Paper digest: Creole and low-resource NLP
AI language tools are built mostly on English text. What research on Jamaican Patois and other Creoles shows, and what local AI teams need next.
Millions of Caribbean people speak a Creole every day, but most AI language tools learn almost entirely from standard English. A small set of research papers shows how to measure that gap, and suggests Creoles may be easier for AI to learn than many other under-served languages.
That matters now because Caribbean groups, such as Jamaica's Amber AI lab, are promising AI models built for local language and culture.
The papers
| Paper | Authors | Year | Link |
|---|---|---|---|
| JamPatoisNLI: A Jamaican Patois Natural Language Inference Dataset | Armstrong, Hewitt and Manning (Stanford) | 2022 | arXiv:2212.03419 |
| CreoleVal: Multilingual Multitask Benchmarks for Creoles | Lent and 20 co-authors | 2023 | arXiv:2310.19567 |
| No Language Left Behind | NLLB Team, Meta | 2022 | arXiv:2207.04672 |
| Neural Machine Translation of Rare Words with Subword Units | Sennrich, Haddow and Birch | 2015 | arXiv:1508.07909 |
| SentencePiece | Kudo and Richardson (Google) | 2018 | arXiv:1808.06226 |
Why this research exists
NLP, or natural language processing, is the field of teaching computers to work with human language. It grew up on English, where text is plentiful and spelling is standard.
Caribbean Creoles are different. They have many speakers but little written text, no single agreed spelling and constant switching between Creole and standard English in the same sentence. A model that has rarely seen Patois will struggle to understand it.
What the papers did
JamPatoisNLI: a first test for Patois
Ruth-Ann Armstrong, John Hewitt and Christopher Manning built the first dataset for testing whether AI understands logic in a Creole. Natural language inference means deciding whether one sentence supports, contradicts or is unrelated to another.
The dataset has 650 examples, about 97 percent drawn from real Patois posts on Twitter, with expert-written follow-up sentences. The team trained English and multilingual models on standard English examples, then gave them up to 250 Patois examples.
The results were encouraging. With those 250 examples, an English model called RoBERTa reached 76.5 percent accuracy. The models improved much faster on Patois than on unrelated Indigenous languages of the Americas. The authors suggest that a Creole's close link to English helps AI transfer what it already knows.
CreoleVal: testing many Creoles at once
Heather Lent and colleagues gathered benchmarks covering up to 28 Creole languages across 8 tasks, including reading comprehension and translation. Having one shared set of tests means researchers can compare methods fairly, rather than each team inventing its own.
No Language Left Behind: translation at scale
Meta's NLLB project built translation models for 200 languages, many of them low-resource. The team interviewed native speakers, mined new training data and checked quality with a human-translated benchmark called Flores-200. The abstract reports a 44 percent improvement in translation quality over the previous best system.
The lesson for Caribbean teams is how much work each added language takes. It also shows the value of checking a model's language list. Haitian Creole is covered, for example, so check whether your language is before relying on a tool.
Tokenizers: how text is cut into pieces
Before a model reads text, a tokenizer splits it into small pieces called tokens. Rico Sennrich and colleagues showed that breaking rare words into smaller parts, using a method called byte pair encoding, improved English to German and English to Russian translation. SentencePiece, from Taku Kudo and John Richardson, made this work directly on raw text in any language.
Tokenizers are trained on whatever text is available, usually English. Creole spellings can get chopped into many awkward pieces. That uses more of the model's memory, costs more on paid AI services and nudges the model towards standard English.
What this means for Jamaica and Trinidad
Most published work is on Jamaican Patois. Trinidad and Tobago's Creole has very little open research or test data. A local AI project would need:
- Written examples labelled for specific tasks, like those in JamPatoisNLI.
- Recorded speech with transcripts, since much daily communication happens in voice notes. Our speech digest covers this.
- Agreed spelling guidelines, made with community input.
- Examples that mix Creole and standard English, because that is how people actually write and speak.
Limits
These datasets are small. JamPatoisNLI has 650 examples, and most come from social media, so they reflect one style of writing. CreoleVal's tests were zero-shot, meaning models were not trained on the Creole first, so the scores show a starting point.
Consent matters too. In small societies, collecting voice notes or radio recordings for training without permission can break laws and trust.
Who should care
If you are building or buying AI for Caribbean customers, ask how it was tested on Creole and on mixed Creole and English. If a vendor can't name a test, treat "supports Caribbean English" as unproven.
For universities and AI labs, the JamPatoisNLI result suggests a practical path. A few hundred carefully labelled examples can go a long way. For many business uses, connecting a model to a local document library may help more than training a new one.
Bottom line
Creoles are under-served by AI, but research suggests they can be learned from small amounts of good data. Watch for the first open datasets for Trinidadian Creole, and for Caribbean AI labs to publish how their models score on tests like JamPatoisNLI.