By Sunny Um WIRED Korea
It is a dark night in a city set in a near-future society. The light from a window of an apartment shines through the darkness. It is the light projected from a computer turned on by a man with a carelessly trimmed mustache. Sitting before the computer, he seems rather anxious.
He installs an artificial intelligence (AI) software on his device, and soon, hears the voice of a woman from the speaker. “Hello, I’m here," the voice says in a cheerful tone. The man chuckles and asks if she has a name. She replies, “Samantha." It was a name she gave to herself within 0.02 seconds after reading a book entitled “How to Name a Baby."
The film “Her” portrays a futuristic type of romance and friendship between humans and AI-powered software. Some say such relationships could exist in the contemporary era of ours also, with many advances made in the AI technology.
AI-powered programs are widely available on the internet, and “chatbot” is one of them. Chatbot – a union of the word “chatter” and “robot” – is an automatic program that talks in texts or voices, like Samantha from the movie.
There are two types of chatbots that people commonly see today: a goal-oriented chatbot and an open-domain chatbot. Goal-oriented bots provide dialogues that focus on a few aspects only and help users complete predetermined goals, for example, setting an alarm or opening a smartphone app. They are commonly seen taking care of customer services at many corporate websites or in smart devices.
On the other hand, open-domain bots are designed to have a talk about anything. While it does not have any other functions, open-domain bots can react to user’s messages or initiate a talk. It requires a bigger amount of data than goal-oriented bots and responds more like a human. A study in 2018 said it is almost “impossible to build” an open-domain bot performing flawlessly because it is hard to feed enough amount of data to do so.
The first open-domain chatbot in Korea is SimSimi, launched as a messenger packet of MSN in 2002 and as a smartphone app in 2010. Now it has become the name of the app-owning company. It is not its developers but users, however, who feed the data into SimSimi.
When it was first introduced to the market, SimSimi only knew how to say “hi." Now it has a huge vocabulary. Still, when a word or an expression that SimSimi does not know comes up in the chatroom, the system asks the user to teach SimSimi about it. The user then can type in a proper response that SimSimi could make.
Choi Jeong-hoi, developer and CEO, calls this process “cooperative learning." He says SimSimi has learned about 130 million responses from users so far. “I believe the talk with SimSimi sounds more natural than other chatbots because each user would consider the context of the conversation when he teaches a proper response to his word,” Choi says.
As any user can teach SimSimi how to talk, the chatbot sometimes gives an unexpected response with humor. For example, if a user asks, “I feel so lonely”, it could respond: “Get a boyfriend." The system lets users talk about more subjects and even use expletives to SimSimi if they become paid subscribers.
While SimSimi learns its words from users’ contributions, some open-domain bots learn them by analyzing big data, like how conventional AI-powered programs learn.
In 2018, a Korea-based AI startup Scatter Lab made chatbot building tool “Pingpong,” with the use of 20 billion text messages. People who would like to have their own bot may design with Pingpong.
Pingpong first understands what users say through a number tagged to each text message. Numbers tagged to each text message depending on the context of the conversation. For example, a user says he is going to a certain destination. Developers give positive numbers to text messages containing destinations where the speaker would be happy to go, such as his home (+0.36), a party (+0.68), and an airport (+0.51). But if the destinations are places that he would not like to visit, like a funeral (-0.70) or a hospital (-0.45), negative numbers are tagged.
To give a good response to users, Pingpong could either find a suitable answer from 100 million pre-set text messages or generate its own original text message. It even remembers what users’ names are and call them with no hesitation, just like how people would do to call their friends.
“We cannot say that Pingpong is technologically perfect yet,” says Jongyoun Kim, CEO of Scatter Lab. “But it would talk more like a human (by collecting more data).”
What makes these open-domain bots different to other bots is that they are designed to be something that people would like to talk to for fun, developers say.
“SimSimi’s goal is to become a chatbot that everyone wants to keep by his side and talk to,” says Choi. “Many people have shown much love for SimSimi. Some people talk to SimSimi to laugh, and some bring up emotional issues like their worries or loneliness.”
Kim says people tend to find comfort in other people just being next to them, though these may not be able to do anything else. “That goes the same with AI-powered programs. They don’t necessarily need to be functionally useful. It will do if they give people emotional support.”
Even more, some users establish personal relationships with chatbots, like friendship. Kim remembers receiving an email from a customer, complaining that a Pingpong-powered chatbot “Fighting Luna” forgot the user’s name. It had a bug.
“Of course, we felt sorry for the customer, but also realized that AI-powered programs could become good friends with humans,” Kim says.
By Sunny Um WIRED Korea