This project demonstrate a practical end-to-end NLP workflow for classifying SMS messages as spam or legitimate text. The aim was to show how a real machine-learning system is built: from raw data and preprocessing to model development, evaluation, and an interactive application. The solution includes a complete text-processing pipeline, TF-IDF feature engineering, two classical machine-learning classifiers, and a Streamlit interface for real-time predictions. This project was organised in a modular way so it remains clear, maintainable, and easy to extend.
SMS Spam Collection dataset was used as the basis for this work. The modelling process is built around two approaches:
- Naive Bayes, used as a simple and reliable baseline
- Support Vector Machine (SVM), which became the final and best-performing model
Every message passes through my preprocessing pipeline (lowercasing, stopword removal, and lemmatization), is converted into TF-IDF features, and then classified by the trained model. To demonstrate the system clearly, everything was wrapped in a Streamlit interface where users can type a message or load an example and instantly receive a prediction.
The number shown under each prediction is the model’s confidence score. It comes from the SVM classifier and reflects how strongly the model leans toward either spam or not-spam. Positive scores indicate a stronger push toward spam, while negative scores indicate a stronger push toward legitimate messages. Scores closer to zero represent borderline cases, and scores further away from zero show higher confidence. For example, a score like +0.55 means the model sees several patterns associated with spam, while a score like -1.02 suggests the message is confidently classified as not spam. Although the score is not a probability, it gives a clear sense of how certain the model is about its decision.
Spam detection is a common and practical NLP problem, but this project to show how a complete solution is built from start to finish. Working through it demonstrated the core skills expected in real NLP work: preparing and cleaning text data, converting messages into numerical features using TF-IDF, training and comparing classical machine-learning models, and organising the code into clean, reusable modules.
Naive Bayes worked well as a quick baseline and handled obvious spam cases effectively. However, the SVM model consistently achieved stronger recall and F1 scores, which made it the right choice for the final version. Both models was evaluated using ROC curves, precision-recall curves, and other metrics. These evaluations helped to understand how each model behaves on real messages and supported the final decision to deploy SVM.
A higher curve indicates that the model maintains strong precision while sccessfully identifying spam messages across different threshold levels.
This project was kept organised into clear folders:
- notebooks/ for exploration, preprocessing, training, and evaluation
- src/ for reusable functions (preprocessing, feature engineering, training utilities)
- models/ for the saved TF-IDF vectorizer and trained classifiers
- app/ for the Streamlit interface
- data/ for the raw dataset and the cleaned version
This structure follows common machine-learning project patterns and makes the workflow easy to follow.
To run the app locally:
- Activate your virtual environment
- Install the project requirements
- Launch the Streamlit interface with: streamlit run app/streamlit_app.py
The app provides a simple text box where anyone can enter a message and instantly see the model’s prediction.
This project was prepared so it can be deployed easily on Streamlit Cloud. This allows the model to run online without any local installation, making it easy to share with others.
This project helped me reinforce several important ideas in NLP and ML engineering, especially the value of clean preprocessing, consistent train/test splits, and proper evaluation beyond accuracy. Moving the model into a real interface also demonstrates how machine learning models can be turned into practical tools.
Python, Scikit-learn, NLTK, Streamlit, Pandas, NumPy, Matplotlib, Seaborn.
There are several enhancements that needs to explore:
- Use transformer-based models such as DistilBERT
- Expand the dataset to include email spam, WhatsApp scams, and phishing messages
- Add explainability with LIME or SHAP
- Introduce threshold tuning and calibration for better interpretability
- Deploy a backend API (FastAPI or Flask)
- Add prediction logging to track usage and monitor model drift
These improvements would make the system stronger and closer to production standards.
When a user enters a message, the system follows four steps:
- Preprocessing The message is cleaned using the same pipeline applied during model training.
- Vectorization The text is transformed into numerical TF-IDF features.
- Prediction The SVM model outputs a class (spam or not spam) along with a decision score.
- Display The result is shown through a clear visual message, along with the score indicating confidence.
This design keeps the interaction simple while still reflecting the underlying ML workflow.
Try the app here: https://spam-detection-nlp-b4gbjrni2anqih42bntlqn.streamlit.app