The challenge
Part-of-speech tags describe how words function in sentences. Balochi needs well-documented tagging tools for language research and downstream NLP.
Approach
- 01Prepare labelled examples using Universal Dependencies tag conventions.
- 02Fine-tune transformer representations for token classification.
- 03Evaluate held-out predictions and report multiple metrics.
Documented outcome
The BalPOS v2 repository reports 87.26% held-out accuracy and 87.22% weighted F1.
Interpretation & limitations
These values describe the reported held-out evaluation dataset, not performance on every dialect, domain or text style.