NLP Major Move – Multi-Task FinBERT

Multi-task fine-tuning of FinBERT to jointly model:

• Financial news sentiment (classification)
• Article importance (classification)

The model uses a shared transformer backbone with task-specific heads and is evaluated using macro and weighted F1 metrics.

---

Installation

pip install -r requirements.txt

A virtual environment is recommended. GPU is recommended for training.

---

Training

Training is implemented in:

ml/notebooks/FinBERT_MultiTask.ipynb

Trained model weights are not included due to GitHub size limitations.
After training, save the model locally and update the model path where necessary.

---

Data Source

Financial news data is retrieved using the MarketAux API.

Base endpoint used:

https://api.marketaux.com/v1/news/all

---

Running the Backend Server

1. Inside the server/ directory, create a .env file:

NEWS_API_TOKEN=your_token_here
NEWS_API_BASE_URL=https://api.marketaux.com/v1/news/all

2. From the root directory, run:

python server/main.py

Ensure the trained model path is correctly configured in the server code.

---

Running the Frontend

No frontend framework or Node server is used.

To run the frontend:

• Open the HTML file using a Live Server extension (e.g., VS Code Live Server),
  or
• Open the HTML file directly in your browser.

---

Notes

• Model weights are excluded due to size constraints.
• Adjust local paths before running.
• GPU recommended for training.