Abstract

This paper studies several discriminative models for spoken language understanding (SLU). While all of them fall into the conditional model framework, different optimization criteria lead to conditional random fields, perceptron, minimum classification error and large margin models. The paper discusses the relationship amongst these models and compares them in terms of accuracy, training speed and robustness.