Integrating various features in hidden Markov model using constraint relaxation algorithm for recognition of named entities without gazetteers

We propose a constraint relaxation algorithm to integrate various features in hidden Markov model (HMM). As an example, a HMM-based named entity recognition system is built without gazetteers. Evaluation on MUC-7 English named entity task shows that our system achieves F-measure of 92.0. It shows that various features can be effectively and efficiently integrated using the constraint relaxation algorithm. It also suggests that gazetteers need not be a bottleneck for named entity recognition in newswire domain.