We select and review products independently. When you purchase through our links we may earn a commission. Learn more.

Meet the Chatbot That’s Learning to Improvise

3D rendering of robots working at computers in office

If you’ve ever tried to have a talk with Siri, Alexa or Google’s, um, Google voice, you know there’s a limit to how naturally AI conversations flow. Now researchers have created a chatbot system based on improv comedy to try to improve things—and you can even give it a try.


To create their system, Jonathan May, research lead at the University of California Information Sciences Institute (ISI) along with Justin Cho, an ISI programmer analyst, and prospective USC Viterbi Ph.D. student, decided to turn to the “yes-and” model of dialogue that is a key element of improvisational comedy. The yes-and method means that one person responding to the other first acknowledges the initial statement and then adds to it. It can even be played as a game

“Because improv scenes are built from almost no established reality, dialogue taking place in improv actively tries to reach mutual assumptions and understanding,” said Cho. “This makes dialogue in improv more interesting than most ordinary dialogue, which usually takes place with many assumptions already in place (from common sense, visual signals, etc.).”

So the duo set out to build a large enough data set of yes-and dialogues to fuel their Selected Pairs Of Learnable ImprovisatioN (SPOLIN) project.   


In his hunt to gather the dialogues (all the improv clubs he contacted ignored him), Cho stumbled upon the improv podcast Spontaneation, which aired from 2015-2019. The 30-minute-long segments proved the perfect place to begin building the database, and about 10,000 yes-ands were eventually mined from it. That helped kickstart things, and then another program was used to learn from those examples and extract even more yes-ands by analyzing movie scripts and subtitles. Cho told Mindbounce that the resulting pairs were then validated byAmazon Mechanical Turk workers, a resource to which time-intensive tasks are often outsourced.

The result is a database of over 68,000 conversational pairs that SPOLIN uses to chat. 

“Ultimately, we want to build a good conversational partner and a good creative partner,” May said, noting that yes-ands are only a small part of any conversation. “Today’s bots, SpolinBot included, aren’t great at keeping the thread of the conversation going. There should be a sense that both participants aren’t just establishing a reality, but are also experiencing that reality together.”

You can check out SpolinBot for yourself, but keep May’s sentiment above in mind about today’s chatbots still falling short of full-on engaging conversations.

May and Cho’s work was presented at the Association of Computational Linguistics conference in July.