This .NET library offers Linear-Chain Conditional Random Fields implementations. Two flavors are provided deriving from LinearChainCRF<I, T>: Abstract classes FullLinearChainCRF<I, T> and ConstrainedLinearChainCRF<I, T>. The I is the type of input, all of it, not an element of an input sequence. This is because the library's implementation can access at any time every part of I instances which need not even be sequences; they can be any kind of object. On the other hand, T is the type of the element of the output sequence. The FullLinearChainCRF<I, T> is a classic implementation; the ConstrainedLinearChainCRF<I, T> takes into account the sparsity of the allowed bigrams in the output sequence of T elements in order to speed up computations. Details about the latter are available in a paper.
In order to use the library, create a concrete descendant of either abstract class specifying the arguments needed by its base constructor. Then train the Conditional Random Field using either OnlineTrain or OfflineTrain methods. You can save and load the trained Conditional Random Field using standard .NET serialization of the concrete descendant, which must also be serializable. Sequence predictions can then be performed using method GetSequenceEvaluator which takes the input of type I as its argument. The returned LinearChainCRF<I, T>.SequenceEvaluator offers properties and methods which provide the predicted sequence and statistical figures.
In general, these implementations follow the notes of Charles Elkan's "Log-Linear Models and Conditional Random Fields" with own corrections included, combined with forward and backward vectors scaling for arithmetic robustness as described in Rabiner's "A tutorial on hidden markov models and selected apllications in speech recognition" with corrections provided by Rahimi's "An Erratum for 'A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition'".
This project relies on the following projects which must reside in sibling directories: