DAFT.punk – DeBERTa for Analysis of Feelings in Texts 🎶

Project for the Natural Language Processing (NLP) exam – University of Salerno

📝 Project Overview

DAFT.punk is a sentiment analysis system for Italian song lyrics, developed as part of the NLP course at the University of Salerno. The project leverages a fine-tuned DeBERTa model to classify lyrics into seven emotional categories, providing both a web interface and scripts for data processing and model training.

🚀 How to Run the Application

python gradio_app.py

The web interface will be available at http://localhost:7860

📁 Project Structure

• gradio_app.py – Main web interface (Gradio)
• scripts/
  • process_lyrics.py – Script for labeling lyrics with emotion using GPT
  • train_finetuned_deberta.py – Script for training the DeBERTa model
  • genius_fragments_scraper.py – Scraper for collecting and fragmenting lyrics from Genius
• model/deberta_italian_v2/ – Fine-tuned DeBERTa italian model
• data/
  • raw/lyrics_enriched.json – Main dataset (Italian lyrics with emotion labels)
  • processed/ – Training reports and metric plots

🎯 Features

• Sentiment analysis on Italian song lyrics with 7 emotions: joy, sadness, rage, love, nostalgia, hope, fear
• Graphical visualization of emotion probabilities
• Scripts for data scraping, processing, training, and evaluation

📊 Model & Dataset

• Model: osiria/deberta-base-italian, fine-tuned on a custom dataset of Italian song lyrics
• Dataset: ~3MB of Italian lyrics, each labeled with one of 7 emotions (labels generated via GPT-4 and manual reviewed)

🏆 Model Performance (DeBERTa Sentiment v2)

Please keep in mind that learning has been heavily limited for economical and hardware reasons.

• Accuracy: 0.6767
• Epochs: 10
• Batch size: 16

⚙️ Installation & Requirements

Install the required dependencies with:

pip install -r requirements.txt

Main dependencies:

• torch
• transformers
• datasets
• gradio
• pandas
• numpy
• openai
• beautifulsoup4

📚 Usage

• To launch the web app: python gradio_app.py
• To process and label lyrics: python scripts/process_lyrics.py
• To train the model: python scripts/train_finetuned_deberta.py