责任导师 : Thierry ARTIÈRES

Conditional Random Fields for sequence labeling

This thesis, entitled "Conditional Random Fields for sequence labeling", deals with sequential recognition. This task includes a wide range of applications such as speech recognition, handwriting recognition, video analysis, action decoding, stock analysis, biological sequences, medical data, meteorological data analysis, industrical process evolution modeling,...
This thesis focused on two axes based on a discriminant statistical model : the HCRF (Hidden Conditional Random Field).
The first axis takes advantage of the recent advances that occurred in the field of neural networks. These models can be integrated into sequential models and allow to overcome the inherent limitations of basic linear models. Moreover, the introduction of hidden states in purely discriminative models allow the modeling of the temporal evolution of the signal within the same class. Based on these ideas, we proposed a new model, called NeuroHCRF.
The second axis start with the fact that most works, both in handwriting recognition and in speech recognition, does not optimize the criterion used to evaluate the model's accuracy for various practical and theoretical reasons. We show that some limitations of the direct optimization of the criterion can be circumvented by using maximum-margin approaches and defined a new training algorithm.

2011-2013 刊物