KingsleyElo/naija-sentiment

Nigerian Twitter Sentiment Analysis — Logistic Regression vs RNN vs LSTM vs AfroXLMR # NaijaSenti — Nigerian Twitter Sentiment Analysis Multilingual sentiment classification across four Nigerian languages: **Hausa, Igbo, Nigerian Pidgin, and Yorùbá**. This project fine-tunes and compares four model architectures on the NaijaSenti corpus, from classical ML to a pretrained African language transformer. ## Live Demo Try it on Hugging Face Spaces > Note: The live demo runs Logistic Regression and LSTM. AfroXLMR (best model, > 0.74 F1) is documented in the results table and available as a trained model > on Hugging Face. --- ## Results | Model | Overall Macro F1 | Hausa | Igbo | Pidgin | Yoruba | |---------------------|------------------|-------|------|--------|--------| | Logistic Regression | 0.69 | 0.71 | 0.73 | 0.44 | 0.68 | | SimpleRNN | 0.69 | 0.72 | 0.74 | 0.38 | 0.69 | | LSTM (V2) | 0.71 | 0.75 | 0.75 | 0.43 | 0.69 | | **AfroXLMR** | **0.74** | 0.77 | 0.78 | 0.49 | 0.70 | > **Note on Pidgin neutral:** The neutral class has only 72 training samples in > Pidgin, causing consistent underperformance across all models. This is a known > data limitation in the NaijaSenti corpus, not a modelling failure. --- ## Setup **Requirements:** Python 3.10, pipenv ```bash git clone github.com cd naija-sentiment pipenv install pipenv shell ``` Run notebooks in order: `01 → 02 → 03 → 04 → 05` > AfroXLMR (notebook 05) requires a GPU. Training was done on Google Colab > (T4 GPU, ~25 minutes). The notebook is fully runnable locally for inference > only if model weights are downloaded from the releases section. --- ## Models ### Logistic Regression Classical baseline with TF-IDF features. Aggressive text preprocessing: lowercasing, URL removal, punctuation, emoji, stopword removal, and lemmatization. ### SimpleRNN Keras sequential RNN with embedding layer. Minimal preprocessing (URLs and emojis only) to preserv …