Rating Classification with DistilBERT – NLP Project

📌 Project Objective

This project aims to automatically classify the rating (1–5 stars) of text reviews for data science books. The model used is DistilBERT, a pre-trained Transformer optimized for efficiency and performance.

The complete workflow includes:

• Dataset analysis
• Preprocessing and balancing
• Tokenization
• Tensor creation
• Model fine-tuning
• Evaluation and error analysis
• Inference using the Hugging Face Pipeline

📊 Dataset

Name: Amazon Reviews Science Size: 20,647 reviews Format: text + rating (1–5) Source: local CSV file

• Main columns: comment → review text stars → numerical rating book_url → link to the book

• Exploratory Analysis: Strong imbalance toward 5-star reviews Average review length: 87 words Outliers: from 1 word to 4,868 words

📈 Included plots (in /images folder):

• Rating distribution
• Review length distribution
• Class balance

🧹 Preprocessing Main operations:

• Convert ratings from 1–5 to 0–4
• Remove empty reviews or reviews with only one word
• Analyze text length
• Undersample the training set to balance classes

Stratified split:

• 80% train
• 10% validation
• 10% test

🔧 Data Processing

Tokenization:

• Model: distilbert-base-uncased
• Parameters: max_length=256, padding="max_length",truncation=True

Tensor creation: Data converted into Hugging Face Dataset objects with columns:

• input_ids
• attention_mask
• label

🧠 Model

• Model used: DistilBertForSequenceClassification
• num_labels=5
• Optimized for text classification
• Trained with Hugging Face Trainer

Training parameters:

• Epochs: 3
• Batch size: 8
• Metrics: accuracy
• Checkpoints saved and evaluated at each epoch

📈 Results

• Validation accuracy
• Test accuracy
• Best performance
• Difficulty

Inference with Hugging Face Pipeline Implemented a pipeline to:

• Classify new reviews
• Visualize probabilities using softmax
• Interpret the final prediction

🛠️ Technologies Used

• Python
• PyTorch
• Hugging Face Transformers
• Hugging Face Datasets
• Pandas
• Matplotlib / Seaborn