Conditional Random Field

A Conditional Random Field ( CRF ) is an undirected graphical model. Often CRFs are used for tagging of sequential data. This means that the CRF receives a sequence X as input and returns an equally long sequence Y. In contrast to Hidden Markov Models (HMMs; another, but oriented model for sequential data ) may access the complete information of the input sequence a CRF at each point, while a HMM provides only the current input.

Training

There are various learning algorithms, such as the gradient method or the quasi -Newton method to train CRFs. In this case, some sequences are determined from which both the input and the desired output is known. The learning process then tries the parameters in the CRF adjusted so that for as many sequences in the training data, the correct output sequence is predicted.

Applications

CRFs have been successfully applied to various problems, such as:

• Information extraction from natural language texts
• Classification of proteins
• Prediction of the secondary structure of DNA and proteins
• Image recognition and image restoration