Fake-Trump-Tweet
A project to generate words and sentences based on Trump's tweets, using an LSTM network.
Dataset
The dataset is from the repo, trump tweet data archive, at 16:37, Mar. 8, 2018. The github address is https://github.com/bpb27/trump_tweet_data_archive
Current Model
At pressent, we apply a simple token-level LSTM model. The model generate one letter each time. By calling it repeatedly, it generates a full sentence. Additional [START] and [END] letter are added to each sequence.
The lettern-level LSTM model is eliminated.
Sample Tweets
Current model is called "Not-Nut-So-Much Fake Trump", which generates some sentences. By studying the sentences, we can know that Nut Fake Trump has already learned some tweet behavior pattern of REAL TRUMP! For example, it says "Real Donald Trump" really often. LOL.
You can see the following sentences are generated by the model.
kind bird opening morgan still . dems icon @ primary fear ! france want to fair drugs any destroyed arab australia ! apart light theapprentice vets date premiere their - pennsylvania building hit im model . came weeks signed . donaldtrump jeffrey career garbage totally reported fighting .
ever @ grow trumpnewyork promised golf further poll apology to the sign facts opportunities certificate general eyes trumpvlog stopping immigrants . colorado agenda masterpiece australia eyes laws secret certificate opens both woman hunt reporters given clear reading christmas officials january finished to unemployment ! this died momentum changes america professional independent public strong wednesday
level guts ! !
room send careful free announcement risk a modern can rate succeed words strategy awards iowa planet israel ! medicare have surprise manufacturing solution pga conflict lives further had ohio to loves moving certificate ground writing words quality florida story lied china to announced she never end nd apart difference . pretty viewers case . release !
coach to rising third !
The so-called "Nut Fake Trump", which generates nothing but non-sence, by using letter level LSTM, is eliminated. You can see it by loading the model in the "letter_model" directory.
How2Use
Download this repo and extract it first. (The raw Trump Tweet data may be very large.)
Train your own new model by calling python3 lstm.py. You may want to set your own hyper parameters in the python file yourself.
Or you can simply generate tweet sentences by calling python3 fake_trump.py.
Requirements
Tensorflow v1.4