Exploring Language Patterns in Multilingual Climate Messages

Overview

During the LUNE-TWO Fellowship, I co-created a multilingual dataset of climate messages in English, Yoruba, Hausa, and Igbo alongside 2 linguists. I wanted to see what more could be learned from it, so I ran an experiment: what happens when we cluster these messages using embeddings? This project is my exploration of that question.

What I Did

• Cleaned and tokenised the climate messages.
• Used XLM-Roberta to generate sentence embeddings.
• Applied K-Means clustering and tested cluster quality.
• Visualised patterns with PCA and t-SNE.

What I Found

• Yoruba and Igbo messages often clustered together, which makes sense given their shared Niger–Congo roots.
• Hausa, from a different language family (Afro-Asiatic), formed its own group.
• The clustering quality was solid (Silhouette Scores: Yoruba 0.605, Hausa 0.579, Igbo 0.618).
• Visual plots gave a clear picture of these language relationships.

Why It Matters

This exercise shows how AI/ML techniques can reveal structure in underrepresented languages and hints at applications like:

• Better multilingual climate communication.
• Tools for translation, retrieval, or summarisation in African contexts.
• Making crucial information more accessible across language barriers.

Next Steps

• Try other models like AfriBERTa and mBERT.
• Expand to more African languages.
• Test downstream tasks such as classification or topic modelling.