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

git clone https://github.com/KingsleyElo/naija-sentiment.git
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 preserve sequence structure. Matched the LogReg baseline at 0.69 overall — underperformed on Pidgin due to the severe class imbalance in the neutral class.

LSTM (V2)

Two stacked LSTM layers with class weighting and EarlyStopping. Tokenizer shared with SimpleRNN. Beat LogReg on 3 out of 4 languages — Hausa, Igbo, and Yoruba. LogReg leads marginally on Pidgin (0.44 vs 0.43).

AfroXLMR

Fine-tuned Davlan/afro-xlmr-base (270M parameters) using PyTorch. AdamW optimizer, linear warmup scheduler, gradient clipping, and class-weighted CrossEntropyLoss. Best overall performance.

---

Key Findings

• AfroXLMR's pretrained African language knowledge delivers consistent gains over classical and recurrent baselines across Hausa, Igbo, and Yoruba
• Yoruba diacritics are handled natively by the XLM-R tokenizer — an advantage over the Keras tokenizer used in earlier models
• Pidgin neutral underperformance (F1: 0.07 with AfroXLMR) is a data problem: 72 training samples cannot support any reliable classifier
• SimpleRNN matched LogReg overall (0.69) — gated architectures (LSTM) delivered consistent gains on Hausa, Igbo, and Yoruba but not Pidgin
• Random baseline for a 3-class problem = 0.33. AfroXLMR achieves 2.2× that

---

Dataset

This project uses the NaijaSenti corpus — a Nigerian Twitter sentiment dataset covering Hausa, Igbo, Nigerian Pidgin, and Yorùbá.

Dataset splits used (after leakage removal):

Split	Size
Train	36,569
Val	7,599
Test	17,654

If you use this dataset in your work, please cite the original authors:

@misc{muhammad2022naijasenti,
      title={NaijaSenti: A Nigerian Twitter Sentiment Corpus for Multilingual Sentiment Analysis}, 
      author={Shamsuddeen Hassan Muhammad and David Ifeoluwa Adelani and Sebastian Ruder 
              and Ibrahim Said Ahmad and Idris Abdulmumin and Bello Shehu Bello 
              and Monojit Choudhury and Chris Chinenye Emezue and Saheed Salahudeen Abdullahi 
              and Anuoluwapo Aremu and Alipio Jeorge and Pavel Brazdil},
      year={2022},
      eprint={2201.08277},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

---

Author

Kingsley Eloebhose
ML Engineer
GitHub · LinkedIn · Portfolio