Who should guide the conversation--the bot or the person?
Q: In the course of my master's thesis, I had the opportunity to analyze a series of conversations between Alexa and humans. I noticed that Alexa imposes some guidance in the course of the conversation, effectively leaving little freedom for the humans to continue the conversation as they prefer. This forced guidance of the bots in the conversation makes the dialogue unnatural and a bit mechanical. In the design of the conversation, is it fair for the bot to have a guiding role or is it fair for the user to have the freedom to hijack the conversation where they want?
- Silvia Gola
Before we get to your main question—who should guide the conversation, the bot or the human—let’s spend a moment on how human-to-human conversations work.
(For anyone interested in a deeper dive into this topic, I highly recommend the books Talk: The Science of Conversation [Stokoe] and How We Talk: The Inner Workings of Conversation [Enfield]).
When humans have conversations with each other, there are different purposes and styles. When you call a customer service agent, or are speaking to a nurse at the doctor’s office, a script is often followed. One person usually has a set of questions they’ll ask, that the other person answers. Despite the scripted nature of the conversation, it can still flow and feel natural.
On the other hand, if you’re at a dinner party, conversational topics most likely change across the course of an evening, going from one thread to another. But even in these types of conversations, people generally stick to the current topic for a certain amount of time, and if they do wish to change it more suddenly, will often explicitly say so with something like “On on an unrelated note…” to let other people know about the switch.
I start with this because I think we sometimes view human conversation as constantly darting from thing to thing, when much of the time, it’s a somewhat guided, closed-domain discussion.
All that being said: many conversations with bots really do force the user into unnatural paths, when they don’t have to. But the answer isn’t to just “let the user hijack the conversation”; instead, it’s to ensure that typical ways in which a user might want to guide the conversation are respected.
One of my favorite examples of this is a restaurant booking bot. Perhaps you’ve designed it to ask the user “How many people are in your party?” and you’re allowing the user to input a number. Makes sense. But what if they respond with “Do you have any highchairs?” This is a perfectly reasonable, related question, even though it did not answer what the bot asked. In this case, the bot should absolutely let the user take the lead for a moment, and provide the necessary info before returning to the question at hand of how many people. But this doesn’t mean your bot should be required to respond to “How tall is Mt Everest?” when it asks “What time is your reservation?” because that’s not how people talk.
To make bot conversations more flexible, and allow humans to do more of the guiding, two things need to happen:
(1) The technology needs to improve. Although NLU (Natural Language Understanding) is increasing dramatically in its ability to figure out what is intended, it still misses the mark a lot of the time. This is why we must make bot conversations much more constrained than human-to-human ones.
(2) Despite these limitations, there are a lot of ways to design bot conversations to allow users to have more control. As above, make sure to allow common in-domain responses, not just exact answers to what was asked by the bot. In addition, make sure to have flexibility in the order in which people can provide information. Some people want to give a lot up front, some people want to break it into multiple turns. Be sure and do a lot of testing and analysis of how people talk to your bot—you will quickly realize common ways things go off track that are fixable.
We still have a ways to go before talking to a bot is as easy as talking to another human, but in the meantime, we can use design best practices to ensure a bot can have a flexible and successful conversation.
For those reading along: have something to add? Feel free to add a comment below!