Ahnouncing adaptive psell chekc ni Watson Assistant.

Mitchell Mason
IBM watsonx Assistant
3 min readApr 9, 2019


Qukic, casualy written messags have beocme the norm for comunication between individuals tody (think iMessage, WhatsApp, SMS, Slack, Skype, etc…). And, let’s face it, most poeple have a super lpw bar for the typos and blatant misspellings in the messages that they snd or recieve from others. While thes may be the absolute BANE of some of our existences, the rest of us are tpyically in a hurry and tend to be furgiving to the fact everything theyv jst read in this article is so annoyingly egregious (ok I’ll stop, promise).

But for a machine, a single misspelled word can completely change its interpretation of a request. In a recent internal study, we found that over 20% of message volume contained a typo. Depending on the relative importance of the typo-ed word to the underlying machine learning model, there’s the very real potential that a misunderstanding could occur. This ultimately leads to a poor experience for your end users, who are typically too impatient to reword their initial request when this happens. Yes, believe it or not, we have found that most users of an assistant would rather drop off or ask for an agent in this scenario.

That’s why we built an adaptive autocorrection feature in Watson Assistant that just works. Gone are the countless hours training Watson with intentionally fat fingered sentences. Gone are the tears shed by your analysts over messages like “change my blling address” being classified as SPAM. Gone are the sleepless nights worrying about — ok you get the picture. Watson now corrects this stuff for you.

Our team has developed an autocorrection algorithm that considers the user’s full sentence and the existing training within your skill to make a real-time determination about each and every correction. More specifically, this means that the same misspelling may be corrected differently depending on the context of the sentence it sits within. It also means that the words you’ve used as training in your dialog skill will not be corrected. We’ve rigorously tested these methods and have found that the accuracy of autocorrection in Watson Assistant consistently outperforms the most widely used open source packages.

Some Examples…

Example of autocorrection working based on the context of the sentence

In the above example, you can see that the typo “frm” takes on three different meanings (form, from, and farm) at different points in the user’s message. Watson, using its understanding of the sentence context, can correct each typo to the proper spelling automatically (without any additional training from you).

Example of autocorrection working against existing training data

In this example, you can see that Watson corrected a number of misspellings on the input “I wnat a lg poperoni pisza” to “I want a lg pepperoni pizza.” We had already trained an entity within our dialog skill to indicate that lg is a synonym of large. Even though “lg” is not a real word, Watson knew not to correct it because it was in the skill’s training data.

Try it Out!

Autocorrection is available now for all Watson Assistant customers! To enable autocorrection, simply go into your Dialog Skill’s options tab and flip it on (see below).

To learn more, visit our docs.

And we’re always happy to get your feedback! Feel free to leave comments here or find us on slack at http://wdc-slack-inviter.mybluemix.net/. We would love to hear how this has changed your customers’ experience with your Assistant!