This directory contains a very simple (and, when it comes to feature computation, very inefficient) C++ program to learn the parameters of a linear chain CRF model using automatic differentiation.

The classical forward algorithm is used to compute the CRF training objective, and rather than explicitly writing the backward algorithm to compute the derivatives, I rely on autodifferentiation. This makes the code much less complicated to write and debug and is only minimally less efficient than a hand-coded backward algorithm

To compile and run this code:

make
./crf sample.conll