ContextClaim: A Context-Driven Paradigm for Verifiable Claim Detection

ACL ARR 2025 May Submission1164 Authors

Abstract: Automated fact-checking (AFC) systems typically follow a sequential pipeline comprising four primary stages: (1) claim detection, (2) matching against previously fact-checked claims, (3) evidence retrieval, and (4) claim verification. While research has progressed significantly in the latter stages of the pipeline, claim detection remains a key bottleneck, often relying on subjective heuristics and lacking integration with broader contextual understanding. We introduce Context-Driven Claim Detection (ContextClaim), a novel paradigm that enhances verifiable claim detection by incorporating context retrieval at the initial stage of the AFC pipeline. ContextClaim leverages a knowledge base, such as Wikipedia, to retrieve, aggregate, and filter information about entity mentions, then, generates supplemental context summaries using GPT-4o and Mistral to enrich the assessment process. Our two variants of ContextClaim improve on the verifiable claim detection task over previous state-of-the-art models as well as over ablated versions of ContextClaim. Furthermore, we investigate the generalizability of the paradigm by applying it across both encoder-only and decoder-only language model architectures and in a cross-domain setting. Experimental results consistently show that ContextClaim enhances claim detection performance under most configurations, suggesting its potential for robust and domain-adaptive deployment in real-world misinformation detection systems.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: fact checking, rumor/misinformation detection

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 1164