Answers are grounded in this paper and cite their sources.
Badini Kurdish, a Kurmanji-group variety written in an unstandardized Arabic-based script, has a single manually POS-annotated resource: the UOZBDN corpus (920 sentences, 54,116 tokens, 38 tags).Expanding it toward the million-token corpus that Badini language technology requires is limited by expert annotation effort. This paper builds a hybrid pre-annotation pipeline and measures what it saves. On a frozen, sentence-disjoint test split we re-evaluate the corpus's rule-based + HMM tagger (82.81%), evaluate a large language model from zero-shot (78.92%) to retrieval-augmented prompting (82.38%, replicated at 82.43% by a second provider), and fine-tune the XLMRoBERTa encoder on the gold training split, which becomes the best single system (85.41%). Because the encoder and the retrievalgrounded LLM err on different tokens, the pipeline auto-accepts tokens on which they agree (83.10% of tokens, 92.44% accurate) and routes disagreements to a human reviewer. A control experiment shows that agreement selects no more accurately than the encoder's own confidence at matched coverage (92.49% vs. 92.44%); what it buys is a denser review interface, worth about a single model's top-2.65 candidates. At an illustrative reviewer accuracy, the routed corpus reaches a projected 90.60% under two-candidate review and 93.29% under free entry, with 16.90% of tokens routed for review (a token share, not a measured effort saving) and $159–$317 of API cost per million tokens. A leave-one-genre-out study on an encoder–HMM pairing shows agreed-token accuracy falling 91.34% to 81.17%, so the projection holds in-domain only. Code, model outputs, and the corpus are available on request.
No takes yet. Share an insight, caveat, or question.
Zebari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: