What makes Tigrinya NLP difficult?

It's tempting to treat 'low-resource language' as a single category with a single fix: get more data. In practice, the difficulty compounds across several independent axes, and Tigrinya is a useful case study in exactly how.

The first axis is script. Tigrinya is written in Ge'ez, a syllabic script (an abugida) occupying its own Unicode block. Most subword tokenizers — the BPE and Unigram models baked into popular pretrained tokenizers — were fit on corpora dominated by Latin and a handful of other high-resource scripts. Applied to Ge'ez text without retraining, they tend to fragment words into long, semantically meaningless subword sequences, which quietly degrades everything built on top: longer sequences, worse compression, weaker representations.

The second axis is morphology. Tigrinya is a Semitic language with root-and-pattern morphology and heavy inflection — a single root can surface across dozens of surface forms depending on tense, aspect, person and number. Tokenization strategies that work fine for morphologically simpler languages tend to either over-fragment (losing the root) or under-fragment (missing the inflectional signal), and which failure mode dominates is itself an empirical question that has to be tested, not assumed.

The third axis is data. There is no Tigrinya equivalent of the large, standardized benchmark suites that make NLP research on English or even mid-resource languages tractable. That means building an NER dataset isn't just an annotation task — it's also a methodology task: what counts as an entity in this context, how much agreement is achievable, and how do you validate quality with no reference to compare against.

None of these problems are unique to Tigrinya, but together they mean the honest starting point for Tigrinya NLP isn't 'apply a pretrained model.' It's building tokenization, representation and evaluation as first-class research problems — because for a language like this, the preprocessing pipeline determines the ceiling on everything downstream.