Classifying Unknown Proper Noun Phrases Without Context

We present a probabilistic generative model used to classify unknown Proper Noun Phrases into semantic categories. The core of the classifier is an n-gram character model, which is enhanced with an n-gram word-length model and a common word model. While most work has depended largely on context or domain-specific rules for semantic disambiguation of unknown names, we demonstrate that there is surprisingly reliable statistical information available in the composition of the names themselves. Using the context-independent probabilities assigned by our domain independent classifier is sufficient to achieve greater than 90% classification accuracy on typical tasks.