Computing boffins strip the fun out of satirical headlines

Real or satire? Teaching machines to tell apart 'BP ready to resume oil drilling' and 'BP ready to resume oil spilling'

Looking for a laugh? You should seek out the ends of satirical headlines and phrases with nouns in them, according to a pair of computer boffins.

The duo, Robert West from École polytechnique fédérale de Lausanne in Switzerland, and Eric Horvitz from Microsoft Research, are the latest academics to try to suck the fun out of life.

In a paper (PDF) published on pre-print server arXiv last week, the researchers compared serious and satirical headlines to find the humour sweet spot.

Their work follows other attempts to predict what might be funny to humans, with an eye to improving artificial intelligence and make chatbots even more creepy easier to interact with.

"Human–computer interactions will never be truly natural without giving users the option to say something funny and have it understood that way," the paper stated.

"It is hard to imagine a computer passing a rich Turing test without being able to understand and produce humor."

But automating comedy is a challenge, the paper continued, and assessing jokes means looking at a "complex narrative structure that is difficult to disentangle".

To avoid this hurdle, and with the aim of unpicking which exact elements make a sentence funny – rather than previous work that aimed to simply predict whether a sentence, paragraph of document is funny – the pair looked to headlines.

Since satirical article toppers are often close to those on serious missives, but with humour being an added bonus, the boffins started there.

"Changing God to Bob Dylan turns the satirical headline 'God diagnosed with bipolar disorder', which was published in the satirical newspaper The Onion, into 'Bob Dylan diagnosed with bipolar disorder', which could appear verbatim in a serious newspaper."

The pair created a corpus of headlines by creating a game,, that asked players to make minimal edits to satirical headlines in a bid to make others believe them. Other players were then asked to rate the seriousness of a range of satirical and player-generated headlines.

"The edit operations used to successfully remove humor pinpoint the words and concepts that play a key role in making the original, satirical headline funny," the paper said.

After a thorough semantic analysis, it concluded that the humorous parts of a satirical headline tend to come at the end of the sentence and in noun phrases. It also found that satirical and serious headlines tend to be opposed to each other, for instance, life and death, humans and animals, or high and low status.

To do the research, the pair used a "chunker" to parse the pairs of sentences into meaningful phrases. In the first example, this would be: [noun phrase - Bob Dylan] [verb phrase - diagnosed] [preposition - with] [noun phrase - bipolar disorder].


Wow, great invention: Now AI eggheads teach machines how to be sarcastic using Reddit


Using 254 pairs where just one chunk had been modified – about half of all pairs that could be chunked – they found that noun phrases were more common in satirical headlines (50 per cent versus 20 per cent for verb phrases).

And, in order to remove the humour, players "overwhelmingly" replaced one phrase with another; they rarely deleted phrases entirely and nearly never introduced a new one.

But the phrase can't be replaced with simply anything, and the new additions tended to be connected in subtle ways – in this case the boffins said it was predominantly a "false analogy" mechanism.

An example given is the opposing good or bad intentions in the combination: "BP ready to resume oil drilling / spilling."

Looking at where the modifications were made, the pair found that most changes were made to the last chunk – leading them to conclude that satirical headlines are often structured in a way that creates a "micro punchline". This is similar to jokes, where the pay off is kept until the end.

The paper did note that satire "is a form of art" and that the humour is often derived not just from following a formula, but also the way it is applied.

However, it did suggest a way of generating satirical headlines using a false-analogy template: First, pick an entity (for example Pepsi), then a central property it has (it is popular) and then choose another entity that holds that property but is opposed to it in some way (Bordeaux wine).

So, a real headline:

“2018 Bordeaux vintage benefits from outstanding grape harvest"

Might generate:

"2018 Pepsi vintage benefits from outstanding high-fructose corn harvest"

The researchers used this formula to generate an alternative header for their paper: "Reverse-Engineering Satire, or 'Paper on Computational Humor Accepted Despite Making Serious Advances'."

As further work, the pair suggested studying the small number of headlines where players got it wrong and thought that satire was serious. This could help understand how people process news, which had implications for our understanding of fake news, they said.

They also argued that the mechanism underlying the game they had created offered a general procedure to identify which parts of a text gave that text its meaning, and that this could be applied to other qualities, such as sexism or hyperbole. The boffins intend to publish their dataset on GitHub soon so that the rest of us can have a play. ®

Biting the hand that feeds IT © 1998–2020