It's all about timing in mission to make AI bots laugh

Laughter is one of the effortless joys of life, something which we share and bond over, as it flows spontaneously through conversations around the world.

However, behind this everyday pleasantry, a hugely complex neurological mechanism is at play, reacting in milliseconds to the specific and yet obscure stimulant we call humor. Scientists trying to recreate laughter in AI bots to simulate natural conversations have run into difficulty. Around the world, chatbots are a booming business, and improving their services by integrating this elusive human quality is proving to be a major technical challenge.

Researchers from Kyoto University in Japan have tried to teach a robot named Erica to laugh. They hope that natural conversations between AI systems and humans can be optimized. Laughter can help lubricate a conversation, and could make customer service experiences much more favorable. Empathy is a key emotion which is explored in this project, and conversational laughter is taught to Erica by gathering training data from more than 80 speed-dating dialogues. These were read by male volunteers, and responses acted out by female actors.

Not all laughter is equal, and solo outbursts are distinguished from other types, such as embarrassed laughter or subtle chuckles made out of politeness. Machine learning programming was then utilized to dictate timing and appropriate responses.

The resulting algorithm mirrored appropriate triggers, although challenges remained. Deciding on when to join in on shared group laughter was a trait which required the most nuance, with of course most laughter occurring in a conversation being singular. Avoiding laughter which just triggered more of the same perpetually was also a challenge to researchers, with such an absurd blip instantly taking users out of the suspension of reality.

Humans were, of course, placed in to check every stage of the process, with clips from Erica played to 130 volunteers. These people reviewed the algorithm for understanding, authentic human simulation, empathy and naturalness. As algorithms improve with ever more data sets, it may even be possible for AI chatbots to have their own unique sense of humor. Always in a cheery mood and willing to chortle away, a consistent near-human experience could be delivered by the chatbot industry in years to come.

More cost-effective and consistent mood-wise than human customer service operators, laughs such as those from Erica may be commonplace in years to come. This, however, may be at least 20 years away, researchers say. Nuances of complex human behavior are not replicated overnight, and physical behaviors such as eye contact, gestures and cadence will all need to be fine-tuned, to say nothing of the challenge of accounting for the cultural differences between international markets.

Studies such as these reveal just how sophisticated a casual human conversation can be. With something that may take the most talented scientists in the world decades to produce, it is worth bearing in mind that most newborn babies have their first laugh in the first three to four months in life. These same babies, however, may be parents themselves by the time it is possible to share their laughter with a robot.

Barry He is a London-based columnist for China Daily