Standard recurrent neural network architectures challenged with translation tasks can only provide a limited range of context which is problematic for complex translations. I used for this project a Long Short-Term Memory architecture for Recurrent Neural Networks cells.
In the paper machine_translation_report.pdf, I focus on: explaining the LSTM architecture for RNNs, describing the trained model and presenting some experiments.
If you run your model on AWS, you will need to make the installations to run TensorFlow with GPU support - but you can also directly use my public AMI (id: ami-12281d05).
Choosing a g2.8xlarge instance seems to be a good performance/price ratio compromise. [EDIT] New p2.xlarge instances are available and are more suitable for deep learning tasks.
After the installation please ensure you have those lines into your .bashrc:
# for CUDA + TensorFlow
export CUDA_HOME=/usr/local/cuda
export CUDA_ROOT=/usr/local/cuda
export PATH=$PATH:$CUDA_ROOT/bin:$HOME/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_ROOT/lib64Enjoy !