Dynamic Conditional Random Fields for Jointly Labeling Multiple Sequences

Conditional random fields (CRFs) for sequence modeling have several advantages over joint models such as HMMs, including the ability to relax strong independence assumptions made in those models, and the ability to incorporate arbitrary overlapping features. Previous work has focused on linear-chain CRFs, which correspond to finite-state machines, and have efficient exact inference algorithms. Often, however, we wish to label sequence data in multiple interacting ways—for example, performing part-of-speech tagging and noun phrase segmentation simultaneously, increasing joint accuracy by sharing information between them. We present dynamic conditional random fields (DCRFs), which are CRFs in which each time slice has a set of state variables and edges—a distributed state representation as in dynamic Bayesian networks—and parameters are tied across slices. (They could also be called conditionally trained Dynamic Markov Networks.) Since exact inference can be intractable in these models, we perform approximate inference using the tree-based reparameterization framework (TRP). We also present empirical results comparing DCRFs with linear-chain CRFs on natural-language data.