts can sometimes distract models with edge cases or make them overly restrictive.

Metadata vs. Co-text

When providing context, speaker metadata (C2) generally provided a more significant performance boost than surrounding sentences (co-text, C1). This suggests that the identity and role of a speaker are primary signals for determining check-worthiness.

LLMs as Calibrated Rankers

Beyond binary classification, we found that LLM confidence scores (token likelihoods) correlate well with actual accuracy. This allows LLMs to effectively rank long lists of claims, prioritizing the most "suspicious" statements for human review. GPT-4 achieved an Average Precision (AP) of 0.951 on ClaimBuster in this ranking task.

5. Conclusion

LLMs demonstrate a remarkable ability to grasp complex annotation guidelines for claim detection and check-worthiness. While supervised models like BERT still hold a slight edge, the ability of LLMs to perform zero-shot classification with human-readable guidelines offers a cost-effective alternative for fact-checking organizations. Future work should explore the impact of prompt structure over content and the performance of larger open-source models in non-English contexts.

Summary updated

Abstract

The increasing threat of disinformation calls for automating parts of the fact-checking pipeline. Identifying text segments requiring fact-checking is known as claim detection (CD) and claim check-worthiness detection (CW), the latter incorporating complex domain-specific criteria of worthiness and often framed as a ranking task. Zero- and few-shot LLM prompting is an attractive option for both tasks, as it bypasses the need for labeled datasets and allows verbalized claim and worthiness criteria to be directly used for prompting. We evaluate the LLMs' predictive and calibration accuracy on five CD/CW datasets from diverse domains, each utilizing a different worthiness criterion. We investigate two key aspects: (1) how best to distill factuality and worthiness criteria into a prompt and (2) what amount of context to provide for each claim. To this end, we experiment with varying the level of prompt verbosity and the amount of contextual information provided to the model. Our results show that optimal prompt verbosity is domain-dependent, adding context does not improve performance, and confidence scores can be directly used to produce reliable check-worthiness rankings.
BibTeX
@misc{majer2024claim,
  author = {Laura Majer and Jan Snajder},
  title = {{Claim Check-Worthiness Detection: How Well do LLMs Grasp Annotation Guidelines?}},
  year = {2024},
  eprint = {2404.12174},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.12174},
}
Figure