“We think http://feministphilosophers.wordpress.com/ is written by a woman (71%).”
From their website: “About GenderAnalyzer BETA: We created Genderanalyzer out of curiosity and fun. It uses Artificial Intelligence to determine if a homepage is written by a man or woman. Behind the scene, a text classifier hosted over at uClassify.com has been trained on 11000 blogs written by men and women. In our lab it seems to works pretty well, we want to see how it performs on the web! We hope you like it!”
See the analyzer here I’d appreciate comments from someone who knows more about how these things work about the assumptions they use to guess the sex of a blog’s author.
[...] Feminist Philosophers (woman, 71%). -34.925770 [...]
It just barely thought my blog was a man. It’s pretty convinced that Jodi Dean’s blog is male, and Megan McArdle’s blog is male.
I can’t help with that expertise question, but I have no shortage of opinions or questions about this and most other topics. My opinion: I don’t know about fun, but it seems more or less harmless.
My question: Is it a hopelessly circular self-perpetuating bias to assume that 2 Swedish men will live up to their international (?) reputation as culturally predisposed to lean gently left on most social issues, including gender equality? I mean, how safe is it to rank-order 2 Swedish men as potentially more female friendly and therefore more likely to be objective and knowledgeable about blind gender analysis because of a whole pile of social factors related to their “Swedishness” compared to other “sensitive” or “not so sensitive” men?
All that aside, 71% accuracy is only a pass in some courses. That was my average on my college level “who cares” courses like marketing. I played the odds and hit the 50/50 thing without even trying. I read somewhere near 1/4 of the text for another 10% of the mark and used “common sense” to get through the rest.
Depending on the subject matter in question, it’s pretty easy to hit a 71% accuracy rate on lucky guesses. I would like to see if it works though. I’ve been on way too many sites where the “author” of the post responded to me personally with something that was inconsistent enough with previous claims (not to mention apparent level of education) that I had serious doubts that the person responding even spoke English.
It guessed mine wrongly, but said that on the whole my blog was “gender neutral.” If you click on the survey results, the Analyzer has so far made about twice as many correct guesses as incorrect guesses.
http://freepollkit.com/index.php?browse=viewresult&pollid=1013057
Gender-izer SEZ about my blog
We guess http://www.imascatterbrain.blogspot.com/ is written by a woman (53%), however it’s quite gender neutral.
Is this correct? Yes No Don’t know
I whimper about my boyfriend’s ex, I go to MyVagina.com, and basically carry on like an hysterical teenage GIRLY GIRL.
SO – how would I answer this; “Is this correct? Yes No Don’t know?”
Now I AM confused!
Looks like uclasify is entirely automated – it works by looking for common patterns. This can produce great results – but also very odd ones.
You train it by passing it examples of whatever types you want to use for classification and it then analyses then for common word usages, phrasing, etc. – and then when you pass it an unclassified web page it will compare that to those patterns and make an “educated guess” based on that.
So, if women tend to use longer sentences, the word “elephant” or some other thing it recognises more than men do, then it will add “women points” on for any of those things that it sees, and vice versa. At the end, add up all the points and declare a winner!
Back in 2002 Bayesian analysis (which is what this would probably be) suddenly became rather big – and I remember using SpamBayes to try and analyse incoming email to spot spam.
Wow, by a mere 55%, it guessed me (correctly) as female.
I guess I’ll have to post more modal logic in my personal blog.
Ha! It guessed me wrong!
We guess http://nickyreiss.blogspot.com/ is written by a man (52%), however it’s quite gender neutral.
… but glad to be ‘gender neutral’!
We think http://churchbells.tumblr.com/ is written by a man (87%). which it isn’t. hum.
The reference to artificial intelligence suggests neural networks to me. Animals, especially human beings, have quite special cognitive features that ordinary digital computers do not. A classic example is pattern recognition, where you may not have necessary or sufficient conditions, but rather something like family resemblance. We’re very good at classifying patterns and neural nets represent a pretty successful attempt to build an artificial system with the same abilities.
Gender is a good example, I’d bet, of a pattern without necessary or sufficient conditions. And it looks as though their system issues something like judgments about degree of the exemplification of “male” or “female.” If so, it says the front page it analysed was fairly strongly female.
The Stanford Encyclopedia has a useful article on neural nets, though it is a bit dated. My cogsci colleagues who work with computer modelling maintain that the whole idea of neural nets as models of how the brain is actually working has fizzled out. Here’s the SE url:
http://plato.stanford.edu/entries/connectionism/
I suspect the algorithm is similar to the one described in this paper by Argamon, Koppel, Fine, and Shimoni; there are several other gender analysis programs online that build on their work.
Brandon et al, the blog says their analyser has been trained on 11,000 blogs. That is another big sign that they are using a neural net. With neural sets you have layers, including an input layer and at least one ‘hidden layer’ and an output layer. learning results in adjusting weights of the connections between nodes in each layer.
They may have started out picking out the relevant inputs and the initial weights according to such a formula, but the net gets changed according to whether the output is correct or not on each trial.
The analyser ought to be significantly more accurate than the continued use of one formula. I suspect it is, since female philosophers tend to look male according to the more simple tests.
that is, the weights adjusted are among the nodes at one leven and those at another level.
I doubt it’s a neural net. It’s using uclassify in the background, which they say is based on the same code as the Cactus Spam Filter, which is using Bayesian techniques according to a few reviews I’ve tracked down.
More info on the kinds of techniques involves are here:
http://en.wikipedia.org/wiki/Bayesian_spam_filtering#Process
AD, thanks! Really interesting.
I checked out the Stanford Encyclopedia for “Bayesian classifier” and “Bayesian network,” and got zero returns. So I’m tempted to say that we don’t know about them, which makes me feel less culpable, but clearly we should. And in fact some philosophers do.
It still sounds like the results are about as reliable as a deck of Tarot cards. I use them to make straightforward yes or no decisions where I have nothing to lose and too many things to choose from–like flipping a coin. About the same “prediction” odds too. Based on a 3/5 card spread with black meaning no and red meaning yes, red falls about 55 to 60% of the time. Which is just slightly better than the odds on a heads is yes coin flip–about 52% with Canadian quarters. Something about the way they’re weighted. If I’m kinda leaning toward a “yes I deserve those new shoes” it’s pretty simple to phrase the questions to get the results I want. Not that I’d use that info to try to delude myself about any kind of psychic ability or anything.
No point in getting all Will Smith in I Robot about it, though. Like I said, this gadget seems harmless enough as long as I’m not spending any money on it.
Xena, you may well be spending money on Bayesian classifiers if you paid for your computer and it came with an email program. They’re also used for facial recognition tasks, which means that when one is probably paying for their use as one pays for products from companies that use them. Ditto handwriting and voice recognition, I would suppose, though those might be done by neural nets, which still might be implementing Bayesian principles.
There are also good reasons to think that our perceptual systems realize Bayesian procedures as we transform indeterminate data into determinate perceptions. See here:
http://neuro.neusc.bcm.tmc.edu/?sct=gfaculty&prf=53
I meant the genderanalyzer; I hope these people are marketing to businesses, not individuals. I’m also not fibbing about how poor I am or how much grief I’ve been through. I haven’t paid for a computer since ’01. I bought it used and had so much trouble with it that I left it in the pile of basement stuff for I don’t know how many years. It was confiscated with everything else the province took when they decided they would rather throw me out of my home than help me with my son’s disability.
I blog from school or from a kind relative’s computer. Does her info show when I blog from there? I’ve also spent a great deal of time taking care of non-academic things like job searches and cooking tips at public libraries. No biggie. I had to ration the influence of the Gadgetocracy on my son anyway. We didn’t even have cable tv for a number of years. Too much visual clutter when he was little. I only allowed him to watch rented movies that we discussed with his therapists first.
Of course, you could say I’m paying for all that stuff with the tech fees my school is adding to my tuition, but I think I’m still paying less here than what it would cost to wire up my house (if I still had one).
I’m going to check that link now. Hopefully it’ll give me some insight as to why the newer faster tv ads with the noticeably amped up visuals give me such a screaming headache. I can’t even watch MTV anymore. Unusual for a GenXer, but you know what they say about that slowly boiling frog vs. the frog that goes from cool to scalded in seconds.
Okay, how about GenderAnalyzer REALLY analyzes how many people will try this?