Iterative multiple sequence labeling with classifier combination

Traditional pipeline approach causes error propagation and cannot share information among multiple tasks. In this paper, we proposed an iterative approach for sequence labeling problems with classifier combination. The approach is beneficial for both cascaded tasks and multiple separate tasks. We discuss feature selection strategy to increase diversity and obtain better oracle for classifier combination. An averaged perceptron algorithm is used as the strategy of classifier combination. Experimental results on POS tagging and chunking problem show that our approach outperforms pipeline, tag combination, and other classifier combination approaches.