Re:How Ninite.com (YC W08) was named by a program -- code in...
by derefr on Friday May 21, 2010 @09:16PM
(#1369501)
> A surprising discovery from this name search was that there seems to be a certain name-iness to some words.
A while ago, I was testing just this assumption in the context of writing a plug-in for a writing program (think Scrivener, not TextMate). I spent a few weeks marking words, names and gibberish as "sounds like a character name", "sounds like a coherent proper/common noun that wouldn't do as a character name", or "is gibberish." I then fed this training data (about 100,000 data points) into a bayesian classifier. The results were okay—it would agree with me to about 80% accuracy—but I gave up on it because I didn't know enough about ML to improve the error, and brute-force scoring another 900,000 words myself just to get a possible 5% improvement seemed a bit bleak.
If anyone is interested in the data set, I could dig it out of storage; it's probably heavily biased toward my preferences in what constitutes a "good" name, though (my backup plan, if it turned out to be a completely subjective evaluation, was to get people to do about a hundred ratings through a website, and then cluster their "creative aesthetics" against others' to help them find people to work with.)
A while ago, I was testing just this assumption in the context of writing a plug-in for a writing program (think Scrivener, not TextMate). I spent a few weeks marking words, names and gibberish as "sounds like a character name", "sounds like a coherent proper/common noun that wouldn't do as a character name", or "is gibberish." I then fed this training data (about 100,000 data points) into a bayesian classifier. The results were okay—it would agree with me to about 80% accuracy—but I gave up on it because I didn't know enough about ML to improve the error, and brute-force scoring another 900,000 words myself just to get a possible 5% improvement seemed a bit bleak.
If anyone is interested in the data set, I could dig it out of storage; it's probably heavily biased toward my preferences in what constitutes a "good" name, though (my backup plan, if it turned out to be a completely subjective evaluation, was to get people to do about a hundred ratings through a website, and then cluster their "creative aesthetics" against others' to help them find people to work with.)