A new training method for sequence data

Under the framework of max margin method, this work proposes a model for training sequence data, which can be solved as a binary classification. However, there are too many samples in the auxiliary classification problem to make the model efficient enough for median to large scale data sets in practice. Therefore, under the additive assumption for the feature mapping and loss function, a simplified model is introduced in order to speed up training. The major advantage of our method is that the new model does not share slack variable for a sequence. This provides the ability to utilize the discriminate information within the sequence and select the discriminative patterns more precisely. Experiment on the task of named entity recognition validates the effectiveness of the new method