Kaweeshwara/sinhalacheck-module1

SinhalaCheck — Module 1: Sinhala News Credibility Classifier

Fine-tuned LaBSE for binary credibility classification of Sinhala news (CREDIBLE = 1, NOT_CREDIBLE = 0).

Part of SinhalaCheck (R26-IT-158), SLIIT final-year research. Author: Kaweeshwara P.D.S. (IT22331304)

Data

LIRNEasia Sinhala Misinformation Corpus (3,000 documents, Excel distribution). CREDIBLE 1,003 vs NOT_CREDIBLE 1,997 (UNCERTAIN + PARTIAL + FALSE).

Results (5-fold cross-validation, n = 3,000)

• Macro-F1: 0.753 ± 0.013
• Accuracy: 0.777
• Beats TF-IDF + SVM by +0.058 macro-F1 (95% CI 0.040–0.076, paired bootstrap)
• Beats SinBERT-large by +0.044 macro-F1

Limitations

• Judges credibility of writing, not factual truth.
• Weak at separating FALSE from UNCERTAIN (only 110 FALSE/PARTIAL examples).
• Max 512 tokens; longer articles are truncated.