Learning Conditional Random Fields from Unaligned Data for Natural Language Understanding

Abstract

In this paper, we propose a learning approach to train conditional random fields from unaligned data for natural language understanding where input to model learning are sentences paired with predicate formulae (or abstract semantic annotations) without word-level annotations. The learning approach resembles the expectation maximization algorithm. It has two advantages, one is that only abstract annotations are needed instead of fully word-level annotations, and the other is that the proposed learning framework can be easily extended for training other discriminative models, such as support vector machines, from abstract annotations. The proposed approach has been tested on the DARPA Communicator Data. Experimental results show that it outperforms the hidden vector state (HVS) model, a modified hidden Markov model also trained on abstract annotations. Furthermore, the proposed method has been compared with two other approaches, one is the hybrid framework (HF) combining the HVS model and the support vector hidden Markov model, and the other is discriminative training of the HVS model (DT). The proposed approach gives a relative error reduction rate of 18.7% and 8.3% in F-measure when compared with HF and DT respectively.