There are a lot of things that chatbots have yet to master and high on the list is small talk. But researchers at Facebook think the best way to make software prattle away is to give it a personality.
The team crowdsourced their chatbot personas from 11,000 online workers on Amazon’s Mechanical Turk. Workers were asked to roleplay in pairs and to give statements describing made-up personas, including their likes and dislikes. The crowdworkers’ chatter was linked to these description statements and used to train the chatbots. This gave the chatbots a rough idea of what kind of person might make what kind of utterance.
The chatbots were evaluated by humans who rated their engagingness, fluency and consistency. “I spend my free time writing,” said one bot, breezily. Its persona included the statement “I am a writer”. When asked if it was married it replied, “haha, no time. I have got a novel to finish.”
Advertisement
Chatbot bores
The personas improved the chatbots performance overall. The most personable was judged to be nearly twice as fluent as a bot trained using film subtitles, for example. But bots that stuck rigidly to their persona struggled to remain interesting. “They ended up less engaging, because they just talk about themselves,” says Jason Weston at Facebook AI Research.
In another conversation, one bot bemused its human conversation partner by talking about its cat immediately after saying, “I am allergic to pets unfortunately.”
“These systems are all very interesting but we’re still far away from the kind of conversational AI where you can get into detailed conversation about a topic,” says at Heriot Watt University in Scotland.
But Weston notes that these chatbots are learning something new – how to get over that initial moment of awkward silence. “It’s for when someone first meets the chatbot and you get to know each other,” he says.
Article amended on 29 January 2018
We corrected Jason Weston’s name and the reality status of the workers’ personae
