Log-Linear RNNs: Towards Recurrent Neural Networks with Flexible Prior Knowledge