Tigrinya NLP

Problem

Tigrinya — spoken by roughly nine million people and written in the Ge'ez script — has almost none of the tooling that high-resource languages take for granted: no widely adopted tokenizer tuned to its morphology, minimal labeled data for named entity recognition, and no public predictive-text or language models trained on it directly.

Why it matters

Language technology gaps compound digital exclusion: search, keyboards, assistants and moderation tools all quietly assume Latin-script, high-resource input. Closing even part of that gap for Tigrinya is a concrete, tractable step toward NLP that works for the languages people actually use, not just the ones with the largest existing datasets.

Constraints

• Little to no publicly available labeled corpora for NER or classification
• Complex Semitic morphology with heavy inflection, which standard subword tokenizers handle poorly
• Ge'ez script occupies a distinct Unicode range most off-the-shelf tokenizers were never tuned against
• No established benchmark datasets to evaluate against, so evaluation methodology has to be built alongside the models

Approach

The project works bottom-up: first a tokenizer tuned to Ge'ez script boundaries and Tigrinya morphology, then a small curated NER dataset built through targeted annotation, then a lightweight language model for predictive text trained on the resulting representations. Each layer is treated as a research question in its own right rather than an assumed solved problem.

Experiments

• Comparing subword strategies (BPE vs. Unigram) on compression ratio and tokenizer fertility for Tigrinya text
• Fine-tuning a small transformer as an NER baseline over the curated annotation set
• Testing a predictive-keyboard model for top-k next-token accuracy on everyday Tigrinya text

Results

• Tokenizer choice measurably changes downstream sequence length and representation quality for Ge'ez script
• Small, carefully curated annotation sets outperform larger but noisier scraped data for NER on this language

Lessons

• For morphologically rich, low-resource languages, tokenization is not a preprocessing detail — it is a first-order design decision
• Building the evaluation methodology is as much of the research contribution as building the model