After the training is done, does the algorithm use a dictionary file to generate haikus, or is it allowed to use entire lines from the dataset? If the former, I’m impressed. If the latter, I’m just annoyed that this is being presented as something of any significance whatsoever.