Artificial IntelligenceNews

New Advances Want AI to Be Funny, Crack Puns

A pun generator may not sound like a critical work for an artificial intelligence engineer; however, for He, who engineered that in her postdoc at Stanford-it’s an entry to a wicked drawback in machine learning. His goal is to develop AI that’s natural and enjoyable to talk to—bots that don’t merely read us the information or tell us the climate, but can crack jokes or narrate a poem, even weave a compelling story. However, getting there, she says, means going against the boundaries of how AI usually learns.

Neural networks are natural followers, learning of language by churning vast quantities of textual content. If cohesion is your aim, that method works effectively—so well, in fact, that latest development has triggered an ethical debate about whether or not individuals will abuse AI to generate convincing faux information.

However, the resulting sequence is as dry as the newspaper textual content, and Wikipedia articles usually used to train them. Neural networks, in different words, are rule-abiding to a fault, and that makes them horrible jokers. A properly-crafted joke wobble on the fringe of coherency without slopping into nonsense, He says, neural networks merely don’t have the sense to strike that stability. Besides, the entire level of creativity is to be, good, novel. “Even when we had a protracted list of puns it could study from, that would miss the purpose,” she says.

As an alternative, He and her crew, which involved Nanyun Peng and Percy Liang, tried to offer their AI some clever wit, utilizing insights from humor concept. To anybody who’s taunted craft a pun, the instinct will sound acquainted. For a pun to work, He determined it must be stunning in a local context (“stopped to get a hair cut” makes little sense by itself) but additionally have an “aha” issue that holds all of it collectively (in this case, due to the phrase “greyhound”). He and her crew anoint this stress with correct academese: the “local-international surprisal principle.” To make a pun, the neural network is given a pair of homophones (hair/hare) and generates a sentence that’s ordinary with the first word, but evokes surprise when the second phrase is swapped in. Then, to tug it again from the cusp of gibberish, it inserts another phrase that makes the overall sentence a bit more logical.

Tags

Cindy Brennan

Cindy Brennan leads the column Artificial Intelligence at DEXGazette. Garnering information from every corner of the world and writing that concisely is a significant part of her work. Cindy has four years of experience as a journalist. Her functional areas are writing reports and project management. She works closely with her team to edit, and help them to hone stories to deliver a message effectively.

Related Articles

Close
Close