Clustering and Classifying Person Names by Origin

Book One

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 20

Natural Language Processing and Speech Recognition

In natural language processing, information about a person’s geographical origin is an important feature for name entity transliteration and question answering. We propose a language-independent name origin clustering and classification framework. Provided with a small amount of bilingual name translation pairs with labeled origins, we measure origin similarities based on the perplexities of name character language and translation models. We group similar origins into clusters, then train a Bayesian classifier with different features. It achieves 84% classification accuracy with source names only, and 91% with both source and target name pairs. We apply the origin clustering and classification technique to a name transliteration task. The cluster-specific transliteration model dramatically improves the transliteration accuracy from 3.8% to 55%, reducing the transliteration character error rate from 50.3 to 13.5. Adding more unlabeled name pairs to the cluster-specific name transliteration model further improves the transliteration accuracy.