The goal of this project is to build and train a deep learning model that can accurately predict the sentiment of a video game review based on its text.
The workflow involves:
- Creating a suitable dataset from raw Amazon review data.
- Training a Bidirectional LSTM sentiment analysis model.
- Source: Amazon "Video Games" 5-core dataset
- Size: 497,000+ reviews in JSON format.
Two CSV files are created from the raw dataset:
small_corpus.csv→ undersampled dataset for quick testing & prototyping.big_corpus.csv→ larger, weighted corpus for training the final model.
📓 Notebook: creating_dataset.ipynb
- Handles raw JSON processing.
- Generates the CSV files for downstream tasks.
The model is a Bidirectional LSTM (Long Short-Term Memory) network with the following architecture:
- Embedding Layer
- 2 × Bidirectional LSTM Layers
- Dropout Layers (for regularization)
- Dense Layers
- Final Dense Layer → Softmax activation for 3 sentiment categories
📊 Performance: Achieves ~73.76% test accuracy.
📓 Notebook: Sentiment_Analysis.ipynb
- Covers data preprocessing
- Model building & training
- Evaluation