Conceptual
Browse Instruct Curate

Transformer Sequence-to-Sequence Transliteration of Romanized Sinhala

Casting the transliteration of Romanized (Latin-script) Sinhala back into native Sinhala script as a sequence-to-sequence task, solved by fine-tuning a Transformer encoder-decoder rather than hand-written rules. Students learn how a data-driven seq2seq model captures the ad-hoc, inconsistent spelling patterns of informal Romanization that a rule-based system cannot, and how the two approaches compare on word- and character-error-rate and BLEU.