Model fusion of Conditional Random Fields

This paper introduces two model fusion methods on a series of sub-models of Conditional Random Fields (CRFs): majority voting and feature fusion. The former performs on the results of each participant without any consideration about the underlying details of each sub-model, and the latter takes place on feature level to produce modified feature weights of CRFs to merge all sub-models into a single one. Experiments on syntactic data and part-of-speech tagging problem shows that by dividing training corpus into small parts and using model fusion techniques, comparable results will be achieved.