README.md

metadata

license: mit
datasets:
  - HausaNLP/NaijaSenti-Twitter
language:
  - ha
metrics:
  - accuracy
  - f1
  - precision
  - recall
base_model: google-bert/bert-base-cased
pipeline_tag: text-classification
library_name: transformers
tags:
  - NLP
  - sentiment-analysis
  - hausa

Model Name: Hausa Sentiment Analysis
Model ID: Kumshe/Hausa-sentiment-analysis
Language: Hausa

---

Model Description

This model is a BERT-based model fine-tuned for sentiment analysis in the Hausa language. It is trained to classify social media text into different sentiment categories: positive, negative, or neutral.

Intended Use

• Primary Use Case: Sentiment analysis for Hausa social media content, such as tweets or Facebook posts.
• Target Users: NLP researchers, businesses analyzing social media, and developers building sentiment analysis tools for Hausa language content.
• Example Usage:

  from transformers import AutoTokenizer, AutoModelForSequenceClassification
  
  # Load the model and tokenizer
  tokenizer = AutoTokenizer.from_pretrained("Kumshe/Hausa-sentiment-analysis")
  model = AutoModelForSequenceClassification.from_pretrained("Kumshe/Hausa-sentiment-analysis")
  
  # Encode the input text
  inputs = tokenizer("Your Hausa text here", return_tensors="pt")
  
  # Get model predictions
  outputs = model(**inputs)
  

Model Architecture

• Base Model: BERT (Bidirectional Encoder Representations from Transformers)
• Pre-trained Model: bert-base-cased from Hugging Face Transformers library.
• Fine-Tuned Model: Fine-tuned for 40 epochs on a Hausa sentiment dataset.

Training Data

• Data Source: The model was trained on a dataset containing 35,000 examples from social media platforms such as Twitter and Facebook.
• Data Split:
  • Training Set: 80% of the data
  • Validation Set: 20% of the data

Training Details

• Number of Epochs: 40
• Batch Size:
  • Per device training batch size: 32
  • Per device evaluation batch size: 64
• Learning Rate Schedule: Warm-up steps: 10, Weight decay: 0.01
• Optimizer: AdamW
• Training Hardware: Trained on Kaggle using 2 NVIDIA T4 GPUs.

Evaluation Metrics

• Evaluation Loss: 0.6265
• Accuracy: 73.47%
• F1 Score: 73.47%
• Precision: 73.54%
• Recall: 73.47%

Model Performance

The model performs well on the given dataset, achieving a balanced performance between precision, recall, and F1 score, making it suitable for general sentiment analysis tasks in Hausa language text.

Limitations

• The model may not generalize well to other types of Hausa text outside of social media (e.g., formal writing or literature).
• Performance may degrade on text containing slang or regional dialects not well-represented in the training data.
• The model is biased towards the examples in the training dataset; biases in the data may affect predictions.

Ethical Considerations

• Sentiment analysis models can potentially amplify biases present in the training data.
• Use cautiously in sensitive applications to avoid unintended consequences.
• Consider the impact on privacy and data protection laws, especially when analyzing social media content.

License

Citation

If you use this model in your work, please cite it as follows:

@misc{Kumshe2024HausaSentimentAnalysis,
  author = {Umar Muhammad Mustapha Kumshe},
  title = {Hausa Sentiment Analysis},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Kumshe/Hausa-sentiment-analysis}},
}

Contributions

This model was fine-tuned by Umar Muhammad Mustapha Kumshe. Feel free to contribute, provide feedback, or raise issues on the model repository.