Have you ever kept yourself awake at night agonizing over an email sent earlier in the day? Did she realize I was being sarcastic? I hope I didn’t come off as harsh or rude. I should have added a smiley face!
IBM would like to help. The company’s Watson artificial intelligence system, most famous for winning Jeopardy in 2011, has a brand new service called Tone Analyzer, designed to help avoid such pitfalls. The service is currently available in an experimental phase.
A user inputs a piece of writing into a field, and the Tone Analyzer breaks the language down into three categories: emotional tone, social tone and writing tone. The emotional tone component looks at feelings expressed in the writing and categorizes them as cheerfulness, anger or negativity. The social tone category analyzes how much the writing expresses the personality attributes of openness, agreeableness and conscientiousness. Then, the writing tone category examines the extent to which the writing is analytic, confident or tentative. From there, the Tone Analyzer offers a percentage “score” for how much of the writing falls into the three categories as well as the total number of words that fit each of the subcategories. It then offers synonym suggestions to bring out different feelings.
“Words such as ‘difficult,’ ‘disappointing,’ ‘blame,’ ‘critical,’ ‘inferior,’ these are the types of words that in general reflect negative tones,” says Rama Akkiraju, the leader of the team that developed the Tone Analyzer. “Words such as ‘together,’ ‘let us,’ and ‘we,’ these are very inclusive words that contribute to agreeableness.”
The Tone Analyzer grew out of the Watson team’s research on how writing reflects personality. Watson researchers have spent several years using linguistic analysis to see how personality traits like agreeableness or neuroticism can be inferred from text messages, emails and other writings. Using this research, building a service that analyzes writing tone was fairly simple, Akkiraju says.
As an experiment, I entered various pieces of writing into the Tone Analyzer. The program still lacks the ability to understand context, which can cause it to falsely flag words as angry or cheerful, and it definitely still needs a context-sensitive human to put its suggestions to use.
A New York Times article about the Greek bailout crisis generated a number of interesting insights about the tool’s ability to make sense of tone. The word “bitter” in the phrase “bitter medicine” was flagged as negative, but the suggested alternatives, including "venomous," "virulent" and "caustic," hardly seemed less so. "Segregated" and "obscure" were both offered as less negative alternatives to "isolated," though their meanings are wildly different. And "hot" and "fashionable" were both suggested as alternatives to "popular," which might make sense if the sentence were about skinny jeans rather than Germany’s finance minister, Wolfgang Schäuble.
I tested a wildly hostile email I received after placing a house-for-rent ad on Craigslist (the emailer was angry that our ad said “no cats”). The Tone Reader skipped over the words “stupidity” and “idiot” and gave the email high marks for agreeableness and conscientiousness. I beg to differ. Another email, this one from my husband, flagged the “missing” in “I’m missing you” as negative.
The Tone Reader correctly identified the negativity in food critic A.A. Gill’s seminal restaurant review of Paris's L'Ami Louis, the "worst restaurant in the world," though it didn’t have the contextual understanding to flag the review’s most damning phrases (the foie gras that “tastes faintly of gut-scented butter or pressed liposuction” or the escargot like “dinosaur boogers”).
Akkiraju says future versions of the Tone Analyzer will almost certainly understand context better.
“Understanding sarcasm, humor and innuendo that are part of human expression is a tough challenge that the natural language processing research community at large is working on,” she says.
While the service can be used by private citizens, Watson’s creators see it as being an important tool for business. A company could use the Tone Analyzer to scan thousands of online reviews of its products, flagging negative ones for human analysis. The tool could also read phone logs at call centers, marking especially good or bad interactions.
Down the line, the Tone Analyzer could be embedded in an email system and personalized, Akkiraju says. It would learn what words and tones you tend to like, and adjust itself accordingly.
“Maybe your style is to be more agreeable, so we would recommend more words that you tend to use and provide a more personal set of suggestions,” she says.
The IBM team has no date for when the Tone Analyzer will be more widely deployed. For the moment, I think I'll use the service to scan my own emails before sending—keeping a watchful eye for wonky suggestions, of course.