🌍 Machine Translation with Seq2Seq Neural Networks

A Neural Machine Translation (NMT) system that translates English to Spanish using TensorFlow, Keras, and a Seq2Seq Encoder-Decoder architecture built with LSTM neural networks.

Originally developed as part of the Codecademy Machine Translation project, this application has been extended into a complete machine learning portfolio project with:

• Interactive command-line translator
• Streamlit web application
• Statistics dashboard
• Unknown vocabulary detection
• Custom sentence translation
• Early stopping
• Modern Keras model persistence (.keras)
• Improved preprocessing and tokenization

---

🚀 Features

Neural Machine Translation

Translate English sentences into Spanish using:

• Seq2Seq architecture
• Encoder-Decoder design
• LSTM neural networks
• Teacher forcing
• TensorFlow / Keras

Example:

Input:
Who?

Output:
¿Quién?

---

Interactive Command-Line Application

Launch the translator directly from the terminal:

python src/translate.py

Features:

1 - Translate custom text
2 - Test dataset examples
3 - Show model statistics
4 - Exit

---

🌐 Streamlit Web Application

The project includes a web-based interface powered by Streamlit.

Run:

streamlit run streamlit_app.py

The application launches in your browser and allows users to:

• Enter English text
• Generate Spanish translations
• View model statistics
• Interact with the translator without using a terminal

---

📸 Application Screenshots

Home Page

---

Translation Example

---

Model Statistics

---

📂 Project Structure

machine-translation/
│
├── assets/
│   ├── home-page.png
│   ├── translation-example.png
│   └── model-statistics.png
│
├── data/
│   └── spa.txt
│
├── models/
│   └── training_model.keras
│
├── src/
│   ├── preprocessing.py
│   ├── train.py
│   ├── translate.py
│   └── utils.py
│
├── tests/
│
├── streamlit_app.py
│
├── README.md
├── LICENSE
├── requirements.txt
└── .gitignore

---

🧠 Model Architecture

Encoder

Input Sequence
      ↓
LSTM (256 units)
      ↓
Hidden State
Cell State

Decoder

Start Token
      ↓
LSTM (256 units)
      ↓
Dense Softmax Layer
      ↓
Predicted Output Token

Total Parameters

9,894,582

Latent Dimension

latent_dim = 256

---

📊 Dataset

Source:

ManyThings.org Language Dataset

Language Pair:

English → Spanish

Training Data Used:

10,000 sentence pairs

Vocabulary Learned:

Encoder Vocabulary : 2516 words
Decoder Vocabulary : 5302 words

---

⚙️ Training Configuration

latent_dim = 256
batch_size = 50
epochs = 100

Training uses:

EarlyStopping(
    monitor="val_loss",
    patience=5,
    restore_best_weights=True
)

This automatically stops training once validation performance stops improving.

---

📈 Model Statistics Dashboard

The application provides real-time model information.

Example:

Encoder vocabulary : 2516
Decoder vocabulary : 5302

Maximum input tokens : 8
Maximum output tokens : 17

Latent dimension : 256

Model size : 75.52 MB

---

🔍 Unknown Vocabulary Detection

Custom translation mode detects words that were not present in the training dataset.

Example:

Input:
My name is Wayne

Warning:
'Wayne' not found in training vocabulary

This helps explain translation limitations.

---

🛠 Technologies Used

• Python 3
• TensorFlow
• Keras
• NumPy
• Streamlit
• Natural Language Processing (NLP)
• Seq2Seq Neural Networks
• LSTM Networks
• Machine Learning
• Deep Learning

---

💻 Installation

Clone the repository:

git clone https://github.com/YOUR_USERNAME/machine-translation.git

Navigate into the project:

cd machine-translation

Create a virtual environment:

python -m venv venv

Activate:

Windows

venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

---

▶️ Usage

Train the Model

python src/train.py

Output:

models/training_model.keras

---

Run CLI Translator

python src/translate.py

---

Run Web Application

streamlit run streamlit_app.py

---

🧪 Example Results

Input

Who?

Translation

¿Quién?

---

Input

Run!

Translation

¡Órale!

---

Input

How are you?

Translation

¿Estabas aquí?

---

🎯 Learning Outcomes

This project demonstrates:

• Neural Machine Translation
• Encoder-Decoder Architectures
• Seq2Seq Models
• LSTM Networks
• Deep Learning Workflows
• Model Persistence
• TensorFlow / Keras Development
• NLP Tokenization
• Streamlit Application Development
• Git & GitHub Project Management

---

🔮 Future Enhancements

Potential future improvements:

• Word Embeddings
• Attention Mechanisms
• Transformer Architecture
• Translation Confidence Scores
• BLEU Score Evaluation
• Larger Training Datasets
• Multiple Language Support
• Docker Deployment
• Cloud Hosting

---

🏷 Topics

Machine Learning • Deep Learning • NLP • Natural Language Processing • Machine Translation • Neural Machine Translation • Seq2Seq • Encoder Decoder • LSTM • TensorFlow • Keras • Streamlit • Python • Artificial Intelligence

---

📄 License

This project is licensed under the MIT License.

See the LICENSE file for details.

Codecademy Machine Translation project and expanded into a standalone machine learning portfolio project featuring a Streamlit web application, custom translation interface, statistics dashboard, and modern Keras workflow.