Sentiment Analysis Methods for Customer Feedback in Nigerian Telecommunications

Customer feedback from Nigerian telecommunications users, expressed across social media platforms, app stores, and customer care records, constitutes a rich but analytically underutilised dataset for service quality monitoring. This study empirically evaluates sentiment analysis methods for processing customer feedback from Nigerian telecom customers, addressing the challenge of analysing Nigerian Pidgin English and code-switched Yoruba-English and Hausa-English text alongside standard English. A dataset of 18,000 customer feedback items is compiled from MTN Nigeria and Airtel Nigeria Google Play reviews, Twitter mentions, and anonymised customer care transcripts, manually annotated for sentiment polarity as a benchmark dataset. Five methods are evaluated: VADER, TextBlob, BERT, AfroXLMR, and a fine-tuned RoBERTa model. AfroXLMR achieves the highest macro-F1 score of 0.82, outperforming VADER (0.61) and BERT (0.76), demonstrating the value of multilingual pre-training for Nigerian code-switched text. Error analysis reveals that idiomatic Pidgin expressions and hyperbolic customer phrasing are the primary classification failure sources. The study contributes a publicly available annotated Nigerian telecom feedback dataset and a model performance benchmark for the NLP research community. Keywords: sentiment analysis, Nigerian Pidgin, telecommunications, AfroXLMR, customer feedback