Developing an Original Multimodal Sentiment Analysis Framework for Nigerian Social Media Content in Low-Resource Conditions

Social media sentiment analysis in Nigeria must contend with multilingual text mixing Nigerian Pidgin, Hausa, Yoruba, Igbo, and English, informal spelling conventions, cultural sentiment expression forms, and increasingly multimodal content combining text, images, and audio, none of which are adequately addressed by existing sentiment analysis frameworks. This study develops an original multimodal sentiment analysis framework for Nigerian social media content under low-resource conditions. The framework, NaiSentiment-MM, introduces five original components: a code-switching aware tokenisation and embedding strategy for Nigerian social media text that handles continuous script switching across four language codes; a cross-lingual transfer learning approach that leverages high-resource language sentiment supervision (English, French) to bootstrap Nigerian language sentiment classification using adversarial domain adaptation; a cultural sentiment lexicon of 8,400 entries capturing Nigerian-specific sentiment intensifiers, insults, praise terms, and political sentiment vocabulary, compiled through crowdsourced annotation; a multimodal fusion architecture combining text, image, and audio sentiment signals using a cross-modal attention mechanism; and a low-compute inference pipeline achieving competitive accuracy at deployable model sizes under 100MB. NaiSentiment-MM was evaluated on a new benchmark dataset of 12,000 annotated Nigerian social media posts across four platforms, achieving a macro-F1 of 0.81 on the multimodal test set compared to 0.73 for the best text-only baseline. The framework provides an original methodological contribution to African social media NLP.

Keywords: multimodal sentiment analysis, Nigerian social media, low-resource NLP, code-switching, African languages