17042845/COS802_AfroXLMR_Project

Explores Afro-XLMR for sentiment analysis in low-resource African languages. Fine-tuned on AfriSenti (F1=0.88), but limited zero-shot transfer to isiXhosa (F1=0.48). Few-shot adaptation (500 samples) improved results (F1=0.53). Highlights need for balanced, culturally grounded data. # COS802_AfroXLMR_Project # 1. Overview Problem: The AfriSenti dataset is a rich source of sentiment analysis data; however, it does not include isiXhosa or isiZulu, which are among the most widely spoken languages in South Africa and neighboring countries such as Mozambique, Zimbabwe, and Namibia. This creates a gap in representation and leads to low model performance due to the lack of exposure to these languages by pretrained multilingual models. To address this challenge, the project will apply transfer learning techniques, specifically cross-lingual transfer, by fine-tuning the multilingual model AfroXLMR on Xitsonga and Swahili, and then testing it on isiXhosa using zero-shot or few-shot setups. This approach enables evaluation of model performance in languages understood by target users, simulates AfriSenti-style sentiment analysis for isiXhosa using transfer learning and contributes to methodological innovation for low-resource African languages. The methodology is inspired by experiments conducted in the AfriSenti benchmark: A Twitter Sentiment Analysis Benchmark for African Languages. ## 2. Summary Results The notebook explores Afro-XLMR for sentiment analysis in low-resource African languages. The fine-tuned AfriSenti resulted in a weighted F1=0.88, but limited zero-shot transfer to isiXhosa with weighted F1=0.48. The few-shot adaptation (500 samples) improved results modestly (F1=0.53). This highlights need for balanced, culturally grounded data. ## 4. Contents of Zip file The zip file contains the notebook that executes the technical parts of the data science lifecycle(as the problem is defined in this readme.md file under 'Overview'), starting with exploratory data analysis, cleaning the data, tokenizing the data, model training, model testing on isiXhosa and zero and few shot testing. The data is read in from online, the afrisenti dataset is read in directly from github and the data was mounted onto my university google drive. The isiXhosa corpus …