Multi-Pipeline Approach for Sentiment Analysis of West-African Pidgin

This paper presents a multi-pipeline approach to sentiment analysis, with the aim of improving both the accuracy and relevance of the results. Sentiment analysis of West African Pidgin has historically been fragmented, often involving the training of new models with pidgin data, and typically focusing on general sentiment polarity. This study seeks to address these gaps by adopting a holistic, multi-pipeline system for the comprehensive analysis of sentiment polarity in Pidgin text. A subject classifier was developed using the Logistic Regression algorithm to predict the relevance of a body of text to the specific subject matter. Data was collected from Twitter and processed into tokens, which were then used for training and evaluation. This enabled the model to handle a wide range of informal and context-specific words commonly found in Pidgin. For the sentiment analysis itself, a cross-lingual model, RoBERTa (XLM-R), was fine-tuned and expanded through transfer learning using the AfriBERTa model, developed by Ogueji et. al (2021). This fine-tuned model achieved an average F1-score of 74.5 over five runs, demonstrating its effectiveness in sentiment classification. The subject classifier also performed efficiently, achieving an accuracy of 0.81 in identifying relevant text. This multi-pipeline system demonstrates significant promise in enhancing sentiment analysis for Pidgin text, being the first to combine subject classification and cross-lingual sentiment analysis techniques. The results show that the proposed approach can be a valuable tool in natural language processing for underrepresented languages such as West African Pidgin.