Enhanced sequence labeling based on latent variable conditional random fields

Abstract

Natural language processing is a useful processing technique of language data, such as text and speech. Sequence labeling represents the upstream task of many natural language processing tasks, such as machine translation, text classification, and sentiment classification. In this paper, the focus is on the sequence labeling task, in which semantic labels are assigned to each unit of a given input sequence. Two frameworks of latent variable conditional random fields (CRF) models (called LVCRF-I and LVCRF-II) are proposed, which use the encoding schema as a latent variable to capture the latent structure of the hidden variables and the observed data. Among the two designed models, the LVCRF-I model focuses on the sentence level, while the LVCRF-II works in the word level, to choose the best encoding schema for a given input sequence automatically without handcraft features. In the experiments, the two proposed models are verified by four sequence prediction tasks, including named entity recognition (NER), chunking, reference parsing and POS tagging. The proposed frameworks achieve better performance without using other handcraft features than the conventional CRF model. Moreover, these designed frameworks can be viewed as a substitution of the conventional CRF models. In the commonly used LSTM-CRF models, the CRF layer can be replaced with our proposed framework as they use the same training and inference procedure. The experimental results show that the proposed models exhibit latent variable and provide competitive and robust performance on all three sequence prediction tasks.