A detailed comparison of two discriminative algorithms on three corpora with different characteristics has been presented. The discriminative approaches generally perform better than the maximum likelihood based approach.
A direct transfer of the maximum entropy framework to multiple prototypes is difficult, as the use of multiple prototypes leads to nonlinearities and the log-linear model cannot be directly applied any more.
The consistent improvements obtained with weighted dissimilarity measures and multiple prototypes in combination with the improvements obtained by using second-order features suggest possible improvements that could be expected from a combination of these two approaches.