What you need to know
'The reason that chatbots are popular right now is not only due to technical innovation, but even more so due to the social media revolution,' says Professor Lee Hung-yi.
Artificial Intelligence (AI) and “chatbots” are changing the way companies around the world talk to their customers, but with only a narrow set of languages used by program developers, some countries have been left behind.
Chatbots are computer programs that use AI to simulate human conversation through voice commands, text or both.
“If you free employees minds from repetitive work, they have more time to do creative work,” says Nelson Chu, the 36-year-old founder of startup HIGH5.
Chu says that most developers of the technology target speakers of English or other European languages. While China’s social media giant, WeChat, has been using chatbots since 2013, its technology “talks” with simplified Chinese characters.
Chu, who has spent the last 20 years in Los Angeles, came to Taiwan earlier this year to further develop a chatbot using traditional Chinese characters.
Very few developers, he says, have built products that communicate in the traditional Chinese characters, common in Taiwan and Hong Kong.
HIGH5 was started in April 2016 by Hong Kong-born Chu, who previously a digital asset manager at large Hollywood studios, and another co-founder who has sinse left the company.
With the help of Taipei startup accelerator, AppWorks, HIGH5's "Robin” helps around 1,100 Taiwanese businesses respond to Facebook messages. In the most successful case, Robin has been able to answer nearly 500 of the 600 questions one company received daily, says Chu.
Samuel Liu (劉家昇), founder of startup e-commerce company Tuanyuannuts, has used Robin since July. For an e-commerce company like Tuanyuannuts, 70 percent of its sales are made online. It receives an average of 20 questions via Facebook’s Messenger service each day. Since it started using Robin, Tuanyuannuts has been able to cut down the amount of time spent replying to customers’ messages from five minutes to one to two minutes.
Retention rates drop when customers notice the chatbot, but most customers do not know they are communicating with a machine, Liu says.
“When customers realize they’re talking to a chatbot, they are less willing to ask questions,” and previous chatbots that reply with "preset robot-like sentences" were so poorly received by their customers that the company stopped using it, says Liu.
According to Lee Hung-yi (李宏毅), an assistant professor of the Department of Electrical Engineering at National Taiwan University, chatbots can “learn” to use more colloquial language by absorbing large amounts of information. To build “Taiwanese-friendly” chatbots, HIGH5 has programmed Robin to learn popular local slang and buzzwords from automatically scanning online content.
During the initial “training” process, Robin first obtains basic company information by asking the company owner approximately 50 questions. It also registers the owner’s tone of voice and mimics their talking habits. After about 10 minutes the chatbot can reply directly to customers based on its “confidence level,” says Liu.
The confidence level is usually based on how similar customers' inquiries are to the questions the chatbots have been trained with, says Professor Lee.
Liu says, if the chatbot calculates its response would be at a confidence level of lower than 90 percent, it asks for the social media manager’s assistance — it will also “learn” from his or her reply … displaying to the person a range of possible answers or learn from his or her reply.
HIGH5’s Chu says that Robin is best suited for companies that receive simple and repetitive questions but it is not as suitable for companies, like architecture firms, receiving inquiries regarding complex design suggestions.
“It really depends on the business, some can take full advantages some can only take some advantages,” he says.
Liu hopes that in the future, chatbots can answer all company-related questions without assistance.
What’s different about Chinese-speaking chatbots?
Professor Lee’s research is focused on machine learning, spoken language understanding, and speech recognition. He teaches two courses in machine learning at NTU and has taught students to design chatbots on their own.
According to Professor Lee, the machine learning process is slightly different for Chinese speaking chatbots from other languages. In Latin-based languages, there is spacing and punctuation between words. In Chinese sentences, however, there is no spacing between words, which are formed by groups of characters.
With advancements in technology, Lee says machines are starting to learn the “natural way.” They can now be programmed with individual Chinese characters and pick up the pattern over time — based on how characters are most commonly grouped to form words — as opposed to programmers deliberately entering segmented Chinese sentences in the training process.
“This is more similar to the learning process of children,” says Lee.
The practice, however, is only common in academia, as it takes longer to train the machine and does not always yield higher performance, the professor adds. It is still easier for companies to train chatbots by programming it with segmented Chinese sentences.
Growing demand for chatbots
In 2016, the chatbot market was said to be worth US$703.3 million worldwide. It is predicted to have a compound annual growth rate of 37 percent from 2017 to 2021, according to PR Newswire.
Chu says growing demand for chatbots comes from increased use of online chatting tools, which is projected to more than double and reach a total of 2.48 billion global users over the next five years.
According to Facebook IQ, “people exchange over one billion messages with businesses every month on Messenger.”
Among the companies interviewed during his market research, Chu found that prior to using chatbots, many companies expect employees to reply “immediately” to customers and go as far as making “less than one-minute replies” a company policy. “If they need to go the bathroom, they need backup,” Chu says.
“The reason that chatbots are popular right now is not only due to technical innovation, but even more so due to the social media revolution,” says Professor Lee. He says chatbots are needed on social media chatting tools because people are used to being able to interact directly with someone from the company. “It’s only natural for chatbots to be used."
In terms of technology, the speech recognition and language understanding abilities of machines have advanced but the ability for machines to make decisions or take follow-up actions remain difficult, says the professor. For example, a chatbot can understand that a customer wants to know the address of the store and send them the address accordingly, but it has a harder time making an online purchase for the customer or tracking product shipment.
Chu believes that this is a “once a lifetime opportunity,” foreseeing a huge market shift in the next decade. “I think chatbots are going to mature really fast so we’re trying to speed up our developments,” he says.
“I have two daughters and I always think about when they grow up what they will do? It’s probably not something that we are doing today,” Chu says.
Editor: Olivia Yang