Automatic Diacritic Restoration for Yorùbá Text

Abstract

Yorùbá is a language of Nigeria with over 50 million speakers. Their writing system uses a combination of Roman characters (a, e, s, m, etc) and diacritics (marks above and below the letters, as in à, ẹ́, ṣ, m̀, etc). However, because these diacritics can be challenging to type, and because not all websites and software support them, much of the Yorùbá text available online lacks some or all diacritics. This leads to issues with web search, machine translation, and more, contributing to the technology gap between low- and high-resource languages. In my work, I build on prior work to develop a statistical, syllable-based method of Automatic Diacritic Restoration (ADR) for Yorùbá. I compare this model to a Seq2Seq model, using metrics like Word Error Rate, precision, and recall. I find that the syllable-based method performs better at a word level, while the Seq2Seq model performs better for each individual diacritic. Both models achieve a Word Error Rate (WER) below 20%, beating metrics from other much larger models. The syllable-based model achieves a WER of 14.32%, and the translation model achieves a WER of 17.95%, which includes a hallucination rate of 3.87%. I ultimately conclude that highly accurate ADR is possible, even for a language like Yorùbá with many diacritics and limited training data.