Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 120–128, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Glen, Glenda or Glendale: Unsupervised and Semi-supervised Learning of English Noun Gender Shane Bergsma Department of Computing Science University of Alberta Edmonton, Alberta Canada, T6G 2E8 bergsma@cs.ualberta.ca Dekang Lin Google, Inc. 1600 Amphitheatre Parkway Mountain View California, 94301 lindek@google.com Randy Goebel Department of Computing Science University of Alberta Edmonton, Alberta Canada, T6G 2E8 goebel@cs.ualberta.ca Abstract English pronouns like he and they reliably re- flect the gender and number of the entities to which they refer. Pronoun resolution systems can use this fact to filter noun candidates that do not agree with the pronoun gender. In- deed, broad-coverage models of noun gender have proved to be the most important source of world knowledge in automatic pronoun res- olution systems. Previous approaches predict gender by count- ing the co-occurrence of nouns with pronouns of each gender class. While this provides use- ful statistics for frequent nouns, many infre- quent nouns cannot be classified using this method. Rather than using co-occurrence in- formation directly, we use it to automatically annotate training examples for a large-scale discriminative gender model. Our model col- lectively classifies all occurrences of a noun in a document using a wide variety of con- textual, morphological, and categorical gender features. By leveraging large volumes of un- labeled data, our full semi-supervised system reduces error by 50% over the existing state- of-the-art in gender classification. 1 Introduction Pronoun resolution is the process of determining which preceding nouns are referred to by a partic- ular pronoun in text. Consider the sentence: (1) Glen told Glenda that she was wrong about Glendale. A pronoun resolution system should determine that the pronoun she refers to the noun Glenda. Pro- noun resolution is challenging because it requires a lot of world knowledge (general knowledge of word types). If she is replaced with the pronoun he in (1), Glen becomes the antecedent. Pronoun resolution systems need the knowledge of noun gender that ad- vises that Glen is usually masculine (and thus re- ferred to by he) while Glenda is feminine. English third-person pronouns are grouped in four gender/number categories: masculine (he, his, him, himself), feminine (she, her, herself), neutral (it, its, itself), and plural (they, their, them, themselves). We broadly refer to these gender and number classes simply as gender. The objective of our work is to correctly assign gender to English noun tokens, in context; to determine which class of pronoun will refer to a given noun. One successful approach to this problem is to build a statistical gender model from a noun’s asso- ciation with pronouns in text. For example, Ge et al. (1998) learn Ford has a 94% chance of being neu- tral, based on its frequent co-occurrence with neu- tral pronouns in text. Such estimates are noisy but useful. Both Ge et al. (1998) and Bergsma and Lin (2006) show that learned gender is the most impor- tant feature in …

Content truncated to 500 words