May 07, 2007

Gresham's Law meets the Law of Group Polarization

Over at the web forum Free Republic, yesterday at 4:53:04 PM PDT, "blam" posted without comment a press release (from the American Roentgen Ray Society, via Science Daily) about how "Voice Recognition Systems Seem To Make More Errors With Women's Dictation".

Within five minutes, "USFRIENDINVICTORIA" posted a link to a Language Log post "Sex and Speaking Rate", 8/7/2006). The cited LL post debunked Louann Brizendine's bizarre claim that women speak twice as fast as men, by showing that none of her cited references said anything about the matter, and presenting data from a large study where the measured difference was a 2% average speaking-rate advantage for men over women.

When I saw the Free Republic citation, following a link from our referrer log, I was pleased -- until I noticed that the LL link was cited in support of Brizendine's claim, not in opposition to it. Here is USFRIENDINVICTORIA's contribution in its entirety:

This might be a reason:

“Girls speak faster on average — 250 words per minute versus 125 for typical males.”

OK, I figured, USFRIENDINVICTORIA did a quick web search and didn't read past the Google snip, or something like that. Somebody in the thread will set the record straight pretty soon.

Fat chance.

Instead, over the next three hours or so, there's a little explosion of sexist jocularity:

Careful guys. This is a landmine waiting to explode! ;O)

Discrimination! We need to lower male voice recognition rates to that of female!

i seem to have the same problem...

Oops! I sense a NOW/ACLU moment here...

chauvinistic computers? lol

Woman say four times as many words a day than men...and during an argument, thst rate soars to ten times more.
Maybe that's why I can't stand to watch a talk show with only women.

Just set "WhineyVoiceLevel" to at least 80 (0-100 scale) and NagOvertones=True...

They don’t listen either...

With gusts up to 375 words per minute...

They may say four to ten times more words, but they certainly don't communicate four to ten times as much information. That's why I can't tolerate women's talk shows.
It's like going into the hen house after all the chickens have laid an egg. A rooster makes a lot of noise, but only once a day.

I bought some noise cancelling headphones, but was dusappointed to find that I still hear my wife just fine when using them.

Perhaps the computer needs a how to interprint “nagging” algorithm
could be the fact there is no “reading your mind” algorithm? (I didn’t MEAN to say that you should have known that!)

In general, I'm impressed by the quality of thought and research on the web. But in some contexts, it seems that an intellectual analog of Gresham's Law applies.

Actually, it's worse than that. It's not only that bad ideas drive good ideas out of circulation, but also that certain kinds of bad ideas reinforce themselves, becoming stronger in the people who believe them to start with, and taking root in the people who don't.

This is a particularly noxious form of the Law of Group Polarization, which says that "members of a deliberating group predictably move towards a more extreme point in the direction indicated by the members' predeliberation tendencies" (Cass R. Sunstein, "The Law of Group Polarization", Journal of Political Philosophy 10(2), 175-195, 2002; working papers version here).

As Sunstein explains, "[G]roups consisting of individuals with extremist tendencies are more likely to shift, and likely to shift more (a point that bears on the wellsprings of violence and terrorism; the same is true for groups with some kind of salient shared identity (like Republicans, Democrats, and lawyers, but unlike jurors and experimental subjects).When like-minded people are participating in 'iterated polarization games' -- when they meet regularly, without sustained exposure to competing views -- extreme movements are all the more likely."

In cases like the Freeper thread that I cited, there seems to me to be an additional factor. In addition to the basic group-polarization dynamic, there's a sort of Gresham's Law effect, whereby people with a taste for the rational evaluation of evidence are likely to withdraw from a forum whose participants are so obviously uninterested in the facts of the matter. As a result, as the group opinion becomes more extreme, the standards of evidence get worse and worse, until we get to the point illustrated in that Freeper thread: a freely-available web link is cited to "prove" the opposite of what it plainly says, and 30-odd participants chime in enthusiastically, over a period of several hours, without even noticing.

It's striking how common this sort of reaction is, right across the political, social and geographical spectrum, to "scientific" evocations of social stereotypes. I've noted this in several previous posts about reactions to Louann Brizendine's book, e.g.

Censorship at the Daily Mail (11/29/2006)
Contagious misinformation (12/1/2006)
Femail again(12/2/2006)
Bible Science stories(12/2/2006)
Fabricated but true?(12/3/2006)
The spread of bogus numbers in the meme pool (12/16/2006)
"Gender myths: letting science mislead" (9/30/2006)

A depressing hypothesis:

  • The Law of Group Polarization applies especially strongly to discussion of social (and individual) stereotypes
  • The role of "science" in such discussions is mainly not to supply prestige, much less evidence, but rather to license the expression of prejudices that are otherwise out of bounds (i.e. "politically incorrect")
  • This applies to "pack journalism" as much as to web forums and bar-room conversations
  • Anyone interested in the facts will tend to withdraw from such discussions

[Note that better performance of ASR algorithms on male vs. female voices is a well-established trend, often observed in the literature. The reason is generally assumed to be that higher-pitched voices have more widely spaced harmonics, with the result that short-time amplitude spectra (the basis of essentially all ASR systems) give a less reliable indication of the shape of the vocal tract and therefore of the identity of the vowels and consonants being produced. As far as I know, human speech perception is not affected by this difference, at least not to anything like the same degree (though a more extreme version of the same effect is part of the reason that operatic sopranos tend to be hard to understand.)

This has nothing to do with the rest of the discussion above, it's just to register the point that there was nothing wrong with the original press release -- it simply triggered an episode of the type that I've described.]

Posted by Mark Liberman at May 7, 2007 07:05 AM