Advancing Fact-Checking in Low-Resource and Multilingual Contexts: Benchmarks, Frameworks, and Evaluations

The growing reliance on Large Language Models (LLMs) and social media platforms for information consumption has increased the practical impact of misinformation and factual errors. Although automated fact-checking has progressed rapidly, most benchmarks, tools, and evaluation practices remain centered on English. This creates a gap for widely used but under-resourced languages, where evidence coverage is uneven and reproducible evaluation infrastructure is limited. This thesis addresses these gaps by developing language-specific benchmarks and evaluation frameworks for factuality in Urdu and Malay. The first contribution is a set of resources for Urdu factuality research. We construct two annotated Urdu benchmarks through an LLM-assisted translation pipeline followed by expert review and dual-stage validation: UrduFactBench (1,006 claims) for claim verification and UrduFactQA (4,926 questions) for factual question answering. Building on these datasets, we introduce UrduFactCheck, a modular fact-checking framework with claim processing, query generation, evidence retrieval, and evidence-grounded verification. To mitigate sparse Urdu web evidence, the framework supports translation-based evidence retrieval and a thresholded fallback that triggers English retrieval only when Urdu evidence is insufficient. We evaluate key design choices through controlled experiments, including evidence threshold tuning, verifier model selection, benchmarking on held-out subsets of UrduFactBench, automatic factuality evaluation of twelve LLMs on UrduFactQA, and a manual error analysis that highlights evidence quality and verifier reasoning as major sources of failure. The datasets and source code for UrduFactCheck are publicly available at github.com. The second contribution extends this direction to Malay. We construct two Malay claim verification benchmarks that cover different domains: a global benchmark translated from existing English claim datasets (Non-SEA, 1,190 claims) and a Southeast Asia focused benchmark curated from professional fact-check articles with human validation (SEA, 1,320 claims). Both benchmarks include English claims and Malay translations, with translation quality assessed using automatic measures and an LLM-based judge. Using these benchmarks, we conduct a controlled, zero-shot claim-only experiment that tests LLM parametric knowledge under two claim-language conditions (English claims and Malay claims). We report accuracy, macro precision, macro recall, macro F1, and invalid output rates to characterize current model reliability and claim-language effects. The Malay fact-checking framework and accompanying manuscript are currently under preparation; for review purposes, the repository remains withheld. Together, these contributions provide benchmarks and modular evaluation frameworks for low-resource and multilingual fact-checking, enabling reproducible comparisons across models and retrieval settings and supporting future work on factuality beyond English.