Hidden Dynamic Probabilistic Models for Labeling Sequence Data

Abstract

We propose a new discriminative framework, namely Hidden Dynamic Conditional Random Fields (HD-CRFs), for building probabilistic models which can cap-ture both internal and external class dynamics to label sequence data. We introduce a small number of hidden state variables to model the sub-structure of a obser-vation sequence and learn dynamics between different class labels. An HDCRF offers several advantages over previous discriminative models and is attractive both, conceptually and computationally. We performed ex-periments on three well-established sequence labeling tasks in natural language, including part-of-speech tag-ging, noun phrase chunking, and named entity recogni-tion. The results demonstrate the validity and compet-itiveness of our model. In addition, our model com-pares favorably with current state-of-the-art sequence labeling approach, Conditional Random Fields (CRFs), which can only model the external dynamics.