Building better conversations

There’s some examples of actually really good experiences where the technology might be the thing that fills it. And so, if you look at the Amazon Alexa skill store and you read some of the reviews for some of the skills, you’ll find things in there that say something like, “didn’t understand what I said”. And it’s not necessarily that the person who created the skill hasn’t catered for what the user said. It’s that the automatic speech recognition- the ASR- might have had some kind of blip. Or the natural language processing might have had some kind of mishap.

These are the instances where those audio snippets and text samples would be sent to those contractors to figure out what was actually said to feedback into the system. But that is an Alexa problem for argument’s sake. It’s not to do with the user. So some of the examples of where things fail- are not because the experience hasn’t been designed very well- it’s because the technology might have had a little bit of an upset. However, there are examples of conversational experiences that are not particularly great. I’m not going to name names because I don’t think that would be very fair.

But if you look- again, the Alexa Skill store or even the Google Assistant action store, there are swathes- more so on Alexa- swathes of really, really poor experiences. And a lot of that is because when the platform was first introduced, there was a whole load of people who found it really interesting. And they were just learning how it worked. They were learning how to build a skill. What does a conversational application look like? And so, all of these people that are learning and learning how things work are creating applications like dad jokes and cat facts and stuff like that, where they’re not really grounded in a proper-use case. It’s more exploratory- can we make this thing work?

A lot of those applications aren’t necessarily created with the user in mind. They’re created from the perspective of, can we technically make this work? And what tends to happen is, there is a lot of common mistakes that people make in those applications from a conversation design perspective. One is the error. We call it error recovery. That’s the wrong term for it. It’s the grounding strategy- so the things that happen that are outside the scope of what you expect the conversation to be.

If you say, “Would you like a another joke,” and someone says something like, “Tell me one about a koala” and you don’t recognise what was said there, you need to respond to that with a grounding strategy and say, “Did you say you want another joke or not” kind of thing. And so, a lot of these old skills- a lot of the things that were built just to see if we can make it work- don’t take into account those kind of recovery strategies to get the conversation back on track.

One of the other examples of conversational techniques, if you like- conversation design techniques which weren’t really taken into consideration a few years back- was allowing the user to move freely, or as freely as possible, within the conversation. If we’re having a conversation right now and it’s on topic- it’s within the domain- and then you, all of a sudden, say, where are you heading to after this- even though we might be right in the middle of another conversation, I’ll be able to understand that. And I’ll be able to say, I’m just heading home, whereas a lot of the conversational experiences that have been created don’t take into account how a user might change tack.

And so, for example, let’s say that you wanted to buy some cinema tickets for a showing. And you’re going through the conversation. You’re going through the process of saying, I want to see the new James Bond, or, I want to see Cats. And it says, what time you want to go? And I say, 8 o’clock. And it says, OK, how many tickets do you want? And I say, three. And it says, it’s going to be X amount of pounds. And if I say, actually, can I have one for half eight- that means I need to step back a few turns, and I’m changing the direction of the conversation.

And quite often, when people are designing these conversational floors, they use a tree-based structure, which means if you don’t say something that’s going to take you next in the tree, it fails. And so, one of the concepts that has been brought about in the practises that we’re employing now in 2019, 2020 is to always leave the conversation open. Most of the time, when people get into the funnel, they’re not going to deviate. But if they do, it’s the times when we fail that we’re judged on- not the times when it goes successful.

And so, we need to be aware of all of the different opportunities and different ways that things might go off-track and be able to cater for them and either take the user to somewhere else to meet that need if their need changes or to get them back on track and get them through that conversation.