Skip to main content

Outline

Imagine walking into a store. You need to buy a new phone, but you don't really know much about phones. You walk up to one of the employees and ask for help. "What kind of phone are you looking for?" They ask. "I don't know," you say. "I need a new phone, but it doesn't have to be fancy or anything."

The employee nods and asks a couple of follow-up questions: do you care much about photography? What about playing games on your phone? And what does your price range look like? By the end of their question spree, they summarise it for you: "Sounds to me like you're looking for a mid-range phone that will last you at least four, five years. You don't need the latest processor because you'll only use the phone to message friends and family, so that means there's more budget for a good camera."

You nod, that sounds about right! The employee smiles: "Perfect. In that case, I know just the one for you."

Communicate understanding, use grounding

What the employee and you just did is called grounding in academic literature. It's an iterative sort of 'game' between (at least) two people. You try to make sure that you understood each other, usually by summarising what it is you think the other said. After your summary, the other has the chance to correct you if you misunderstood them, you get to speak, and the roles are reversed.

Grounding helps us show engagement towards our conversational partners, and in the context of product search - like the phone example - it also boosts our competence: it shows that we understood the actual question, even if people struggled expressing themselves.

Humans are so good at grounding that you probably never even noticed you did it (although you may after reading this!). For Virtual Agents (VAs), though, grounding does not come naturally. If you use generative AI you'll have to instruct it to do so, or you may have to write the acknowledgements entirely - activities that will take time and resources.

So is it worth your time to create grounding VAs?

Image by Annie Spratt on Unsplash (click image to go to source!)

Testing a VA that uses grounding

To investigate what happens when VAs engage in grounding, Papenmeier and Topp (2023) set up an experiment within the domain of e-commerce. They made three mock-up conversations with a VA that acted as a shopping assistant for a new laptop. In each conversation, the VA would either use:

  1. No grounding at all.
  2. Acknowledgements, e.g. words like "right" or "okay".
  3. Paraphrasing.

They invited 105 participants to compare the different conversations, split into three groups. Each group would read and compare two versions. After reading the conversations, participants had to respond to a scale for perceived conversational engagement and competence of each VA in assisting to purchase a new laptop.

Paraphrase to show engagement & competence

I'll cut to the chase: acknowledgements did nothing. The VA that used acknowledgements seemed as engaged and competent as the VA that used no grounding at all. If anything, the VA that used acknowledgements seemed slightly more polite or friendly - but also less to the point than the VA that used no grounding at all, making neither better than the other.

Paraphrasing showed the best results. The VA that paraphrased its users showed better scores on engagement and competence, compared both to the VA that used no grounding and the VA that used acknowledgements. Participants remarked that they felt more understood when paraphrased:

“I don’t really need them to feel like they’re talking, I just want to know that they actually understand what I’m saying” (Papenmeier & Topp, 2023)

Paraphrasing in action

When to use grounding

So what does that mean for your VA? Should you never use acknowledgements and paraphrase every message your users send?

Although that may be a tempting conclusion, reality is of course more nuanced. Acknowledgements may still be relevant to use. For example, they may create valuable buffer time in voice conversations for users to listen to the rest of the message. And paraphrasing every user message is definitely not a good idea either, as it would just slow the conversation down and make your users frustrated.

Instead, use paraphrasing whenever your VA has to do something that involves any risk for your users. People are less likely to use a VA for tasks that are risky (which you can read more about in the article below). For example, when your VA is recommending a product to purchase, or conducting a transaction on your users' behalf, its perceived competence becomes critical - and that's exactly where a good paraphrase can give your VA the boost it needs.