Publications

Research on low-resource NLP and machine translation, with a focus on Hausa. ORCID: 0009-0006-9440-1711.

Preprints

Silent Errors in Low-Resource MT Data: Auditing the Hausa Subset of MAFAND-MT Shuaib Shuaib Yusuf, Ndefreke Sam, Jeremiah Digho, Elizabeth Mallim In preparation. arXiv · PDF · Code · BibTeX

Abstract

Parallel corpora for low-resource African languages are often built from web-scraped sources, where truncated or incomplete translations are common and standard filters miss silently omitted content. We present a deterministic, three-stage heuristic auditing pipeline (language-ID, length-ratio, and numeric/hashtag alignment checks) that runs on a consumer CPU with no neural components and logs every pair’s status to a SQLite ledger. On the web-scraped Hausa split of MAFAND-MT (5,865 sentence pairs), it ran in 0.89 s on average (σ = 0.02, ten runs) and flagged 6.0% of pairs (350), mostly (4.7%) numeric/hashtag omissions; manual review confirmed all 350 as true errors. We fine-tuned M2M-100 (418M) on three variants: the raw data, the data with the flagged pairs removed, and the data with those same pairs manually corrected. Removal improved English→Hausa by +0.26 BLEU and +0.46 chrF++ over raw, and also exceeded the manually corrected variant, while Hausa→English was essentially unchanged (−0.24 BLEU, +0.07 chrF++). These single-run differences are small, but they suggest lightweight, auditable heuristics can support data curation for Hausa; significance testing and other languages remain future work.