Will you marry me? It’s not a question you would typically associate with home loans. Yet, it was the top query asked of Ubank’s RoboChat virtual assistant when it first launched in May 2017. The chatbot had been developed using IBM Watson’s Conversation API to meet customer demand for home loan information outside of UBank’s call centre office hours.
Born out of a Hackathon at the IBM Watson Summit in Sydney in mid-2016, the app was developed to answer a range of common questions about home loans. The conversational tone of the chatbot was brought in line with UBank’s value proposition: “simpler, better and smarter”, which included having a response for any cheeky questions thrown its way.
Anticipating curveball questions is easier said than done when creating a chatbot or virtual agent application. This is something we learned at our own Hackathon, held in collaboration with Amazon Web Services, this month.
Teams comprising developers, speech experts and marketers came together from our international operations in Sri Lanka, the Philippines, New Zealand and Australia to build a new Salmat service or capability using the Amazon Echo/Alexa Virtual Assistant. Salmat has had a long-established presence in the Australasian Speech market so our teams were excited to get their hands on the devices and build the apps.
Salmat’s winning Hackathon team with Salmat's CEO: (left to right) Dhanushka Jayakody, Rebecca Lowde (CEO), Peter Nann and Lisa Jackson (Nawaiz Khan, Richard Collins and Ben Blowers not pictured).
Aside from the satisfaction of seeing our hard work come together in working Alexa apps, one of the key takeaways from the experience was the challenge posed to developers of anticipating such endless permutations of questions and answers. Humans are, after all, unpredictable beasts. The phrasing of our questions, our intonation when speaking, the slang or regional dialects we use – there are many factors that influence our language, person to person.
This goes to the heart of the challenge with AI – how do you predict human behaviour?
Learning the language
AWS has broken down the language used to speak with chatbots and virtual agents into four key components. These four components are the starting points for developers when building their virtual assistant applications. The four key features consider how the user phrases their original question, what the implied intent is of the user, what action needs to be taken, any specifics around how or when the action needs to be taken, and then how the conversation between the bot and user is maintained. And you thought all you had to do was programme a question and response.
In technical terms, these four characteristics are defined as follows:
Utterances: The spoken/typed input that you say/type to the AI
E.g. Alexa, what time is it in New York?
Intent: What you want or intend to achieve with the AI
E.g. You want to find out the time in New York
Entity: An item (often a thing) you mention that makes your intent more specific
E.g. In the utterance, 'What's the weather like in Seattle?', the word 'Seattle is an entity. Entities can usually be any one of a list of similar items.
Dialogue: The back and forth conversation between the user and the bot.
User: Alexa, what are the odds on Australia to win?
Alexa: Is that for Rugby Union or Cricket?
Alexa: OK, the odds on Australia to win the Rugby Union against New Zealand are ...
What Alexa says should also not be underestimated. Defining Alexa's 'prompts' is possibly one of the most under-estimated challenges in the Speech space.
Setting customer expectations
Setting customer expectations upfront is one way to set the limits of how customers interact with your chatbot/virtual agents. On Alexa, you can programme the dialogue to give an explanatory line to the user upon opening the app. The same is true of chatbots. However, the challenge for the creators of the app or chatbot is to summarise the instructions concisely and quickly (so that the user doesn’t lose interest or become overwhelmed with information).
Considering which prompts to include in the app (and for which utterances they relate to) is another challenge in itself. For developers and marketers alike, it means really getting into the head of your customer – How do they speak? How are they likely to respond to a question? Would they use industry terminology or conversational language? Are there specific terms or slang you need to consider?
Consider a betting app, for example. Not only do you have to define the language (utterances) used to find out the odds and then place a bet, but you have to consider how you define the teams in question. The All Blacks, for example could also be ‘New Zealand’ or ‘the New Zealand team’ or even ‘the Kiwis’ ... the list goes on. It takes considerable time to map the many variables – worth considering when building your own app.
AI experts recently spoke on the subject at an Interactive Minds, event: The Rise of the Machines: AI in Marketing. Sponsored by Salmat, the event brought together speakers Michelle Zamora, Head of Marketing at IBM Watson; Ben Leane, Group Manager at Avanade; and Noor Hammad, Head of Product Marketing at Stackla to discuss their experience of AI.
IBM Watson’s Zamora described AI as “the modern day URL” as it understands, reasons and learns. In her role, Zamora works closely with businesses to help them harness the power of AI to improve their own systems. She and the other speakers pointed to a number of working examples that are already in the market, from H&M’s chatbot, which recommends outfits in the slang terminology used by the younger audience, through to Soul Machine’s realistically human, virtual agent Rachel.
Ben Leane talks Interactive Minds delegates in Sydney through the technical side of AI.
Leane, whose business Avanade is a Microsoft partner, spoke about the technical side of AI. He summarised the three key design principles to consider when designing an AI app.
3 design points to consider: