Bots are the big trend in government, but how easy are they to build – and how useful can they be? GovInsider entered our testing lab to find out more about the process, how people interact with them, and the value they can provide to public service delivery.
“Won’t it be nice if you could ask questions about where you pay your parking fines, or where you get subsidies and grants?”, Jacqueline Poh, CEO of Singapore’s GovTech Agency, asked at Innovation Labs World. “This chatbot idea would allow us to do it”, she said. Singapore has announced bots for its Tourism Board, Municipal Services Office, and is investigating many other options.
There are plenty of tools on the market that allow people to build bots without extensive coding. Some integrate with chat apps like Facebook Messenger or WeChat. Others create a standalone service that can be embedded into a website.
There are a couple of tools in the market, including Botsify and Chatfuel, which promise a bot almost instantly. We tried Rebot, the most rudimentary, but it allowed us to track the interactions in the back-end and didn’t need to be synced to a third party messaging system.
We decided to pretend that we are Singapore’s Art Musuem, feeding information into the bot from a Frequently Asked Questions page. This allowed us to create an interactive service that mirrored common enquiries from citizens.
We found that the bot only understood the exact question it was fed. It generates a random response if it doesn’t understand your question.
The ‘knowledge’ tab tracks what users are saying, and this allows adaptation of the bot. We saw huge amounts of different enquiries, even when testing internally, including random messages of affection and one word queries. Some will try to test the limits of the bot’s intelligence. This is where machine learning comes in – the most advanced bots need good AI that can understand the many different ways that people will ask questions.
If a bot doesn’t understand your response, it generates a random response. This needs a generic customer service message to guide people to the right answer.
The ‘knowledge’ tab is useful on this software, because it logs questions that users have asked, and you can then key in custom replies based on those. This is the most important thing for a bot to have – the ability to learn about customer enquiries and continually improve. We didn’t have AI, so we manually made some changes to go through the process.
It’s plain that chatbots are easy to create, and can be built free-of-charge with all the tools available on the web. But the greatest limitation is language – our bot couldn’t yet grasp semantics. This is a problem with free tools.
We have informal ways of saying things, and responding to this needs advanced AI. There are examples where this is effective. Microsoft has built Xiaoice in China, a bot that has been used over the past year to interact with citizens with great effect.
A journalist in China even interviewed it without alerting Microsoft, drawing responses to light-hearted questions like ‘what is the happiest thing in your life?’.
People enjoy chatting with bots, partly to see where these automated responses slip up. Equally, it’s really difficult to predict when someone is asking. The fact that this machine could answer those responses is extremely impressive. It stands in marked contrast to Tay, another Microsoft bot on Twitter that quickly learned from online stimuli and started to share racist blog posts.
“Our goal is to extend the conversation,” Dr. Hsiao-Wuen Hon, Corporate Vice President, Microsoft Research Asia of Microsoft Asia-Pacific R&D Group said at a recent briefing session on AI. Researchers want to increase the ability of these tools to interact with humans.
Translation into other languages is still “a long way off”, he noted, with difficulties in finding out the right context. Bots, therefore, are limited in medical professions because you can’t guarantee 100% accuracy, he noted. “If your life depends on it, you better make sure it is 100% accurate.”
Bots can run simple tasks though, like checking traffic congestion or an online savings account. In our case, we suspect a simple Q&A would be better, unless the museum wanted to invest in a tool that displays art works through the messenger and showcases the latest exhibitions. We suspect that people are mostly likely to want the opening hours, admissions prices and information on eating options.
Bots show great potential in areas where there are a great many enquiries that need a dynamic response. For example, a calculation based on your personal information, or something like traffic where the AI can crunch different options and tell you an optimal route.
Sadly (or perhaps luckily), they can’t yet write articles about government innovation. Although we see that Xiaolce has given it a go!