Apappu writeup on Collins

• This paper describes some novel parameter estimation algorithms for tagging problem as an alternative to CRFs. These new algorithms are based on voted or average perceptron

• Author highlighted positive aspects of averaged perceptron in comparison to ME (with GIS ), it would be interesting to see any comparisons to ME with other popular training methods.

• I think the mistake bounds for inseparable data may suffer if the data is sparse.

• Overall, this paper has been a good read primarily because it stands-out of all the readings related to different variants of markov models.