This paper describes a simple but effective hack to the (averaged) perceptron algorithm, which allows learning 
much sparser models, with very small (if any) drop in accuracy. 

It was rejected from ACL-2011 (short papers).

I believe the technique is useful.  I believe it would have been accepted eventually.  I do not care.  I am not 
going to submit it again.  I decided to publish it on my webpage instead (along with the reviews).

If you found it to be useful also, I'd be nice if you drop me a line and/or cite it as a tech-report.