EGV-Si: Evidence-Grounded Verification for Sinhala via Hybrid NLI and Factual Rules

Abstract

This record contains the paper and related materials for EGV-Si, a hybrid system for Sinhala claim verification. EGV-Si combines a multilingual Natural Language Inference (NLI) model with eleven high-precision factual rules tailored to Sinhala. We introduce a 600-example evaluation dataset of Sinhala claim-evidence pairs labeled as Supported, Refuted, or NotEnoughInfo, covering topics in Sri Lankan history, geography, language, culture, and national symbols. On this dataset, EGV-Si achieves 87.0% accuracy, a 7.3-point absolute improvement over a strong multilingual NLI baseline (79.7%). The gains are concentrated on the Refuted class and come entirely from the factual rules, which operate on 28 unique claim patterns. The work addresses a resource gap for Sinhala, a low-resource language that previously lacked a dedicated NLI or claim-verification dataset. The paper includes detailed ablation analysis, human validation results, error analysis, and a discussion of limitations (including claim repetition and the generated nature of the data).