I've been playing with stylometry for a different (and completely unrelated to authorship attribution) project recently and this is very interesting. Thanks!
But I do notice this in the JSAN tutorial pdf[1]:
"Building a Better Corpus with Amazon Mechanical Turk
Yes and no. I've seen published authorship attribution (AA) papers that used 13 submissions, so compared to that it is big.
I don't know why this branch of NLP/machine learning has been fairly okay with small corpora. When I did my last project we used 100 authors which was a lot compared to most of the literature at the time.
On the flip side, the only AA papers I've seen that use a large corpus are projects that scrape blogs, and I think it is safe to say that the traditional stylometric features are not optimized for that kind of language.
3 comments
[ 3.7 ms ] story [ 16.2 ms ] threadBut I do notice this in the JSAN tutorial pdf[1]:
"Building a Better Corpus with Amazon Mechanical Turk
"• Only 45 of 101 of submissions are usable!
"• 45 Accepted Submissions."
[...]
"• This corpus is large, diverse, and unique."
(Page 14.)
45 7000-word submissions is large and diverse?
[1] http://events.ccc.de/congress/2011/Fahrplan/attachments/2019...
I don't know why this branch of NLP/machine learning has been fairly okay with small corpora. When I did my last project we used 100 authors which was a lot compared to most of the literature at the time.
On the flip side, the only AA papers I've seen that use a large corpus are projects that scrape blogs, and I think it is safe to say that the traditional stylometric features are not optimized for that kind of language.