Describes an algorithm for improving the performance of unknownproper noun recognizers, using a statistical framework. We present abootstrapping technique that starts out by using a training set toacquire contextual classification cues, and then uses the results of theinitial phase to acquire additional training data from an unlabeledcorpus. The training set (tagged proper nouns in contexts) is obtainedtrough an application of standard knowledge-based techniques for propernoun tagging, commonly used in information extraction systems
展开▼