ghanaopenai/gloss-mt-parallel-en-twi

English-Twi Glossary-Annotated Parallel Chunks

Author: Ghana Open AI Supported by: Ghana NLP

English → Twi (Akan) parallel text in short, aligned chunks, each annotated with dictionary lookups for its key noun phrases, the Twi words the reference translation uses for them, and versions of the English with those phrases already pretranslated in place, in angle brackets. It is built for training translation models that get dictionary terms with the source text, such as encoder-decoder models (umT5, NLLB) or translation-specialised LLMs (TranslateGemma), which have no tool calling.

It is a re-cut of ghanaopenai/twi-translation-tool-calling: the same documents, splits and gold Twi references, without the tool-call trajectories and reasoning traces.

How it was built

1. Each document is split into lines; English and Twi are line-aligned (documents whose line counts differ are skipped: 5 in total).
2. Consecutive lines are packed into chunks of at most 192 umT5 tokens of English, cutting only at line boundaries, so every English chunk matches its Twi chunk. Markdown line structure (headings, lists, tables) is kept inside a chunk.
3. The chunk's noun phrases are extracted with spaCy (en_core_web_sm, nkrane-style stopword filtering) and looked up in an offline English→Twi dictionary of 545,633 phrases with exactly one translation each. It starts from the Gemini translations in AfriSpeech/shola-machine-translations: entries that had a single translation (gemini-3.6-flash-one) are kept, and the ~412k entries that had 2-3 candidates (the superseded gemini-3.6-flash fallback) were re-translated with gemini-3.8-flash into one translation each (scripts/rebuild_dictionary_single.py). Phrases not in the dictionary keep their English as the translation (found is false), telling the model to carry them over as they are.
4. Gold alignment. gemini-3.8-flash is given the gold Twi chunk and the English noun phrases (not the English chunk) and maps each phrase to the exact span of the Twi chunk that translates it (or nothing). A span is kept only if it occurs verbatim in the gold Twi chunk on the same line as its English phrase and differs from the English.
5. Pretranslation. A phrase is replaced in the chunk by its Twi in angle brackets, <twi> (case-insensitive, whole words, longest phrase first). In the gold pair a phrase is bracketed only as many times as its Twi span is bracketed in the gold chunk, so both sides carry the same terms. Phrases without a translation stay English.
   • english_pretranslated uses the gold-aligned spans, and twi_bracketed is the gold Twi with the same spans in <brackets>: every inserted term is what the reference really says and sits in the same place on both sides, so a model trained on this pair learns to keep the bracketed terms and translate the English around them. Built from the reference (brackets are removed from the model's output).

The build script is scripts/build_chunked_dataset.py in GhanaOpenAI/gloss-mt, which uses the same twi_translate.seq2seq helpers as training and inference.

Columns

column	content
english	English chunk
english_pretranslated	English chunk with its phrases replaced by <the Twi span the gold chunk uses> (model input)
twi_bracketed	gold Twi chunk with the aligned spans in <brackets> (model target)
twi	gold Twi chunk (human reference, no brackets)

Every <term> in english_pretranslated has exactly one matching <term> in twi_bracketed (checked at build time).

Statistics

split	documents	chunks	chunks with ≥1 glossary hit	glossary phrases (found / total)	gold terms (aligned / noun phrases)
train	15,129	31,909	31,676	529,763 / 710,368	563,500 / 710,368
validation	332	701	694	11,414 / 15,526	12,471 / 15,526
test	538	1,100	1,092	18,819 / 25,212	20,315 / 25,212

Example input format

Models read english_pretranslated and produce twi_bracketed (with translate English to Twi: in front for umT5; NLLB uses its eng_Latn/twi_Latn codes instead):

english:                    ## Mawuli Bruce Spetua, a man who is twenty-seven years old, and he has a UCC certificate, and he wants to serve this nation as a soldier.
english_pretranslated: ## <Mawuli Bruce Spetua>, a <ɔbarima> who is twenty-seven years old, and he has a <UCC abodin krataa>, and he wants to serve this <ɔman> as a <asrafoɔ>.
twi_bracketed:              ## <Mawuli Bruce Spetua>, <ɔbarima> a wadi mfeɛ aduonu nson, na ɔwɛ <UCC abodin krataa>, na ɔpɛ sɛ ɔkɔsom <ɔman> yi sɛ <asrafoɔ>.
twi:                        ## Mawuli Bruce Spetua, ɔbarima a wadi mfeɛ aduonu nson, na ɔwɛ UCC abodin krataa, na ɔpɛ sɛ ɔkɔsom ɔman yi sɛ asrafoɔ.

The model learns to carry the bracketed terms over and to translate the rest; the brackets are stripped from its output (clean_output). Validation and test inputs use dictionary terms in brackets, and twi is the plain reference.

License

CC BY-NC 4.0, same as the source dataset.