Brendan O'Connor wrote:
On Wed, May 27, 2009 at 8:08 PM, Nancy M <[email protected]
<mailto:[email protected]>> wrote:
I do like the maps, but 50% error -- you would not possibly get on an
airplane with that kind of error rate, would you? And I don't think
I'd want to make decisions about my demographics on something with
that error rate either. Why not take the IPS and bounce them against
whois or something?
This app isn't about that; it's about what places a person is talking
about. You can't use their IP's, the point is to identify locations
in the text of their tweets. (I asked whether the app was looking at
the author's location to help disambiguate because i thought it could
be used to improve accuracy; but this is hypothetical.)
Thanks, that is exactly the point, as explained in the only text on the
page:
"TweetLocations analyses twitter updates and checks if they contain any
geographical locations. Instead of relying on the Twitter location in
your user profile TweetLocations finds the locations you talked about."
:-)
In defense of error rates, if the task is just to get a sense about
what regions of the world someone tends to talk about, then something
like a 10% or 20% error rate might be ok; and it was lower than that
for Chris's and some of the other example twitter users the app was
suggesting.
Well, error rates are a good question. How would a dumb computer know
what the context is in 140 characters? Notice that if you use "My name
is Jack London and I live in Toronto" PlaceMaker ony shows Toronto,
which is impressive!
But here's one case where errors are very bad. One thing I thought
was great about the map UI was that you can see a flag all by itself
out in mexico or something, and be curious what the person is saying
about mexico, and click on it to see the message. If errors tend to
be geographic outliers then they really hurt this use case since
geographic outliers are easy to see and are interesting simply because
they are unusual ("oh, brendan's always boring and talks about
california, but look, one time he talked about switzerland! oops, not
really.")
How could I work around that?
I think the issue with some of the errors the yahoo placemaker thing
was making with my tweets is, is that it's not integrating very well
prior information about how commonly those locations are talked about.
I think "scala" is only rarely used to mean the switzerland canton,
but is quite often used to mean the programming language; but
placemaker is happy to use a rare, unlikely sense of "scala" here.
Well, PlaceMaker is a DB of geographical locations (which you can even
download - http://developer.yahoo.com/geo/geoplanet/data/) and doesn't
compare with a DB of programming languages. It would be interesting to
see how it differs from the other (less open) services out there. Maybe
I'll use Simon Willison's geocoders and only return if there is a match.
http://github.com/simonw/geocoders/tree/master
regards
Chris
