Back to index
Language

The Syntax of Survival: Why Nigerian Pidgin Challenges Traditional NLP

Nigerian Pidgin develops through language contact and everyday use. My work at Sussex examines how well models understand its meanings and the context of code-switched conversations.

Nigerian Pidgin is a language born from diversity. Nigeria has more than 500 languages, and Pidgin gives many of their speakers a common way to communicate. It carries features of Nigerian languages and borrows much of its vocabulary from English, with words adapted through local use.

That adaptability is what I mean by survival. Pidgin develops through contact between languages, while its speakers can draw on several languages within a conversation. Understanding it requires following that flexibility.

Where models struggle

I became interested in this problem after noticing people using Pidgin words and descriptions to discuss moderated topics freely online. I also saw LLMs produce nonsensical translations of meaningful Pidgin and mixed-language messages. These observations made me question how well the systems understood what people were saying.

Familiar English words can be misleading. Consider this message from an online group discussing a problem:

“hmm... flick moment ni Chief instigator, you gather liver to enter this meeting ooo.”

The message mixes Pidgin with “ni,” a word from another language meaning “is.” The speaker calls “flick moment” an instigator and rebukes them for attending the meeting despite the trouble they caused.

A literal English reading of “gather liver” suggests collecting organs. In this conversation, it means having the nerve to attend. The words become meaningful through their use and the context of the exchange.

Research highlight

A model can get the English meaning right and still miss the Nigerian Pidgin meaning.

What I want to measure

My work at Sussex focuses on evaluation frameworks for code-switched environments. I want to assess how well models handle changes in language and understand the cultural context of a conversation.

The metrics we use should assess that understanding. For a message like this, the evaluation needs to establish whether the model recognises the speaker’s accusation and what “gather liver” conveys in context.

Debby Oluwatuyi

Debby Oluwatuyi

PhD researcher in NLP