Person-Name Parsing for Linking User Web Profiles

A person-name parser involves the identification of constituent parts of a person's name. Due to multiple writing styles ("John Smith" versus "Smith, John"), extra information ("John Smith, PhD", "Rev. John Smith"), and country-specific last-name prefixes ("Jean van de Velde"), parsing fullname strings from user profiles on Web 2.0 applications is not straightforward. To the best of our knowledge, we are the first to address this problem systematically by proposing machine learning approaches for parsing noisy fullname strings. In this paper, we propose several types of features based on token statistics, surface-patterns, and specialized dictionaries and apply them within a sequence modeling framework to learn a fullname parser. In particular, we propose the use of "bucket" features based on (name-token, label) distributions in lieu of "term" features frequently used in various Natural Language Processing applications to prevent the growth of learning parameters as a function of the training data size. We experimentally illustrate the generalizability, effectiveness, and efficiency aspects of our proposed features for noisy fullname parsing on fullname strings from the popular, professional networking website LinkedIn and commonly-used person names in the United States. On these datasets, our fullname parser significantly outperforms both the parser trained using classification approaches and a commercially-available name parsing solution.