Bidirectional Sequence Labeling Via Dual Decomposition

In this paper, we propose a bidirectional algorithm for sequence labeling to capture the influence of both the left-to-right and the right-to-left directions. We combine the optimization of two unidirectional models from opposite directions via the dual decomposition method to jointly label the input sequence. Experiments on three sequence labeling tasks (Chinese word segmentation, English POS tagging and text chunking) show that our approach can improve the accuracy of sequence labeling tasks when the two unidirectional models individually make highly different predictions.