Abstract:This paper proposes a CRFs-based joint labeling approach for cascaded tasks. In the approach, we construct a multi-layered CRFs based on multiple linear chain CRFs, and introduce the relation among multiple layers in both training and decoding phases. In the training phase, the labels of the preceding layer act as the observed feature of the current layer so that the inter-layer relation is introduced bottom-up, while the global optimal joint labeling is obtained in the decoding phase through a multi-layered Viterbi algorithm. The approach does not complicate the structure of linear chain CRFs, so that relatively lower computational complexity can be preserved. Therefore, the approach has the potential to be a kind of simple and effective solution for joint labeling of practical cascaded tasks. © 2011 ISSN 1881-803X.

A CRFS-based joint labeling approach for cascaded tasks