Multilingual Sentiment Classification API using BERT

This project demonstrates the development of a multilingual sentiment classifier for customer reviews using a fine-tuned BERT model. The classifier predicts one of four sentiment categories: very negative, negative, positive, and very positive.

Overview

The goal was to create a robust and scalable NLP pipeline for sentiment analysis, with the following components:

• Data Preprocessing: Cleaned and filtered reviews, removed neutral samples, and standardized label encoding.
• Model Training: Fine-tuned a pre-trained bert-base-uncased model with Hugging Face Transformers.
• Evaluation: Measured performance using accuracy and macro F1-score over multiple epochs.
• Deployment: Exposed the trained model as an API using FastAPI.

Model Performance

Epoch	Train Loss	Val Loss	Accuracy	Macro F1
1	0.5536	0.9056	0.6466	0.6486
4	0.0423	1.6631	0.6241	0.6245
8	0.0015	2.1568	0.6391	0.6409

Final Test Metrics (after 8 epochs)

• Accuracy: 65%
• Macro F1-score: 65%
• Performance by class:
  • Very Negative: F1 = 0.70
  • Negative: F1 = 0.61
  • Positive: F1 = 0.62
  • Very Positive: F1 = 0.67

Operation Example

Example Request:

curl -X POST http://127.0.0.1:8000/predict
-H "Content-Type: application/json"
-d '{"text": "The product was terrible and arrived broken."}'

Example Response:

{ "sentiment": "very negative", "confidence": 0.92 }

API Usage

The trained model is served via FastAPI. To run the API:

uvicorn main:app --reload

# Notes

# Random seed (42) was set to ensure reproducibility.
# Trained and evaluated using Apple Silicon (MPS acceleration).
# Tokenizer and model saved for reuse and inference.