Unbiased conjugate direction boosting for conditional random fields
Kristian Kersting, Bernd Gutmann · Lirias · 2006
Conditional Random Fields (CRFs) currently receive a lot of attention for labeling sequences. To train CRFs, Dietterich et al. proposed a functional gradient optimization approach: the potential functions are represented as weighted sums of regression trees that are induced using Friedman's gradient tree boosting method. In this paper, we improve upon this approach in two ways.First, we identify an expectation selection bias implicitly imposed and compensate for it. Second, we employ a more sophisticated boosting algorithm based on conjugate gradients in function space. Initial experiments show performance gains over the basic functional gradient approach.