Text Classification with Deep Learning

TensorFlow sentiment-classification study using the IMDB Reviews dataset and a trainable text-embedding pipeline.

Problem Statement

The project examines how raw review text can be transformed into a compact representation and classified as positive or negative without manual feature engineering.

Architecture

flowchart LR
    REVIEW[Raw IMDB Review] --> TOKEN[TensorFlow Hub Text Encoder]
    TOKEN --> EMBED[20-dimensional Representation]
    EMBED --> HIDDEN[Dense ReLU Layer]
    HIDDEN --> OUTPUT[Binary Sentiment Probability]

Loading

Results

The recorded notebook run reports:

• 85.1% test accuracy;
• 0.320 test loss; and
• 400,373 trainable parameters.

The validation curve improves through the recorded 20 epochs, with a growing train/validation gap near the end that should be treated as an overfitting signal.

Tech Stack

• Python
• TensorFlow 2 and TensorFlow Hub
• TensorFlow Datasets: IMDB Reviews
• Jupyter Notebook

Repository Structure

Text_Classification_with_Deep_Learning.ipynb  Training and evaluation
text_classification.tar.gz                    Exported TensorFlow SavedModel
tests/                                        Repository smoke checks

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook Text_Classification_with_Deep_Learning.ipynb

Testing

pytest -q

Limitations

• The notebook uses an older TensorFlow Hub text-encoder API.
• The model evaluates binary English-language movie-review sentiment only.
• No robustness, calibration, subgroup, or out-of-domain evaluation is included.
• The SavedModel archive is a reproducibility artifact, not a production service.

Future Improvements

• Migrate to a current text-vectorization or transformer encoder
• Add precision, recall, F1, calibration, and confusion analysis
• Use early stopping and controlled hyperparameter experiments
• Evaluate domain shift and adversarial phrasing
• Package inference behind a versioned API with monitoring