Sentiment Analysis

Problem statement

Sentiment models are easy to demo but hard to evaluate reproducibly. The pipeline needed clear stages from raw reviews through vectorisation, model training, validation and error inspection.

Engineering objective

1. Selected the IMDB Large Movie Review Dataset for labelled review classification.
2. Compared pooling and BiLSTM architectures with shared evaluation metrics.
3. Defined error analysis workflow for misclassified reviews.
4. Scoped reproducible scripts for training, validation and export.

Architecture and approach

An end-to-end TensorFlow sentiment classification pipeline covering exploration, cleaning, vectorisation, embeddings, model training, validation, testing and saved model export.

• Raw data to cleaned review pipeline.
• Pooling model and BiLSTM comparison.
• Accuracy, precision, recall and F1 reporting.
• Misclassified review inspection for explainability.

How it works

Python, TensorFlow, vectorisation and embedding layers.

IMDB Large Movie Review Dataset (50,000 labelled reviews).

Reproducible training and evaluation scripts.

Controls and constraints

• Public research dataset only.
• No production deployment or client data handling claimed.

What the system includes

• Data exploration and cleaning workflow.
• Model training and comparison scripts.
• Validation, testing and error analysis outputs.
• Saved model artefact.

Product and system views

An end-to-end NLP classification pipeline designed around reproducible training and explainable error analysis.

Evaluation and results

• Demonstrates classical NLP pipeline engineering and evaluation hygiene.
• Useful as a secondary research reference rather than a commercial product case study.

Validation and testing

• Accuracy, precision, recall and F1 on held-out test data.
• Explicit misclassified review analysis.

Limitations

• Movie review domain may not transfer directly to operational text classification tasks.
• Research prototype, not a deployed production service.