Introduction

In this project, I built multiple deep neural networks (simple RNN, RNN with Embedding, Bidirectional RNN, Encoder-Decoder RNN, Encoder-Decoder Bidirectional RNN with Embedding) that can function as part of an end-to-end machine translation pipeline. The pipeline accepts English text as input and returns the French translation. The pipeline consists of:

• Data Preprocessing
• Model Training
• Model Predictions

Setup

This project requires GPU acceleration to run efficiently.

Local Machine (Option)

If you are planning to run on a local machine, I recommend only doing this option if you have a powerful GPU meant for deep learning. For instance, ASUS - ROG GU502GV 15.6" Gaming Laptop - Intel Core i7 - 16GB Memory - NVIDIA GeForce RTX 2060 - 1TB SSD + Optane - Brushed Metallic Black.

Amazon Web Services (Option)

Launch a GPU EC2 instance. For instance, you can choose to launch AWS Deep Learning AMI (Ubuntu 18.04) as a GPU instance.

Install

• Python 3.5
• NumPy
• TensorFlow GPU 1.3.0
• Keras 2.0.9
• Jupyter
• Cython

conda create --name machine-translation python=3.5 numpy
conda activate machine-translation
pip install Cython
pip install tensorflow-gpu==1.3.0
pip install keras==2.0.9
pip install jupyter
cd path/to/project
jupyter notebook

Machine Translation Pipeline

Data Preprocessing

• Tokenization: Tokenize the words into ids
• Padding: Add padding to make all the sequences the same length

Model Training

The models that were built and then trained include:

• Model 1: simple RNN
• Model 2: RNN with Embedding
• Model 3: Bidirectional RNN
• Model 4: Encoder-Decoder RNN
• Model 5: Custom Encoder-Decoder Bidirectional RNN with Embedding (Final Model)

For models 1 - 4, I set them to be trained over 10 epochs with a validation split of 80% for training and 20% for validation. For model 5, I set it to be trained over 30 epochs with the same validation split as models 1 - 4. I increased the epochs for training model 5, so my model correctly predicted both sentences used for testing.

Model Predictions

Each model was trained to translate from english to french text. I included the validation accuracy as it was a good indicator to determine how well each model would do when translating from new english text it had not seen before to french text.

Model	Validation Accuracy
Model 1	61.85
Model 2	79.47
Model 3	67.07
Model 4	68.23
Model 5	95.32

By examining the table, we can determine Model 5: Custom Encoder-Decoder Bidirectional RNN with Embedding (Final Model) would perform the best when translating from new english text it had not seen before to french text. Through the testing in the project, Model 5 was most accurate.

Future Enhancements

• Add the ability to translate from English text to other languages, such as Spanish, Japanese, etc.
• Develop a web application or mobile application
• Integrate with Audio Recognition to translate audio English text to another language, such as French, Spanish, etc

Resources

• Encoder-Decoder Long Short-Term Memory Networks
• A Bouquet of Sequence to Sequence Architectures for Implementing Machine Translation
• Language Translation with RNNs