Year: 2025
Type: dissertation
Abstract: This thesis evaluates the performance of different BERT-based models in conducting sentiment analysis on Nigerian Pidgin, a widely spoken but understudied low-resource language. Although sentiment analysis has advanced considerably for high-resource languages such as English; low-resource languages remain underrepresented in natural language processing research. This comparative study evaluates the performance of English and Nigerian Pidgin-specific pre-trained BERT models under various fine-tuning conditions to address this gap. We conducted four controlled experiments: (1) English BERT fine-tuned on English data, (2) English BERT fine-tuned on Nigerian Pidgin data, (3) NaijaBERT fine-tuned on English data, and (4) NaijaBERT fine-tuned on Nigerian Pidgin data. Each model was evaluated in both in-language and cross-lingual test sets to assess transfer learning capabilities. Our experiments used 7,000 English sentiment texts and 3,500 Nigerian Pidgin comments for training, with appropriate validation and testing splits. The results show that the fine-tuned models in in-language data outperform zero-shot cross-lingual transfer approaches, with the NaijaBERT model fine-tuned in Nigerian Pidgin data achieving the highest F1-score of 83.6%, compared to 78.9% for English BERT fine-tuned on Nigerian Pidgin. Zero-shot cross-lingual transfer showed a significant performance decline, with F1-scores dropping by 15-21 percentage points when models trained on one language were tested on another. This research contributes to the growing body of work on low-resource language processing and provides more understanding into the effectiveness of cross-lingual sentiment analysis, with implications for developing more inclusive NLP technologies for African languages.
Author: Bossou, Felicien
Institution: Uppsala University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7043171956