http://www.theguardian.com/technology/2016/mar/04/donald-trump-deep-drumpf-twitter-bot
MIT project uses artificial-intelligence algorithm to learn Republican frontrunner's speech patterns before publishing 'remarkably Trump-like statements'
Donald Trump may be a "really smart person" by his own estimation, but his speeches are now fuelling a really smart Twitter bot, which uses artificial-intelligence technology to copy the Republican frontrunner.
[...] "The bot creates Tweets one letter at a time. For example, if the bot randomly begins its Tweet with the letter 'M,' it is somewhat likely to be followed by an 'A,' and then a 'K,' and so on until the bot types out Trump's campaign slogan, 'Make America Great Again.' It then starts over for the next sentence and repeats the process until it reaches the 140-character limit.
The Tweetbot's creator, CSAIL postdoc Bradley Hayes, used techniques from 'deep-learning,' a field of artificial intelligence that uses systems called 'neural networks' to teach computers to to find patterns on their own. Hayes was inspired by an existing training model that can simulate Shakespeare, as well as a recent report that analysed the presidential candidates' linguistic patterns to find that Trump speaks at a fourth-grade level."
I'm not quite sure, but did @realDonaldTrump just replied to this one?
** [Lightweight #LittleMarco] no campaign chance. Believe me, we start winning, winning, winning. That's the mindset.
Related Stories
Microsoft's new AI Twitter bot @tayandyou was shut down after only 24 hours after it began making "offensive" tweets.
The bot was built "by mining relevant public data and by using AI and editorial developed by a staff including improvisational comedians," and designed to target 18-24 year olds.
Shortly after the bot went live, it began making offensive tweets endorsing Nazism and genocide, among other things.
As of this submission, the bot has been shut down, and all but 3 tweets deleted.
The important question is whether or not it succeeded in passing the Turing test.
takyon: This bot sure woke fast, and produced much more logical sentence structures than @DeepDrumpf.
(Score: 3, Touché) by Anonymous Coward on Tuesday March 08 2016, @12:05AM
I wonder if you could do the same with the Soylent comment sections to Trump articles?
(Score: 3, Funny) by Anonymous Coward on Tuesday March 08 2016, @12:06AM
MAKE COMMENTS GREAT AGAIN
(Score: 5, Funny) by maxwell demon on Tuesday March 08 2016, @10:54AM
"I will build the best firewall, the biggest, the strongest, not penetrable, they won’t be hacking through it, like giving it a little ping and they’re over the firewall, it costs us trillions. And I’ll have Microsoft pay for the firewall. Because Microsoft is screwing us so badly. I will take it from out of just a small fraction of the money they’ve been screwing us for over the last number of years."
The Tao of math: The numbers you can count are not the real numbers.
(Score: 3, Insightful) by takyon on Tuesday March 08 2016, @12:10AM
I looked at the Twitter account [twitter.com] when the story came out. I'm not sure whether it means deep learning is still crap, or if the Donald's grammar is just terrible.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by Gravis on Tuesday March 08 2016, @01:49AM
pff! clearly you don't have experience reading the gibberish that most bots spew.
(Score: 2, Insightful) by redneckmother on Tuesday March 08 2016, @04:28AM
s/bots/politicians/
Mas cerveza por favor.
(Score: 2) by maxwell demon on Tuesday March 08 2016, @09:26PM
botliticans?
The Tao of math: The numbers you can count are not the real numbers.
(Score: 4, Funny) by frojack on Tuesday March 08 2016, @02:37AM
If, ah, you simply, ah substitute Obama's name, ah for Trump, ah
then this whole story, ah, would never have made it, ah, to the pages of SoylentNews.
And , ah, even if it did, ah, it would be modded, ah, to to oblivion, by the army of
Obama defenders.
No, you are mistaken. I've always had this sig.
(Score: 4, Informative) by takyon on Tuesday March 08 2016, @02:46AM
If Obama's tweeting, he's not making spelling and grammar mistakes every 10th tweet.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2, Informative) by frojack on Tuesday March 08 2016, @04:34AM
Because he has people for that.
No, you are mistaken. I've always had this sig.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @06:24AM
10 billion dollars and donald drumpf can't afford a proof reader for his tweets?
your posts clearly delineate your bubble of delusion
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @07:43AM
So, according to you, it's not Trump making the mistakes, but his assistants?
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @08:56AM
yeah, thats exactly what i'm sayin
::rolleyes::
(Score: 2) by frojack on Tuesday March 08 2016, @05:50PM
10 billion dollars and donald drumpf can't afford a proof reader for his tweets?
ACs: Too lazy to sign in, and too stupid to follow a thread.
Obama does not do his own tweets. His staff, paid for by taxpayers, does the actual tweeting, (to the extend twitter is used at all).
Trump, (foolishly in my view) insists on doing his own tweets.
No, you are mistaken. I've always had this sig.
(Score: 3, Touché) by Anonymous Coward on Tuesday March 08 2016, @02:49AM
Well, OK, if you say so. But I just wonder...would an analysis of Obama's speech patterns reveal that he speaks at a fourth-grade level? Just asking.
(Score: 4, Touché) by jimshatt on Tuesday March 08 2016, @09:44AM
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @03:13PM
For decades his public speaking has been on that level.
Who knows how he sounds in private.
(Score: 2) by DeathMonkey on Tuesday March 08 2016, @04:58PM
People make jokes about Obama here all the time!
Like that one time you pretended to be against Net Neutrality because Obama came out out in favor of it. That was hilarious, man!
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @02:15PM
The Donald's grammar is terrible. I thought there was a little weirdness here and there, but otherwise it's fairly accurate.
(Score: 3, Insightful) by DeathMonkey on Tuesday March 08 2016, @04:53PM
I looked at the Twitter account when the story came out. I'm not sure whether it means deep learning is still crap, or if the Donald's grammar is just terrible.
Yes
(Score: 2) by Snotnose on Tuesday March 08 2016, @12:19AM
You took a text file, piped it through the dialect filter, and got something hilarious out. First one turned text into black people talking. That one also raised a stink, some thought it was racist. People soon created others so your text would turn into spanglish, I remember a scottish one, etc.
As an example, if you fed it "Hey Joe, how are you today" it would give back "Yo Bro! How dey hangin dis fine day!"
When the dust settled America realized it was saved by a porn star.
(Score: 3, Informative) by hemocyanin on Tuesday March 08 2016, @03:02AM
"I speak Jive" -- from Airplane. Skip ahead to about 1 min (or just watch, Airplane is always worth it): https://www.youtube.com/watch?v=g0j2dVuhr6s [youtube.com]
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @04:19AM
... where you goin' with that gun in your hand?
Sorry - flashback.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @12:26AM
So would the Tweetbot be close to passing a simplified Turing test? Or is just imitating Trump the bar?
Posting anonymously because I don't want to be knifed in self defence.
(Score: -1, Offtopic) by Anonymous Coward on Tuesday March 08 2016, @12:26AM
http://www.godlikeproductions.com/forum1/message3106640/pg1?disclaimer=1 [godlikeproductions.com]
(Score: 1, Informative) by Anonymous Coward on Tuesday March 08 2016, @12:31AM
The Guardian article links to the programmer's blog,
https://www.csail.mit.edu/deepdrumpf [mit.edu] which in turn links to
https://karpathy.github.io/2015/05/21/rnn-effectiveness/ [github.io]. That author used recurrent neural networks to generate an essay on algebraic geometry, and to write C code.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @12:45AM
Look at this twitter bot and how it mocks Trump.
This is what passes for news here? Did some DICE Holdings employees move over when Slashdot got bought?
I'm still not going to vote for some cuck.
(Score: 3, Touché) by Anonymous Coward on Tuesday March 08 2016, @01:05AM
People who think "cuck" is a meaningful term are what they hate.
(Score: -1, Troll) by Anonymous Coward on Tuesday March 08 2016, @02:42AM
So I'm a liberal or jew or something? Seriously I wish you would think before you pull something out of the liberal retort catalog. I use cuck because it humiliates so-called conservatives who vote against the interests of their voters.
(Score: 1, Informative) by Anonymous Coward on Tuesday March 08 2016, @04:05AM
No, you are a cuck, ya dumbcuck!
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @05:51AM
It doesn't really humiliate anyone. It makes you look like subhuman trash from 4chan and other such cesspools.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @01:51AM
Drink another six pack and try again, Ethanol.
(Score: 1, Touché) by Anonymous Coward on Tuesday March 08 2016, @02:07AM
Je Suis Ethanol.
You might be surprised to find that many persons are Ethanol.
(Score: 1) by redneckmother on Tuesday March 08 2016, @04:25AM
s/are/are on/
Mas cerveza por favor.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @06:07AM
Only now, at the end, do we realize the true power of the Anonymous Coward! Hey, your bigotry is showing through your invisibility cloak! Pay no attention to the man behind the curtain! I am the Grate and Glorious Donald!!! From Kansas, of course. You know, if we know who you are when you post AC, there really is no point in posting AC. Unless you are looking for plausible deniability, like Reagan with the Contras, and selling weapons to Iran. Then it might work.
(Score: 1, Funny) by Anonymous Coward on Tuesday March 08 2016, @12:46AM
They tried to automate Hillary also, but it kept deleting users' email.
(Score: 1, Touché) by Anonymous Coward on Tuesday March 08 2016, @01:07AM
That's what happens when you talk out of the wrong orifice. Wait... That's how she talks anyway.
(Score: 3, Funny) by Tork on Tuesday March 08 2016, @01:25AM
🏳️🌈 Proud Ally 🏳️🌈
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @05:43AM
It also claimed that wiping a server meant cleaning the outside with a soft cloth. Oh, wait, that was the real Hillary.
(Score: 1, Interesting) by Anonymous Coward on Tuesday March 08 2016, @01:41AM
The fourth-grade level is what it takes to get elected in the USA. If you sound elitist and try to make your audience feel dumb, they won't be voting for you.
Private conversations can be different.
(Score: 4, Interesting) by Non Sequor on Tuesday March 08 2016, @02:04AM
You can put this kind of thing together without too much effort, then you keep running it and you cherrypick the better sounding ones to post. Maybe this guy has cut down on the cherrypicking at this point by investing sufficient time into training a neural net, but honestly I don't see that much difference in cherrypicking from an untrained model and training a simplistic learning model.
I recently spent some time working the Shakespeare generator angle. I ginned up a Markov program I found online here: http://www.devjason.com/2010/12/28/shakespeare-sonnet-sourced-markov-text-generation/ [devjason.com]
After the modifications, it will generate something like this:
(Note that I didn't do any cherrypicking to get this one)
You see, poetry's actually a little easier to generate, once you organize your generation strategy in terms of the structure of the poetry. If you train a Markov model from iambic meter, it generates iambic meter. You just have to constrain it to generate 10 syllables per line to make it iambic pentameter.
Actually, I just generate more lines than I need and pick out the 10 syllable lines based on a syllable count list compiled from CMU Sphinx text-to-speech data. I added code so that when it encounters a word not in the dictionary it asks the user and at this point the list is fairly complete for the Shakespeare sonnet corpus.
Then I used the set up a rhyme list based on the Shakespeare sonnet corpus and the standard ABAB CDCD EFEF GG rhyme scheme, carefully adjusting for a small number of non-standard sonnets. Then I closed the rhyme sets into rhyme classes. From there, I can generate, say 500 10 syllable lines, then pick couplets out of the set and arrange them as a sonnet. The last couplet is constrained to end with a punctuation mark.
Early on, I was disappointed that every time I saw a really good line, it would turn out to be a direct quotation from a sonnet. So each line is checked against a hash of the actual sonnet lines. I also hybridized the Markov model between chain length 2 and chain length 3. A chain length 3 model generates each new word based on looking at all of the possible words that occur after the last two words and a chain length of 2 just looks at the last word. I used both types of transitions, but gave the chain length 3 transitions a substantially higher weight. This gives it a way of breaking out of uncommon runs of words while still preferring the transitions that result in words making more sense in context.
What I'm mulling at this point is whether I can sufficiently group words in Shakespeare's sonnets based on some kinds of themes and pick words relating to the theme without overconstraining the Markov model. Another avenue for modification might be to change from generating whole lines to generating pairs of line fragments centered on a line break and match compatible line fragments, in order to get better flow across line breaks. Another option to improve text quality might be to look at parsers or other mechanisms for matching words to parts of speech in context and identify text that matches syntax used in the corpus.
Write your congressman. Tell him he sucks.
(Score: 1, Informative) by Anonymous Coward on Tuesday March 08 2016, @04:24AM
Those are full lines from the sonnets, or half lines. That's quite different from the Drumpf thing and not very entertaining.
I actually recognized three w/o googling: "Music to hear, why" was Sonnet 8, and a couple others that I don't know the number of: "papers, yellow'd with their age" and "When wasteful Time debateth with decay".
(Score: 2) by Non Sequor on Tuesday March 08 2016, @12:52PM
Well, actually there are no entire lines quoted, although in this one the "Time debateth" and "yellow'd with their age" lines are one word off from a real line. Infrequently used words have fewer things that can come after them and "debateth" and "yellow'd" are hapax legomenon. I may want to work on a method to exclude lines that have a certain percentage similarity to an actual line, although I need to think about how to do that efficiently.
These are word-based transitions rather than character-based transitions. "yellow'd with their age" could have evolved into "yellow'd with their stars", or it could have jumped off into any two syllable line ending that follows "with", although "with their age" and "with their stars" were given higher weight.
What's interesting about the RNN approach used for the Trump bot is that the model basically picks up that words are an appropriate unit to split the source into. That's interesting and it's less work to get good generation out of it, but you get less of an idea of what elements of the text need to be modeled in order to satisfy human readers that it's good. It acts more like a Markov model of Markov models that mimics the shape of a training data set.
To me, tuning a generation method to a corpus is a more interesting exercise and it gives me some thought on the identifiable elements of the corpus and how to capture them. Getting rhymes that didn't occur in Shakespeare but were still valid was very satisfying to me even though it was a relatively simple refinement. So I start thinking about things like how if I could isolate meter and part-of-speech equivalence classes, I could allow word substitutions. I keep trying to think about how I can look at ways of algorithmically working with relationships between words that gets elements of Shakespeare (and also hopefully cuts down on direct quotation at the same time).
Ultimately I think that poetry composition is about word choice subject to constraints and I want to study what subproblems within that can be captured by an algorithm.
Write your congressman. Tell him he sucks.
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @03:41PM
Don't let that AC bring you down. Keep going even if others have done it before. In fact write a blog post somewhere and post it here. I am sure we will find it cool.
(Score: 2) by takyon on Friday March 11 2016, @10:09PM
If you make any progress, publish a journal about it. And try my extension.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Tuesday March 08 2016, @04:07AM
DeepDrumpf @DeepDrumpf 8h
He's not going to be able to build, weak on immigration, depart manufacturing. And you know what happens. A disaster
DeepDrumpf @DeepDrumpf 24h
I have two sides. The Republican Party. The middle class. We need a lot of the last thing.
In reply to James Hohmann
DeepDrumpf @DeepDrumpf Mar 6
@jameshohmann @tedcruz [Lyin' Ted] He's a lightweight. And we accept it. All out of it. I mean the record. It's totally the reason I won 10
DeepDrumpf @DeepDrumpf Mar 6
Everybody can be amazing. They are because I'll bring back a hell of a big business. Don't forget you. Thank you.
In reply to Fox News
DeepDrumpf @DeepDrumpf Mar 5
[@FoxNews] That's when we become a country that's unsalvageable. They have losers. I'm against the lobbyists. I am very important right now.
DeepDrumpf @DeepDrumpf Mar 5
Mark my words. We're going to beat ISIS. Come replace the big lie, Obamacare. Believe me.
DeepDrumpf @DeepDrumpf Mar 5
Right now, think of this: We owe China $1.3 trillion. We owe Japan more than that. We have gun laws. I'll bring back our money.
DeepDrumpf @DeepDrumpf Mar 4
This it business. Our president is Obamacare. now, in this is what it's not. Thank you very much. We’re not a cheerleader we're interesting
DeepDrumpf @DeepDrumpf Mar 3
We will never be negotiate, look at the border with Yemen. Again I came out their deals — there we need people. Thank you very much.
DeepDrumpf @DeepDrumpf Mar 3
The military, make it great again. And we’re 26th in the world on President Obama’s illegal executive order.