Separating literal from figurative speech is actually quite complicated. A proper interpretation of a statement depends on shared knowledge between speaker and listener, the ease of communication and knowledge of a speaker’s intentions. It’s relatively easy for humans to do this in an instant, but computational models aren’t as adept at identifying non-literal speech.
Researchers from Stanford and MIT set out to create a program that could. They began by asking 340 individuals, recruited through Amazon’s Mechanical Turk, to determine whether a series of statements were literal or hyperbolical. The statements described the prices of an electric kettle, a watch and a laptop. For example, “The laptop cost ten thousand dollars.”
The results seemed intuitive: A statement claiming the kettle cost $10,000 was viewed as hyperbolic, but a price tag of $50 was interpreted as a literal statement. Interestingly, when the number was precise, like $51 or $1,001, participants were more likely to view those statements as literal. In other words, round numbers led to fuzzy interpretations.
The researchers then used this data to build a computational model that took into account a) how near the price was to a reasonable price, b) whether the number given was precise or fuzzy, and c) how big the number was (prices that are higher being deemed more likely to be exaggerations).