How Microsoft Editor Can Make Your OWA Messages More Polite

Artificial Intelligence at its Finest

In message center MC281368 (August 27), Microsoft announced the introduction of “tone detection” as a new capability within the Microsoft Editor component in OWA. Microsoft explains that tone detection will “offer writing refinements to suggest changes to tone. These suggestions will allow the user to communicate more clearly and in a desired tone to deter misinterpretation.” In other words, Editor monitors what you type as you compose messages, and if it considers that your tone is getting a tad snarky or offensive, Editor steps in with suggestions to make your text less provocative with an “improved conversational tone.” This is Microsoft 365 roadmap item 86539.

Microsoft said that the roll-out of tone detection should finish at the end of September 2021, so the feature should be available to all tenants now. However, “English – United States” appears to be the only language which supports tone detection in the set of Editor options exposed by OWA. This is usual as it takes time to refine the basic models to detect certain kinds of text in different languages.

Behind the Scenes

Microsoft Editor (a writing assistant that provides spelling, grammar, and other intelligent writing suggestions) already analyzes text as users type messages in OWA and the Microsoft 365 apps for enterprise to deal with spelling and grammatical errors. In 2020, Microsoft introduced text predictions, first in OWA and later in the desktop apps. Predicting what text comes next is a complicated business and depends on plenty of server horsepower to generate and refine the trainable algorithms from user input. Recurrent neural networks analyze the patterns in user writing to recognize how people put sentences and phrases together. This knowledge creates learning models to deliver the suggestions made as Editor observes what’s input into messages. Over time, the models improve as they process more messages. After a while, Editor will know the probability that someone will follow one phrase with another in email and how they stitch words together in sentences.

Outlook stores model data in a hidden folder in user mailboxes. The data is inaccessible to clients and isn’t shared with Microsoft. The model data is anonymized and does not hold personal details and “no message content is transmitted or stored outside of your organization,” which is how it should be.

Enabling Tone Detection

Not everyone is impressed with the thought of intelligent software analyzing and improving their writing. Fixing spellings and making sure that crass grammatical errors don’t creep into text is fine but telling someone that an automated assistant can improve the tone of their writing is a different matter.

Before Editor applies tone detection, you must enable it in the Editor settings. In a message compose screen, you can expose the Editor settings by clicking the Editor icon (if not, click the […] menu and choose Editor). Tone detection is the last option. So far, “Politeness” is the only setting under tone detection (Figure 1). The fact that tone detection gets its own category makes me think that Microsoft has some additional settings in the works.

Enabling tone detection for politeness (U.S. English)
Figure 1: Enabling tone detection for politeness (U.S. English)

After consenting to allow Editor to monitor message text for politeness, we can try things out by creating a new message complete with some potentially rude statements (Editor might not share your definition of what is rude). To be sure to get Editor’s attention, I composed some blunt and forthright commentary on the usefulness of a recipient (Figure 2).

Editor detects a lack of politeness in an OWA message
Figure 2: Editor detects a lack of politeness in an OWA message

Editor duly stepped in to advise I might be impolite and suggested some alternate text, and that’s when the weakness of the algorithm revealed itself. The suggested alternative did not capture the meaning or intent of the original text. I said something impolite under my breath and tried again. This time Editor was more successful, and the suggested text matched the impolite version more closely (Figure 3).

Editor comes up with a more polite way of saying something
Figure 3: Editor comes up with a more polite way of saying something

Time Needed for Model to Improve

My email writing style tends to be blunt. I don’t like wasting words (too many articles to write). As such, I thought that Editor would pick up more instances of impoliteness than it has done over a week’s use. As noted above, some of the recommendations are off base. I think the number of flawed suggestions will decrease over time. At least, that’s the promise of artificial intelligence and learning models, isn’t it?

Learn more about how Office 365 really works on an ongoing basis by subscribing to the Office 365 for IT Pros eBook. Our monthly updates keep subscribers informed about what’s important across the Office 365 ecosystem.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.