After the previous experiment with a character based recurrent neural network (RNN) for romantic novel titles I wanted to find more details about word level RNN:s. I was happy to find that Sung Kim has made it easy to explore word-level RNN:s with Tensorflow in this repository. Training text is 50/50 Christmas carols and death metal lyrics.
Output examples below. (added line breaks and bolded lines that seemed like reasonable song names :-). The resulting texts seem choppier than the character level RNN I tried previously. But still fun! Continue reading