Project for generating speeches of austrian politicians with recurrent neural networks.
The project structure is as follows:
backend: Django application of the providing necessary endpoints
frontend: static files composing the
intelligence: deep neural network architecture and data set utilties
infrastructure: utility file for the necessary infrastructure components
neural-politician: Django settings
In general a stacked recurrent neural network (RNN) was applied to protocol of speeches in the Austrian parlament. Underlying LSTMs had a size of 1526 units, which were trained with a sequence length of 15. The following figure depicts the applied architecture.
Training was done on GPU instances on the Google Compute Engine. The utility script
./intelligence/start.sh provides neccessary actions for training and executing the neural network.