This is an experiment in using a double stacked LTSM to build a character model that can generate novel tweets. Most of the code is modified from the excellent examples provided in the Keras Example Folder.
Here's some examples of this trained on political tweets, and then fine tuned on the tweets of Donald Trump...
diversity: 0.2
----- Generating with seed: "ioofficial i will be"
ioofficial i will be a great deal with any marcon of the country would be a great president who did the republicans we must be sure to change the right things d
----- diversity: 0.5
----- Generating with seed: "ioofficial i will be"
ioofficial i will be a great leader and i am there we are on the race in the us will lead to say what i was in las vegas! #trump2016 #makeamericagreatagain #tru
----- diversity: 1.0
----- Generating with seed: "ioofficial i will be"
ioofficial i will be on @foxandfriends! berand that you're best fichert maryn? http://tmitt.com/cardzaro/20/trump_speaks-theners in america's threet turnberry,
----- diversity: 1.2
----- Generating with seed: "ioofficial i will be"
ioofficial i will be on the hollyfomplianzelecial
via @dlougdgitt
rave bad sovie for syria - will be great to help!
trump chanking iss tweeting in the out-he is
Obviously, the seeds are pretty bad, and the tweets are almost nonsensical. However, it's pretty cool that it learns to form links, makes hashtags, is all about #makeamericagreatagain, and can form words.
This was trained for about 6 hours on a NVIDIA TI 980Ti. Increasing the length of the seed greatly increases training time, so we can't use full tweets as seeds, nor can we do tweets -> replies without running this for a loooong time.
It also has trouble dealing with novel/unseen seeds, which makes me think its severly overfitting. More investigation needed (and more understanding of LTSMs).