A matter of tone
January 16, 2016 12:44 AM   Subscribe

The Tone Analyzer uses linguistic analysis to detect emotional tones, social propensities, and writing styles in written communication. Then it offers suggestions to help the writer improve their intended language tones.
posted by Gyan (22 comments total) 23 users marked this as a favorite
 
I can see how this tool would be helpful when used by conscientious people to facilitate better communication... but cynically speaking, I foresee it mostly being used by tone police to shut down conversations, especially where the sides involve social justice versus status quo.

Sorry pal, you may say you're suffering injustices that are killing you by inches but unless you can phrase it with the same cold, distant tones you'd use while analysing the crenellations on a scale model of a medieval fort I'm gonna assume you're just making shit up due to feels. I mean, there couldn't possibly be any rational arguments in a post with more than 12% emotion tones! Don't you know how to Logic™?
posted by timd at 2:40 AM on January 16, 2016 [6 favorites]


linguistic analysis to detect emotional tones

It considers the word "emotional" here negative (pejorative I suppose), but it has no suggested alternative.
posted by Segundus at 2:47 AM on January 16, 2016 [2 favorites]


is it broken? i fed it this long comment from cortex (not picking on cortex - thought it was a good comment and something to learn from) and it seems to think he's angry (emotional tone 35%: anger 88%). yet none of the words seem to be angry words (mousing over "angry" doesn't seem to highlight any words) (and the comment is pretty clearly NOT angry).
posted by andrewcooke at 3:22 AM on January 16, 2016


Robo-tone-police enforcing emotional labor quality is a future I am looking forward to!
posted by srboisvert at 5:38 AM on January 16, 2016 [7 favorites]


Sorry to be vague, but I recently heard a piece on NPR, about diversity in Silicon Valley, and they were using text analysis of their job descriptions to see if they were skewing job listings to the point that they would pull a gender-biased applicant pool.
posted by puddledork at 6:15 AM on January 16, 2016 [1 favorite]


This is a cool idea, but the demo shows we've still got a long way to go before computers can accurately detect and analyze "tone," which is such a slippery and context-dependent aspect of language. A lot of these word classifications seem baffling and arbitrary to me, though it's apparently all based on legit psycholinguistics research. I'll have to take a closer look at that.

The question on my mind here, is what dialect of English is our reference for tone? Presumably it's standard American English, but this isn't stated by the developers. To them it probably goes without saying, but to me, that's a part of the problem--the unspoken assumption that other speakers share your dialect is one of the miscommunication-factors that makes this kind of app relevant in the first place--cf. the recent thread on splaining and what it means to different speech communities.
posted by zchyrs at 6:16 AM on January 16, 2016 [4 favorites]


Pasted in a rant from my own blog and it came up with 100% Anger, 92% Cheerfulness and 100% Negative for Emotion Tone. Sorta kinda mostly not quite right.
posted by tommasz at 6:59 AM on January 16, 2016 [1 favorite]


Individual words don't have a fixed tone (or even a fixed meaning) out of context. The way to make this work better might be what they do with translation - use a massive corpus and match big chunks of text. But you'd need the corpus pre-rated for tone and unlike parallel language texts, that doesn't exist.
posted by Segundus at 7:33 AM on January 16, 2016 [2 favorites]


I ran the US declaration of independence through it, just to get an idea how it works. Results: 100% angry, 85% negative and 96% confident. But also 86% agreeable. Ha, that's the USA I know and love find agreeable.
posted by sively at 7:45 AM on January 16, 2016 [4 favorites]


ndividual words don't have a fixed tone (or even a fixed meaning) out of context. The way to make this work better might be what they do with translation - use a massive corpus and match big chunks of text. But you'd need the corpus pre-rated for tone and unlike parallel language texts, that doesn't exist.

It kinda sorta exists. Reviews (e.g. movie reviews, book reviews, hotel reviews, etc) are popular sources of data for what's called sentiment analysis (i.e. determining whether a given text is more positive or negative about its subject), using a combination of the text of the review and the numerical rating given by the reviewer. But the language used in each type of review is fairly domain-specific, and so sentiment analyzers trained on that data don't necessarily work well on other corpora.

There are also some (small, by machine learning standards) human classified corpora, such as this corpus of 5513 tweets.
posted by jedicus at 7:47 AM on January 16, 2016 [4 favorites]


The Sermon on the Mount is 100 percent angry, 96 percent negative and 62 percent cheerful (and 100 percent applicable to any manufacturer of milk products).

I do not consider this tool 100 percent useful.
posted by Devonian at 8:18 AM on January 16, 2016 [5 favorites]


These numbers make no sense. How can the same email be simultaneously 77% confident and 73% tentative? And if I look at the percentiles, my email is 19% social tone, but if I look at the word count, it is 94% social tone. I think this thing is just broken.
posted by jacquilynne at 8:49 AM on January 16, 2016


"I fed it this long comment from cortex... and it seems to think he's angry." -- andrewcooke

Cortex has had it up to here with all the politely-phrased pony requests from The Tone Analyzer.
posted by rokusan at 9:36 AM on January 16, 2016 [2 favorites]


is it broken? i fed it this long comment from cortex...seems to think he's angry...

Under the honeyed words, it knows the true rage within your soul.
posted by BlueHorse at 9:46 AM on January 16, 2016 [1 favorite]


I'm a very polite atom bomb that's just a couple of stray cosmic rays away from a criticality event.
posted by cortex at 10:07 AM on January 16, 2016 [7 favorites]


While this may be based on legitimate research into the emotional valence of certain vocabulary items, it's a toy--not an accurate tool. The field simply isn't developed enough to accurately do this type of analysis; even if it was, you wouldn't go about it this way (see above comment about sentiment analysis).

Anyone promoting or using this as an accurate tool is engaging in junk science.
posted by Kutsuwamushi at 10:41 AM on January 16, 2016 [2 favorites]


This is ridiculous. Words do not exist in a void, outside the context of sentences. This "tool" is useless.
posted by kozad at 1:16 PM on January 16, 2016 [1 favorite]


Uh, isn't this pretty much how the Avogadro Corp/AI Apocalypse series begins?
posted by quinndexter at 4:15 PM on January 16, 2016


If you click off "Percentile" (at the top of the results), then back on, you get this explanation: "Percentile compares the raw output score to a preset scorecard. For this demo the scorecard is Business Email."
posted by eruonna at 4:31 PM on January 16, 2016 [1 favorite]


Word Count returns the percentage of words per tone and the total number of words for each individual tone trait.

This is gibberish to me.. anyone explain the percentages?

The graph is nice, but needs a few drawing/labeling changes to make it more intuitive.
posted by polymodus at 9:29 PM on January 16, 2016


MetaFilter: Politely-phrased pony requests from The Tone Analyzer.
posted by rhizome at 10:58 AM on January 17, 2016 [1 favorite]


The way to make this work better might be what they do with translation - use a massive corpus and match big chunks of text. But you'd need the corpus pre-rated for tone and unlike parallel language texts, that doesn't exist.

Simple, just build a separate corpus of criticism about the works from which the fragments are taken.
posted by rhizome at 11:00 AM on January 17, 2016


« Older Marriage is like money – seem to want it, and...   |   Happening recently on porches in Houston: Good... Newer »


This thread has been archived and is closed to new comments