Three ways to improve your conversational AI

Techniques to streamline your AI’s dialogue, increase its effectiveness, and satisfy your users.

Conversational AI is a great way to help your users get what they need, when they need it, the way they want to get it. Still, there are several common shortcomings in Conversational AI systems.

Figure 1 We want users to get what they need from conversational AI. Photo by Daria Nepriakhina.

This post will walk you through three simple tricks to improve your conversational AI:

· Ask for less

· Give clear choices

· Copy editing

Ask for less

Conversational AI systems are designed to collect all information required to complete a task. This information can come from contextual information or asking the user a question.

Every question the AI asks the user is an opportunity for the conversation to fail. Wherever possible, ask fewer questions, or ask easier questions.

Let’s consider a car insurance example. The user is calling to check on the status of a claim they just filed. Fundamentally we need to figure out which claim the caller wants to talk about and fetch the details for the claim.

Figure 2 Process flow for getting claim details

Technologists may ask too many questions

A technology driven approach might start with the knowledge that a “claim details API” is needed. A developer might even say they have an API that retrieves claim details by claim ID. Great! All we need is for the user to enter their 11-digit claim ID and we’re set. The design works backwards from what’s available technically.

Figure 3 The availability of an API dictating a conversational flow.

While easy for the developer, this may be challenging for the user. They must enter an 11-digit ID accurately. In voice environments, it may be challenging to correctly capture all 11 digits even if the user knows them.

Friendlier design can make the questions easier to answer

We can make it easier for the user to enter the claim ID. Rather than asking for all 11 digits, we can ask for fewer digits (for instance, the last four digits). The backend still needs a claim ID, but we can make the question easier for the user. There are less chances to make a mistake speaking or typing the few number of digits.

Figure 4 Ask an easier question and do a little more backend work.

There are several variations on “ask an easier question”:

· Ask a different question: For a claim, that may be the claim date. For many use cases either an ID or a date is unique for a single user.

· Take a picture: Some mobile apps find and extract the necessary information from a photo.

The downside of “ask an easier question” is that you are still asking a question. For an automobile claim, you are expecting the caller to have the claim in front of them. This constraint may cause the user to fail at their task. Can we make it even easier?

Make a reasonable assumption (and possibly confirm)

Let’s flip the conversation around. Instead of starting from the technical perspective, put the user front-and-center. Why is a user calling about an automobile claim? They were probably just in a car accident and following up on the claim they submitted.

A user may have multiple claims, but they are probably calling about the most recent one.

Figure 5 Make a reasonable assumption and let the user confirm it.

Figure 5 removes the burden from the user. Now they don’t have to have the claim in front of them. We can provide information to the user, instead of demanding information up front, and still get the information needed for our backend data retrieval. It’s much easier for the user to confirm information than to provide it. Plus, this makes the conversational AI seem much smarter, and the user’s confidence in the system will increase.

This pattern requires use of additional context so that information is not given to a nefarious third party. For instance, the claim search example is driven by knowing who the caller is (perhaps tying an incoming phone number to a claim policy).

Skipping a question by making a reasonable assumption is not always possible, but it is a powerful technique for making more effective conversational AI systems.

Next, let’s look at a way to improve questions that you need to ask.

Give clear choices

In the previous section we improved AI effectiveness by removing questions. Sometimes you can’t remove a question and need to ask it. Some questions are explicitly choice questions — requiring the user to select from a fixed list. Imagine asking the user to select an apple or an orange. This feels simple, but how can we ensure users successfully choose an option?

Figure 6 Choosing an apple or an orange. Apple photo by Marek Studzinski on Unsplash. Orange photo by Xiaolong Wong on Unsplash.

Choice questions can be more difficult to implement than you’d expect. Let’s look at a couple of pitfalls.

Beware of yes/no confusion

It’s not always obvious how a user should answer a choice question. If you ask the user “Would you like an apple or an orange?”, they may say yes!

Rewording the question can increase clarity for the user:

· “Please choose apple or orange”

· “Which fruit would you like — apple or orange?”

Conversational AI works best when the dialogue copy is adapted for the conversational medium. While technically equivalent, these variations do not perform equivalently in dialog.

Buttons make things even easier

Some users prefer not to type or speak their responses. In a chat use case, it may be easier to click a button than to type the response. In voice use cases, users may infer the ability to press buttons on their keypad — 1 for the first option, 2 for the second option, etc. When assessing the user’s response to “Please choose apple or orange”, accept “apple” or “1” as equivalent.

Figure 7 Don’t be afraid to let users choose on their keypad. Photo by Charisse Kenion on Unsplash

Buttons may not “feel” high-tech, but used appropriately they can help your users be more effective.

Be resilient to small mistakes in choice selections

When you ask your users to make a choice, either the user or the AI can still make a mistake. In chat interfaces, the user may typo “appel” for “apple”. Rather than saying “I did not understand”, the AI should assume the user meant apple. For instance, Watson Assistant has autocorrect for spelling and fuzzy matching as well which is resilient to these kinds of minor mistakes.

Figure 8 Don’t turn using an AI into a spelling test. Photo by Jessica Lewis on Unsplash

In voice applications, if your choices are domain jargon you may need to train a custom model to recognize the choice options. You know what users are going to respond with — whatever choices you give them, be sure that your speech service can accurately transcribe them. For this fruit use case, be sure to include some “apple” and “orange” entries in your custom model training.

Figure 9 Speech to text services may make transcription mistakes. Will your conversational AI be resilient to minor mistakes?

After training a custom model, you may still encounter mistranscriptions. Your speech service may occasionally transcribe “apple” to “appeal” or “hackle”. In this case you can add some of these mistranscriptions to your Watson Assistant implementation.

Figure 10 Updating Watson Assistant to be resilient to speech mistranscriptions. Here we allow “appeal” and “hackle” transcriptions to mean “apple”.

Choice questions are common in conversational AI systems. Use these tips to make sure your choice questions are as effective as possible.

Copy editing

Earlier we reviewed how precise wording improves AI effectiveness by generating better user responses. Better dialogue text improves the user experience. This increases user satisfaction and helps them get their tasks done more quickly.

Figure 11 Are your users just waiting for the AI to stop talking? Photo by Elyas Pasban on Unsplash

Word diet

Every word of dialogue increases the work for your users. In chat use cases they need to read the words; in voice they need to listen to (and wait for) the words. Review your dialogue for unnecessary text. Consider these examples.

· “Please listen carefully as our menu options have recently changed”. When was the last time this text was useful? If you only call once per year, you don’t remember the options anyway. If you call every day, you only need to hear this once. This message is 99% useless and my number one IVR pet peeve.

· “Your call is important to us”. This is something you should show, not tell. Better to actually treat your users with respect than to talk about it.

· (Long explanation of how to proceed). If you need to give a lengthy explanation so that the user “does it your way”, consider a technical change the requires less explanation.

Use correct punctuation

Conversational AI systems use text to speech engines to speak dialogue to users. These engines use clues from punctuation on where to add stresses and pauses within a sentence.

I remember a clever T-shirt reading “Let’s eat, grandma” with a bolded comma, and the tagline “punctuation saves lives”. Punctuation may not actually save lives in your AI system, but it will make your dialogue text sound better.

Review your dialogue for punctuation:

· All sentences end with a period.

· Compound statements are broken by commas.

· Choice lists include the “Oxford comma”. (The final comma in “Please choose A, B, or C.”)

You can even use automation to create audio files of all dialogue in your virtual assistant.

Conclusion

This article included several techniques for improving your conversational AI. Consider these techniques as you build a new conversational AI or improve an existing system. Any AI system benefits from continuous improvement and iteration. These practices have been proven to improve conversational AI systems.

Find more tips in my Conversational AI book!