Trumpbot v0.1

Trumpbot was my attempt at creating a RNN trained on Donald Trumps(DT) tweets. I used this as a sort of practice project for learning a bit about RNN's and Tensorflow 2. The result was a chaos and a learning experience so let's dive in.

Run with Containers

Docker

If you don't want to install dependencies to your host, you can build a Docker container with the included Dockerfile:

$ docker build -t trumpbot .

The entrypoint is the script to generate the tweets:

$ docker run trumpbot
...
 obamas Top and France at 900 PM on FoxNews. Anderson Congratulations to the House vote for MittRomney o

 hillary Clinton has been a total disaster. I have an idea for her great speech on CNN in the world  a great honor for me and his partisan hotel and every spor

 friends support Trump International Golf Club on the Paris About that Right School is started by the DNC and Clinton and the DNC that will be a great show with t

If you want to interact with the container (perhaps training first) you can shell inside instead:

$ docker run -it --entrypoint bash trumpbot
root@b53b98f12c34:/code# ls
Dockerfile  README.md  __init__.py  learn.py  raw_tweets.txt  requirements.txt	training_checkpoints  trumpbot.py  tweets.txt

You'll be in the /code directory that contains the source code.

Singularity

For users that want to perhaps use GPU (or better leverage the host) the recommendation is to use a Singularity container, and a recipe file Singularity is provided to build the container.

$ sudo singularity build trumpbot.sif Singularity

And then to run (add the --nv flag if you want to leverage any host libraries).

$ singularity run trumpbot.sif

If you need to change the way that tensorflow or numpy are installed, you can edit the Singularity or Docker recipes.

Setup

Setup is pretty straightforward. It only needs numpy and tensorflow 2 alpha just run the start pip install:

pip3 install -r requirements.txt

Dataset

The entire dataset was just tweets scraped from the DT twitter account. I used Jefferson Henrique's library GetOldTweets-python that I modified a little bit. All the raw tweets can be found in the raw_tweets.txt file FYI all the links in any tweet have been removed.

The first thing about using Tweets as a dataset for training is that they are filled with garbage that wreaks havoc when training. Heres what I did:

Removed any links or urls to photos
Simplified all the puncuation, with Trump this is a big thing, his tweets are a clown fiesta of periods and exclemation marks.
Cleaned out any invisible or non-english characters, any foreign characters just casuases trouble.
Removed the '@' symbol, I'll explain why later.
Removed the first couple of months of tweets, they were mostly about the celebrity apprentice and not really core to what I was trying to capture.
Removed any retweets or super short @replies

The final training text is in tweets.txt which altogether is about 20,000 tweets.

Training

I trained the model twice, the first time for 30 epochs which took around 6 hours. The result was absolute garbage, at the time I hadn't removed hidden or foreign characters so it took 6 hours to spit out complete nonsense. So after I cleaned out the tweets again, I ran the training overnight for 50 epochs this time.

Just run the learn.py file to train it again if you want, the model check points are stored in the 'training_checkpoints' folder

python3 learn.py

Generating Tweets

So now the fun part, you can run the command:

python3 trumpbot.py

This will generate 10 tweets from a random group of topics. If you open the trumpbot.py file theres a few things you can play with:

tweets - Number of messages you want generated

temperature - This controls how predictable the tweet will be, by 
    default its random from 0.1 -> 0.4, anything above about 0.7 generates
    garbage.

talking_points - Is a list of inputs to feed the network, try out 
    differnt words and see what works.

num_generate - This controls the length of the message you want to
     get generated.

Result

For my first crack at text generation Im happy with the results. Here are some sample tweets:

hillary Clinton has been a total disaster. If you cant admit that 
the U.S. more than has been treated big baster I am a g

Donald Trump is 45% Iran

healthe lobbyist now wants to raise taxes for our country in the 
first place! If only one thing is clea

friends support Trump Rally Anger Golf Club of Caporate legislation 
at the WhiteHouse today! #MakeAmericaGreatAgain Thank you for your
 support! #Trump2016 

koreau like you it was great being in the last election then will be
 a great show. I have a fan o

koreau lies and losers and losers will be a great show with the U.S.
 The President has a various past c

What I learned

Tweets make for a tough training set. Things like @ mentions just pollute the hell out of the text so unless you want your bot to be constantly @ing everything I need to find a better way to deal with that.
Things I thought the bot would love talking about stuff like #MAGA, Russia, China, and collusion just generate garbage strings.
Text generation is really hard, and takes a ton of training time.
I could probably get a bit better results if I let it train a bit longer but for any drastic improvements I probably need to try another method or spend alot more time tuning the training set.
Pick a subject that doesn't tweet like hes a dad yelling at a little league game. I think because his tweets are short little outbursts its hard to generate a predictable pattern across them.
The words it groups together for differnt topics is probably worth looking at, like whenever you use 'hillary' as a input it usually has the words 'liar' or 'disaster' in the sentence. or how it loves telling you when its gonna be on @Foxandfriends
With the method I used spelling its like to add random 'u' infront of words.

I feel like this is good starting point, and with some work we might have a digital orange man bot in our future.

📮 Contact & Support

Created by Wyatt Ferguson

For any comments or questions message me on Mastodon

☕ Buy Me A Coffee