The unscripted world of improvisational theater (improv) challenges performers to spontaneously craft a story from almost no established reality. These actors rely on the power of dialogue to build coherent scenes and expand their shared vision. This is in great contrast to modern dialogue systems (chat-bots), whose non-committal and closed responses often prohibit the progression of conversations.
Spying room for improvement, computer scientist Jonathan May of the University of Southern California (USC), realized that the collision of these two worlds could prove fruitful. “I'd done some improv in college and pined for those days,” May said in a statement. “Then a friend who was in my college improv troupe suggested that it would be handy to have a 'yes-and' bot to practice with, and that gave me the inspiration – it wouldn't just be fun to make a bot that can improvise, it would be practical!”
May decided to pinpoint one of the pillars of improv to help generate a collection of conversational prompts and responses that a bot could be trained on. “The yes-and principle is a rule-of-thumb that suggests that a participant should accept the reality of what the other participant has said (“yes”) and expand or refine that reality with additional information (“and”),” May and co-researcher Justin Cho, also of USC, wrote in a paper presented recently at the Association of Computational Linguistics conference.
Finding a source of “yes-and” dialogues proved difficult for the pair, but eventually, they landed on an improv podcast, Spontaneanation, hosted by actor and comedian Paul F. Tompkins between 2015 and 2019. “Spontaneanation was a great resource for us, but is fairly small as data sets go; we only got about 10,000 yes-ands from it,” May explained. “But we used those yes-ands to build a classifier (program) that can look at new lines of dialogue and determine whether they're yes-ands.”
Movie scripts and subtitles were then inputted into the program and tens of thousands more yes-and examples were added to the SPOLIN (Selected Pairs Of Learnable ImprovisatioN) data set. Now armed with over 68,000 pairs of prompts and yes-and responses, May and Cho could use SPOLIN to train the first ever improv bot (named SpolinBot). Capable of turning a safe and boring chat to funny and wacky, SpolinBot can also generate five response options to help keep the conversation flowing.
To further evaluate the abilities of their bot, the researchers asked a group of people to compare the “yes-and” qualities of four responses given to a prompt.
For example, in response to the prompt “I know alotta women and I’m sure she remembers me,” a standard dialogue system (Persona-chat in this case) said “oh my goodness, I don’t know her.” SpolinBot replied “Yeah she’s a bit of a mystery.” A bot trained with both a standard dialogue compilation and SPOLIN said “So you remember her? I remember her in the shower,” whilst the actual “yes-and” response featured in the development set was “She does. From when you were a boy.”
Overall, SpolinBot fared better than standard dialogue systems, but was still nowhere near the “yes-and” quality of the actual responses themselves. May and Cho have grand plans to improve their improv bot and extend its conversational abilities beyond the yes-and realm. “We want to explore other factors that make improv interesting, such as character-building, scene-building, 'if this (usually an interesting anomaly) is true, what else is also true?,' and call-backs (referring to objects/events mentioned in previous dialogue turns),” Cho said.