The fact that posts coming from Twitter bots are becoming more realistic is exemplified by Deep Drumpf. The Deep Drumpf account on Twitter is built on a deep-learning algorithm known as a recurrent neural network, that is trained on speech transcripts, tweets and debate quotes. The Twitter bot has over 20,000 followers as well as 12 million views. The bot is named after a remark from John Oliver about the laste name of Trump’s ancestors.
The bot was created by Bradley Hayes, a robotics researcher at MIT. Trump was specifically chosen to study as a response to a training model that can simulate Shakespeare, as Hayes argued that it would be a fun process to model while learning modeling techniques. How it works is that the bot creates tweets starting with a random letter and consequently adds the letters that are most likely to follow, until it hits Twitter’s word limit. While the tweets that the bot produces do not always make sense, most of the produced sentences are coherent and recognizable as Trump’s. The bot has even sent replies to messages posted through Trump’s real account, while taking the context of those messages in account.
Hayes has recently added another purpose to the bot. Deep Drumpf campaign website www.deepdrumpf2016.com is now used to raise money for GirlsWhoCode, an organization with the aim of bridging the gender gap in fields such as technology and mathematics.
In the future Hayes wants to develop Twitter accounts for other presidential candidates and make them talk to each other to simulate debates. If you want to judge yourself how good the bot’s algorithm works, go to https://twitter.com/DeepDrumpf. Also feel free to let me know what you think about this development in the field of Twitter bots and its future.
https://www.technologyreview.com/s/602682/why-im-backing-deep-drumpf-and-you-should-too/
http://www.deepdrumpf2016.com/about.html
http://cs.stanford.edu/people/karpathy/char-rnn/shakespear.txt
Dear Stephan, thank you for your blog! I had not heard of the DeepDrumpf twitter account yet, and I got curious so I checked it out. While some tweets don’t make much sense at the first look, they do resemble what Trump indeed stands for/could have said. Thus, I think the algorithm is quite good, and I also think in the future, more and more bots will be used for different things, such as answering customer concerns or indeed stimulate debates. Very interesting!
Hi Stephan,
Thanks for your post! I also hadn’t heard of DeepDrumpf yet. Looking at the Twitter messages, it’s quite scary how realistic some of these tweets are – some very much look like something Trump would say. It’s amazing how accurate a machine learning model can be when there is an abundance of information available to analyze patterns. I agree with Ananda that in the future, an advanced model could even replace the need for human customer support!