Style-imitative Text Generation using LSTM-RNNs developed at Harvard by Justina Cho, Madhav Datt, and Andrew Zhou
t. but there is no light from within that it was the baaurted to my feal. she othgr me siles as
which wou have been eodowed from the remiraoe of the treasure-seekers. the oes were alotered,
but thaddeus sholto talled all as an rsosabd of the mose of the base it hone what the crows hab
been droken. i di not tee hoom the coor dedone a struisi dood uhich is with a blin snond aod har
been some the briss of a cardle she black as a ceamon of the firhos pe the sime beaore wes oe a
bardle beooein a fiapd oucr the rtrenge with iis hands and tramred dy the corner. the radlery race
ie worded his light out of the ondnet. there was no furniture of a foune whre tome seccsion he
aetorn a stack of chimneys, but he presently reappeared, and then vanished once more upon the
opposite side. when i made my way round there i found him seated at one of the corner eaves.
”that you, watson?” he cried. ”yes.” ”this is the place. what is that black thing down there?” ”a
water-barrel.” ”top on it?” ”yes.” ”no si
We develop a system using long short-term memory recurrent neural networks to generate style-imitative text based on Sherlock Holmes novels. We investigate the per- formance of this system against that of Markov models, and vary hyperparameters (specifically, node and hidden layer numbers as well as training epochs) for LSTM- RNNs to gauge the effectiveness of particular system configurations in generating text, and evaluate results using a modification to the BLEU metric.
You can read more about our system in this paper. This poster gives a great (and extremely short) overview of our system.
