Conditional Random Fields (CRFs) have received a great amount of attentions in many fields and achieved good results. However, a case frequently encountered in practice is that the test data's domain is different with the training data's. It would affect negatively the performance of CRFs. This paper presents a novel technique for maximum a posteriori (MAP) adaptation of Conditional Random Fields model. The background model, which is trained on data from a domain, could be well adapted to a new domain with a small number of labeled domain specific data. Experimental results on tasks of chunking and capitalizing show that this technique can significantly improve performance on out-of-dornain data. In chunking task, the relative improvement given by the adaptation technique is 56.9%. With two in-domain sentences, it also can achieve 30.2% relative improvement.
展开▼
