Heh. Makes sense, as people tend to speed up their speech when they're nervous, and become nervous in response to deliberate speech modulation of the sort you describe.
To the GP I'm sure you could 're-tone' many messages, but as presented here the results would likely be unsatisfactory. Doing so effectively would require second- and third-order analysis and now you're getting into Hard Problems.
I also noticed this, and it seems like a pretty big flaw. It analyzed one of my emails as very confident because it saw the words "clear" and "sure", but in context, they were used in the phrases "None of this is clear to me" and "I am not totally sure"
My initial impression is it doesn't seem analyze groups of words very well. It will pick emotional words out of the middle of a sentence and apply the emotion to the whole text, when the context of the sentence changes the meaning of the word.
Quick example, "I hate your guts and I hope you die." -- Hope is picked out and noted as cheerful.
I think it's aimed at business communication. It picks out terms that might be taken the wrong way, instead of trying to rewrite your whole communication.
when I was doing my stab at this problem (see comment above), the contextual power structure was important.
that's why I coupled my model with flow analysis on the mail server end to assign subordination relations between parties when possible.
A general solution is OK for liaisons but that's about it.
The IBM demo appears to be just positioning Watson as a general business concierge here and doesn't seem to be concerned with a true assist on what Gardener calls interpersonal intelligence.
Being able to write concisely and well doesn't necessarily mean you can permeate communication walls.
A valid computer solution, in brief, involves continuous kernel application during the authorship; how the user responds to the system is itself part of the analysis. This isn't just binary sentiment analysis of product reviews using an SVM - that's childs' play compared to this.
Even with the best solution I came up with, it's still just culturally specific ... cultural to the geolocation of the persons and the heritage of the people, but also to the industry itself.
Something that may show weakness in one industry may show wisdom in another, hostility in another, and humility in another.
For instance, say you met someone with whom you had a serious romantic interest. Would you exchange business cards, and then arrange a 9am skype call the following tuesday? It's just all so contextual.
I'd be interested in whatever else you can talk about. I find this topic interesting and can see the usefulness of it (though I'm imagining something more comprehensive than this demo). The kind of nightmare that I imagined when going into this article though, was some kind of genetic/optimization algorithm that will just iterate on some text until it has achieved the optimal, "most emotionally/etc correct", communication. Obviously, this assumes some highly accurate model of people's emotional perceptions [not to mention, what the words actually mean :)], but I think it's still a somewhat scary thought.
Anyway, would love to hear more about what you're working on.
I have an open source version, but it's intentionally about 800 commits behind my current work (which I haven't touched in a while tbh). If you are really interested, let's take this offline. kristopolous (at) gmail.com
I tried just that: "your work is simply not engendering confidence. If I were working for a sewerage works I'd recommend you for promotion, since what you are producing is more of the proverbial. By Friday things will have improved one way or another. I am sure you understand what I mean." The particular things that Watson saw in this just aren't what I see in it.
my frustrometer project, which does the same thing and predates this by a few years (albeit less flashy), at http://frustrometer.com/ can handle that fine.
I didn't finish all the build out, because I got really cold feet in the early pitches ... moved on to other things
should I get back on this? It's always so hard to tell if discouragement is a true representation or just people being internety. I also grossly lack self confidence. I don't know if I need pills, therapy, or what.
Keep building it. Especially if it interests you. It interested you enough to start it in the first place! Just don't build it with the expectation that you'll garner fame or fortune.
I imagine a lot of people could use some kind of OSS sentiment detection, even if it's not perfect.
It's interesting and I will keep it bookmarked for occasional use, but the classification into emotional, social and writing tones seem very arbitrary. At least part of the problem is that it seems to be working on single words, most of which are weighted more toward one classification or another based on aggregate use rather than any real context. Tools like this rarely seem to do any analysis at the phrase level, even though phrases of two or three words often contribute significant substantive or structural context. This is odd considering that we already have techniques for identifying statistically improbable phrases as intermediate lexical units (often used for detecting plagiarism).
Is anyone aware of work going on in this area? I'm quite interested in lexical analysis, but I'm wary of investing a lot of time just to reinvent the wheel.
There is a race to create a virtual assistant and do question answering.. the big players are IBM Watson, Apple w/ Siri, Google, Microsoft w/ cortana, facebook M, and Amazon Echo.
I kind of like it how the example on this page is helping someone make their email more harsh, when it's already using exclamation points and starting sentences with "but" and is already heeps of negative.
"We can't blame the economy" is passive aggressive, but it gets coded as green.
Not being passive aggressive is good, but perhaps writing the sample email in a way that makes people want to work for this person would be a good idea instead.
I'm not sure what you mean "It gets coded as green". It doesn't consider an entire statement, it considers word choice.
Similarly, it says that green means that it's placed in the "writing tone" evaluation category, in that it tries to determine whether the writing comes across as analytical, confident or tentative. More like instructions or analysis and less like judgment or conversation.
It's better to look at the demo in a different way. I mean, if I put your post into the analyzer it tells me that it's impassioned, and of that it's not very cheerful. If I put mine in, it looks very matter of fact. It's more analytical and matter-of-fact.
The two posts imply that kind of difference in tone, and it's pretty cool to see.
It's got a way to go, but trying different examples is interesting.
I mean that they are building a tool and the primary example is showing a typical passive-aggressive business email and the user is trying to make it worse.
This is a strange reason to write a tool, and perhaps symptomatic of the corporate culture that sought to produce it.
Teaching people to be nicer would be more interesting, and it's clear it is not dwelling on the right elements since it neglects to detect the overall tone of the email as being problematic.
I can see a more advanced version of this being deployed as part of automated admin packages for commenting sites such as reddit and HN. Any comments of sufficiently negative tone gets auto-downvote/hidden/flagged, etc.
In fact I wouldn't be surprised if it's already part pf the internal tooling.
I think IBM really inspired people with Watson, but they have failed to productize it and now they are using "Watson" as a marketing label for any A.I. related technology at all -- be it something very good that they started work on years ago (Speech Recognition) or something really bad that they aquired like AlchemyAPI.
28 comments
[ 4.4 ms ] story [ 66.0 ms ] threadSeems like the logical progression, if not full outright automatic generation of messages.
To the GP I'm sure you could 're-tone' many messages, but as presented here the results would likely be unsatisfactory. Doing so effectively would require second- and third-order analysis and now you're getting into Hard Problems.
Quick example, "I hate your guts and I hope you die." -- Hope is picked out and noted as cheerful.
that's why I coupled my model with flow analysis on the mail server end to assign subordination relations between parties when possible.
A general solution is OK for liaisons but that's about it.
The IBM demo appears to be just positioning Watson as a general business concierge here and doesn't seem to be concerned with a true assist on what Gardener calls interpersonal intelligence.
Being able to write concisely and well doesn't necessarily mean you can permeate communication walls.
A valid computer solution, in brief, involves continuous kernel application during the authorship; how the user responds to the system is itself part of the analysis. This isn't just binary sentiment analysis of product reviews using an SVM - that's childs' play compared to this.
Even with the best solution I came up with, it's still just culturally specific ... cultural to the geolocation of the persons and the heritage of the people, but also to the industry itself.
Something that may show weakness in one industry may show wisdom in another, hostility in another, and humility in another.
For instance, say you met someone with whom you had a serious romantic interest. Would you exchange business cards, and then arrange a 9am skype call the following tuesday? It's just all so contextual.
This rabbit hole runs deep.
I'd be interested in whatever else you can talk about. I find this topic interesting and can see the usefulness of it (though I'm imagining something more comprehensive than this demo). The kind of nightmare that I imagined when going into this article though, was some kind of genetic/optimization algorithm that will just iterate on some text until it has achieved the optimal, "most emotionally/etc correct", communication. Obviously, this assumes some highly accurate model of people's emotional perceptions [not to mention, what the words actually mean :)], but I think it's still a somewhat scary thought.
Anyway, would love to hear more about what you're working on.
I didn't finish all the build out, because I got really cold feet in the early pitches ... moved on to other things
should I get back on this? It's always so hard to tell if discouragement is a true representation or just people being internety. I also grossly lack self confidence. I don't know if I need pills, therapy, or what.
I imagine a lot of people could use some kind of OSS sentiment detection, even if it's not perfect.
Try therapy before pills
The synonyms for acknowledge on their example are not very good.
Here is a link to the demo to test it out yourself. http://tone-analyzer-demo.mybluemix.net/
You are a retarded goat.
Is apparently a positive sentence.
Is anyone aware of work going on in this area? I'm quite interested in lexical analysis, but I'm wary of investing a lot of time just to reinvent the wheel.
http://cs224d.stanford.edu/syllabus.html
There is a race to create a virtual assistant and do question answering.. the big players are IBM Watson, Apple w/ Siri, Google, Microsoft w/ cortana, facebook M, and Amazon Echo.
"We can't blame the economy" is passive aggressive, but it gets coded as green.
Not being passive aggressive is good, but perhaps writing the sample email in a way that makes people want to work for this person would be a good idea instead.
Similarly, it says that green means that it's placed in the "writing tone" evaluation category, in that it tries to determine whether the writing comes across as analytical, confident or tentative. More like instructions or analysis and less like judgment or conversation.
It's better to look at the demo in a different way. I mean, if I put your post into the analyzer it tells me that it's impassioned, and of that it's not very cheerful. If I put mine in, it looks very matter of fact. It's more analytical and matter-of-fact.
The two posts imply that kind of difference in tone, and it's pretty cool to see.
It's got a way to go, but trying different examples is interesting.
This is a strange reason to write a tool, and perhaps symptomatic of the corporate culture that sought to produce it.
Teaching people to be nicer would be more interesting, and it's clear it is not dwelling on the right elements since it neglects to detect the overall tone of the email as being problematic.
In fact I wouldn't be surprised if it's already part pf the internal tooling.