Person Name Segmentation with Deep Neural Networks

Abstract

Person names often need to be represented in a consistent format in an application, for example, in <Last Name, Given Name, Suffix> format in library catalogs. Obtaining a normalized representation automatically from an input name requires precise labeling of its components. The process is difficult owing to numerous cultural conventions in writing personal names. In this paper, we propose deep learning-based techniques to achieve this using sequence-to-sequence learning. We design several architectures using a bidirectional long short-term memory (BiLSTM)-based recurrent neural network (RNN). We compare these methods with one based on the hidden Markov model. We perform experiments on a large collection of author names drawn from the National Digital Library of India. The best accuracy of $94\%$ is achieved by the character-level BiLSTM with a conditional random field at the output layer. We also show visualizations of the vectors (representing person names) learned by a BiLSTM and how these vectors are clustered according to name structures. Our study shows that deep learning is a promising approach to automatic name segmentation.