Godwithus/nigeria-tax-reform-sentiment

🇳🇬 Nigeria 2025 Tax Reform — Sentiment Classifier

A fine-tuned cardiffnlp/twitter-roberta-base-sentiment-latest model for classifying public sentiment toward Nigeria's 2025 Tax Reform Act as expressed on Twitter (X).

This model is the primary artefact of the research paper:

Twitter Sentiment Analysis of Nigeria's 2025 Tax Reform Act: A Transformer-Based Comparative Study of Public Fiscal Policy Discourse
Ezerioha Ifeanyi Emmanuel — 2026

---

📊 Model Performance

Model	Macro-F1	F1-Negative	F1-Neutral	F1-Positive
RoBERTa (this model)	0.6073	0.9436	0.2857	0.5926
BiLSTM (baseline)	0.4763	0.9078	0.1000	0.4211
Logistic Regression (baseline)	0.4376	0.9199	0.0000	0.3929

Evaluated on a held-out stratified test set of 330 tweets (Negative: 288, Neutral: 12, Positive: 30).
RoBERTa significantly outperforms BiLSTM (Welch's t-test, p = 0.0000).

---

🗂️ Dataset

Property	Detail
Source	Nigerian Twitter (X) — manually collected
Collection period	July 1 – September 30, 2025
Total tweets (post-cleaning)	2,197
Labels	Positive / Neutral / Negative
Class distribution	Negative: 87.9% · Positive: 8.7% · Neutral: 3.4%
Train / Val / Test split	70% / 15% / 15% (stratified, seed=42)

---

🚀 Quick Start

from transformers import AutoModelForSequenceClassification, AutoTokenizer, AutoConfig
import numpy as np
from scipy.special import softmax

MODEL = "Godwithus/nigeria-tax-reform-sentiment"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
config    = AutoConfig.from_pretrained(MODEL)
model     = AutoModelForSequenceClassification.from_pretrained(MODEL)

def preprocess(text):
    tokens = []
    for t in text.split():
        t = "@user" if t.startswith("@") and len(t) > 1 else t
        t = "http"  if t.startswith("http")                else t
        tokens.append(t)
    return " ".join(tokens)

def predict(text):
    encoded = tokenizer(preprocess(text), return_tensors="pt",
                        truncation=True, max_length=128)
    output = model(**encoded)
    scores = softmax(output[0][0].detach().numpy())
    labels = {config.id2label[i]: float(scores[i]) for i in range(len(scores))}
    return labels

# Example
print(predict("This development levy will kill businesses in Nigeria!"))
# → {'Negative': 0.94, 'Neutral': 0.03, 'Positive': 0.03}

---

⚙️ Fine-Tuning Details

Hyperparameter	Value
Base model	cardiffnlp/twitter-roberta-base-sentiment-latest
Learning rate	2e-5 (AdamW, linear warmup 10%)
Batch size	16
Max epochs	8 (early stopping, patience=3)
Max sequence length	128 tokens
Imbalance handling	WeightedTrainer + 1.5× Neutral boost
Random seed	42
Hardware	NVIDIA Tesla T4 (Google Colab)

---

🏷️ Label Mapping

ID	Label	Description
0	Negative	Critical, opposed, or concerned reaction
1	Neutral	Informational or non-evaluative commentary
2	Positive	Supportive or optimistic reaction

---

⚠️ Limitations

• Trained on Nigerian Twitter discourse from a specific 3-month window (Jul–Sep 2025). Sentiment patterns may shift as implementation progresses.
• The Neutral class is severely under-represented (3.4% of corpus); Neutral F1 (0.2857) should be interpreted cautiously.
• The model handles English and code-mixed English/Nigerian Pidgin; purely non-English tweets may produce unreliable results.
• Twitter demographic skew (urban, younger, digitally engaged) means findings may not represent broader Nigerian public opinion.

---

📄 Citation

If you use this model or dataset, please cite:

@article{ezerioha2026nigeria,
  title   = {Twitter Sentiment Analysis of Nigeria's 2025 Tax Reform Act:
             A Transformer-Based Comparative Study of Public Fiscal Policy Discourse},
  author  = {Ezerioha, Ifeanyi Emmanuel},
  year    = {2026},
  note    = {Fine-tuned model available at HuggingFace}
}

---

🔗 Links

• 📓 Research Notebook (Google Colab)
• 🤗 Live Demo (Hugging Face Space)
• 📧 Contact: [email protected]