Conversational UX - Avoiding The 'Trough of Disillusionment'
Technology: "it's not what you've got, it's how you use it"
More often than not getting technology to work as you hoped takes a little 'wizardry' - it's all about the execution. The right tools, the right services, and the right strategies for implementation. For conversational technologies this is especially important at this point in the technology adoption cycle.
Having recently returned from Web Summit in Lisbon, one of the things that struck me was how how right the keynotes from the Chatbot Summit were when they pitched Conversational UX as exiting the 'trough of disillusionment'*, and other research supports this, including the accelerated growth of voice commerce.
Quite apart from the "AI with everything" at Web Summit, you could see Conversational UX being surfaced in everything from manufacturing for process efficiencies, ecommerce and information discovery through to machine interactions and cars. BMW showcased their new Series 3 due out in March 2019 complete with a new virtual assistant.
Last week I attended IBM's #ThinkLondon event, predominantly to see what was new with IBM Watson Assistant. They are now clearly chasing Amazon with a 'Watson in everything' strategy, showcasing how this could be integrated into a car (or anything else). They have also improved their dialog management suite and introduced analytics too. The Watson narrative is now remarkably similar to the narrative I remember at Artificial Solutions many years ago, even down to the importance of data ownership.
My point with this preamble is that many of the technologies available for building out smart agents and chatbots are very similar, whether you choose technology from Amazon, Google, IBM, Microsoft or any number of startups, and all the big players are tooling up for the Voice First revolution. 'Proper' and useful NLP analytics is perhaps one area where there is still clear differentiation and only a handful of solutions, like Teneo provide truly useful NLP analytics that help you understand and improve the customer experience. Samsung's recent launch of Bixby developer tools may surprise us all with a new paradigm take on how bots are constructed, but there is currently a steep learning curve for developers so it may take some time to gain traction.
Despite the march towards ubiquity, conversational technologies are still viewed by many CxO's as not ready for mainstream adoption. With so many less than inspiring implementations this is understandable, but much of the failure to inspire lays with implementation rather than the technologies.
Whilst user adoption of conversational interfaces is happening far faster than that of mobile, it is a paradigm shift compared to the jump between desktop and mobile, and that shift is far bigger for developers and providers than it is for users. For any chance of successfully launching a smart agent or chatbot it is all about execution. Knowing which NLP strategies to use in various situations is key. How to manipulate context, memory and multiple data sources, and knowing which pieces of the technology stack to use where, in order to deliver a great customer experience requires experience.
These conversational technologies have been around for decades, and the core streams of the tech like automatic speech recognition and natural language processing have come on in leaps and bounds over the last 5 years or so. ASR is so much more fault tolerant than it was when Siri launched, and the full range of NLP technologies is astounding.
Yet when any of these are implemented without a good understanding of their best use or without regard for their dependencies or limitations the user experience can fall flat or worse still completely fail to deliver. Different NLP technology strategies can lead to varying degrees of success or failure, no single strategy will satisfy all the complexity of language and how we use it to communicate. Semantic & Morphological parsers will not understand metaphors, and string and pattern matching alone will not capture all the possible recognitions for most utterances.
Often a hybrid approach is required to deliver the best experience and like any good user experience this requires a little extra work behind the scenes to provide the simplicity needed for a dreamy UX.
To illustrate take a look at this example from a healthcare app. The app knows my age and gender from my profile, yet has guessed I’m talking about pregnancy when it’s missed the context of my question. A robotic process automation tool, like rainbird, or even a simple rule inserted into the data architecture would have alleviated this complete failure in the user experience.
A balanced implementation of technology and techniques is required to generate a truly human-like interaction, and that requires a lot of experience.
The extra work under the hood to hide, manage and exploit the complexity, leaves the UX simple, effortless and natural.
Knowing when and how to handle disambiguation for example can be the difference between an immersive experience or one that simply follows a decision tree approach, and becomes laborious for the user.
All too often you see implementations of 'slot filling' turned into an unnecessarily lengthy series of questions for the user, where a little more thought and creative effort put into language recognition models will deliver a better user experience.
Disambiguation doesn't have to be just about asking the user to explain what they mean. A variety of other methods can be used, eg. using clues from other datasets or other user utterances, or even environmental data can be used to discover intent and context. In much the same way as we do when talking to each other.
As a simple example of the principle, if Siri compared my 'recently played' list or even a list of recent releases, she could have immediately played Thunder by the Imagine Dragons, instead of presenting me with a list of video options to perform disambiguation. This list is useless to me when I'm driving.
"You've got to start with the user experience and work back toward the technology"
Steve Jobs
These strategies can be universally used by applying conversational principles, and understanding how and where to employ different technologies and techniques.
Start with the user experience and work back.
Scalability is often tabled as the reason to use some of the latest tools like automatic natural language generation, transfer learning and one-shot learning etc, but these more 'self-propelled' technologies require much more mature user experiences with exceptional datasets for machine training and testing, otherwise you end up scaling the wrong things - problems.
It's very tempting to use the latest shiny tech and say, "look what I made it do", and over look nuances of the user experience, but it's far better to increase tech complexity iteratively inline with the maturity of your user experience.
Execute a few things really well and users will evangelise your application, be more receptive to your iterations, and more forgiving of the hiccups along the way.
You can achieve a lot with some very basic tools by employing the right strategies. For example, when Alexa first launched the developer tools were very barebones, the technology was far from sophisticated, and they had some annoying deployment rules around session management. I undertook the task of building a skill for Badminton Horse Trials which had numerous datasets or domains to sift through and access in order to deliver a seamless conversation like you might have with someone from the event. We even built a little bit of additional language logic to help with topic switching whilst maintaining context.
The end results was an Alexa Skill that accesses ten different data sources to answer questions about the event in a conversational style, winning critical acclaim from its users and an award at Amazon's developer conference.
Analyse the video below carefully and you'll notice how it switches effortlessly back and forth between topics and entities whilst maintaining the correct context, just as you would when holding a conversation with a person.
Experienced implementation, and knowing which tools to use for which job will deliver the dream user experience every time. Alternatives will leave you stuck in the 'trough of disillusionment'.
With this in mind I'm hosting a small breakfast briefing in central London on Thursday 17th January for CxO's to help them select the right tools, services and strategies when creating their next smart agent (or fix the issues in their current one!), and deliver a best in class user experience.
*see also: Gartner Hype Cycle