Sentiment Analysis with Neural Networks 🤖

This notebook demonstrates a full pipeline for multi-class sentiment analysis using neural networks and DistilBERT embeddings. Given a dataset of text reviews, it preprocesses the data, extracts embeddings using a pre-trained transformer, trains a custom neural network classifier, and evaluates its performance. You can also use the trained model to predict the sentiment of your own custom input!

Features

• Data Preprocessing: Loads and cleans review data, merges summary and review text, and maps sentiment scores to five classes.
• Label Engineering: Casts sentiment labels to human-readable categories: negative, somewhat negative, neutral, somewhat positive, positive.
• Efficient Tokenization: Uses DistilBERT's tokenizer for fast, robust text encoding.
• Embedding Extraction: Extracts semantic embeddings from DistilBERT for each review, capturing contextual meaning.
• Custom Neural Network Classifier: Trains a multi-layer perceptron on top of the embeddings for accurate multi-class sentiment prediction.
• Training & Validation: Includes stratified dataset splitting, batch loading, and live training progress with accuracy and loss stats.
• Evaluation: Reports both strict and relaxed accuracy (±1 sentiment class) on the test set.
• Custom Prediction: Easily predict the sentiment of any input text using the trained pipeline.

Prerequisites

Before running the notebook, make sure you have the following installed:

• Python 3.8+
• PyTorch: For deep learning and model training.
• Transformers: For DistilBERT and tokenization.
• datasets: For easy data manipulation and splits.
• tqdm: For progress bars during processing.
• CUDA (optional): For faster training if you have a GPU.

Install all dependencies with:

pip install -r requirements.txt

How to Run

1. Prepare the Data
   Place your review dataset as Reviews.csv in the working directory. The CSV should have at least Summary, Text, and Score columns.

2. Open the Notebook
   Launch Jupyter Notebook and open the provided notebook file.

3. Run All Cells
   Execute the notebook cells sequentially. The pipeline will:

   • Preprocess and split the data
   • Tokenize and extract embeddings
   • Train the neural network classifier
   • Evaluate performance on the test set

4. Try Custom Predictions
   At the end, you can enter your own text and see the predicted sentiment class!

How It Works

1. Data Preparation

   • Loads review data from CSV.
   • Merges summary and review content.
   • Converts scores from 1–5 to 0–4 and assigns sentiment labels.

2. Tokenization & Embedding

   • Uses DistilBERT tokenizer to encode text.
   • Extracts the [CLS] token embedding for each review, representing the whole input.

3. Neural Network Training

   • A simple feed-forward neural network is trained on the embeddings.
   • Training and validation accuracy and loss are displayed per epoch.

4. Evaluation

   • Reports standard accuracy and relaxed accuracy (prediction within ±1 class is also counted as correct).
   • Example: If the true label is "neutral" and the model predicts "somewhat positive," it's considered correct under relaxed accuracy.

5. Custom Input Prediction

   • Enter any text and instantly get its predicted sentiment class, mapped to a readable label.

Example Output

Test Accuracy: 72.82%
Relaxed Test Accuracy (±1): 90.50%
Predicted Sentiment Class: 5
positive

Sentiment Classes

Label	Description
0	negative
1	somewhat negative
2	neutral
3	somewhat positive
4	positive

Why This Approach?

• Transfer Learning: By leveraging DistilBERT, the model benefits from rich, pre-trained language representations, making it robust even with limited data.
• Custom Classifier: Training a lightweight neural network on top of embeddings allows for fast experimentation and adaptation to new sentiment categories.
• Transparent Pipeline: Each step—from raw data to prediction—is clearly separated, making it easy to understand, modify, or extend.