Accelerated Estimation of Conditional Random Fields using a Pseudo-Likelihood-inspired Perceptron Variant

We discuss a simple estimation approach for conditional random fields (CRFs). The approach is derived heuristically by defining a variant of the classic perceptron algorithm in spirit of pseudo-likelihood for maximum likelihood estimation. The resulting approximative algorithm has a linear time complexity in the size of the label set and contains a minimal amount of tunable hyper-parameters. Consequently, the algorithm is suitable for learning CRFbased part-of-speech (POS) taggers in presence of large POS label sets. We present experiments on five languages. Despite its heuristic nature, the algorithm provides surprisingly competetive accuracies and running times against reference methods.