Sparse Higher Order Conditional Random Fields for Improved Sequence Labeling

by Xian Qian, Xiaoqian Jiang, Qi Zhang, Xuanjing Huang and Lide Wu

This paper adds higher-order features to the CRF model with exact inference. The feasibility is caused by the sparsity. That's to say although we include higher order features in the feature vector, it seldom fires. So in the inference stage (since we have to calculate the gradient), the computational complexity will not go exponentially and the sparsity makes the exact inference possible. This paper extends the terms for linear-chain CRF into configurations which can be used for higher-order features. That's the main contribution.

I will write something related to this problem soon. First I have to finish some optimization code.