AfriScience-MT: Parallel corpus for science translation across six African languages
davlanade · x · 2026-10-25
- AfriScience-MT (EMNLP 2026): a parallel corpus covering six African languages (Amharic, Hausa, Luganda, Northern Sotho, Yorùbá, isiZulu) across 11 scientific domains, addressing the dominance of colonial languages in African scientific communication.
- Professional translators with science communicators translated plain-language summaries into target languages, coining new terms where none existed.
- Benchmarks: closed-source models lead — GPT-5.4 and Gemini-3.1-Flash-Lite top sentence-level COMET at 68.3 and 68.0, tying at 48.3 document-level; best open systems are fine-tuned NLLB-1.3B (67.3 sentence) and TranslateGemma-12B (44.0 document, 1-shot).
- Corpus released to support document-level scientific African MT benchmarking.
More from Research
- Murray Shanahan on disembodied AI consciousness: language of awareness bends toward Buddhist emptiness — mpshanahan · 2026-10-25
- Sebastian Raschka releases 11-hour 'Build Your Own Reasoning Model' video course — rasbt · 2026-10-25
- VISPA: Training-Free Pluralistic Alignment via Automatic Value Selection Lands EMNLP 2026 Oral — preslav_nakov · 2026-10-25
- Apple & CMU's SSR cuts multimodal agent reasoning latency by 90% — jiqizhixin · 2026-10-25
- A free-host heuristic alone scores insanely high on PhishNChips — antoine_chaffin · 2026-10-25
- RFdiffusion 4 generates putative binder for blood-brain-barrier receptor CD98hc with ipTM 0.88 — AllThingsApx · 2026-10-25