Developing a Novel Deep Learning Architecture for Real-Time Pidgin English Sentiment Analysis in Nigerian Social Media Crisis Communication

Nigerian Pidgin English is the most widely used lingua franca in Nigerian social media, yet its sentiment analysis remains computationally underserved due to the absence of large labelled datasets and the inapplicability of standard NLP architectures to Pidgin's highly variable orthography and code-switching characteristics. This study develops a novel deep learning architecture for real-time Pidgin English sentiment analysis applicable to crisis communication monitoring in Nigerian social media contexts. Three original technical contributions are presented: a Pidgin-adaptive tokenisation scheme that handles orthographic variation through character-level fuzzy matching; a domain-adversarial training approach that transfers sentiment knowledge from English to Pidgin representations using parallel corpora collected from 180,000 Pidgin social media posts; and a distilled transformer architecture (PidginBERT-Mini) capable of real-time inference on mobile server environments relevant to Nigerian crisis management organisation deployments. The architecture was evaluated on a newly constructed Pidgin English Sentiment Corpus of 45,000 manually annotated posts from Twitter, Facebook, and Nairaland covering three Nigerian crisis events (EndSARS, flood response, and COVID-19 vaccination period). Available Pidgin NLP literature from Nigeria identifies the absence of standardised orthography and severe domain shift between formal Pidgin text and crisis-period social media language as the two principal technical barriers. The Transfer Learning Framework and the Crisis Informatics Theory by Imran and colleagues provide the theoretical basis. PidginBERT-Mini achieves a macro-F1 of 0.834 on the test set, significantly outperforming multilingual BERT and AfroXLM-R baselines. Keywords: Pidgin English, sentiment analysis, deep learning, crisis communication, Nigeria.