Gender and Animacy Knowledge Discovery from Web-Scale N-Grams for Unsupervised Person Mention Detection

Gender and Animacy Knowledge Discovery from Web-Scale N-Grams for Unsupervised Person Mention Detection* Heng Ji a and Dekang Lin b aComputer Science Department, Queens College and Graduate Center, City University of New York 65-30 Kissena Boulevard, Flushing, NY 11367, USA hengji@cs.qc.cuny.edu bGoogle, Inc. 1600 Amphitheater Parkway, Mountain View, CA 94043, USA lindek@google.com Abstract. In this paper we present a simple approach to discover gender and animacy knowledge for person mention detection. We learn noun-gender and noun-animacy pair counts from web-scale n-grams using specific lexical patterns, and then apply confidence estimation metrics to filter noise. The selected informative pairs are then used to detect person mentions from raw texts in an unsupervised learning framework. Experiments showed that this approach can achieve high performance comparable to state-of-the-art supervised learning methods which require manually annotated corpora and gazetteers. Keywords: Knowledge Discovery, Mention Detection, N-Grams, Gender, Animacy 1 Introduction The task of detecting entity mentions (references to entities) is very important to the downstream processing of information extraction such as coreference resolution and event extraction. Entity mentions can be divided into name mentions (e.g. “John Smith”), nominal mentions (e.g. “president”) and pronouns (e.g. “he”, “she”). Typical mention detection systems are based on supervised learning (Boschee et al., 2005; Zitouni and Florian, 2008) or semi- supervised learning (Ji and Grishman, 2006). Achieving really high performance for mention detection requires deep semantic knowledge and large costly hand-labeled data. Many systems also exploited lexical gazetteers such as census data with gender information. However, such knowledge is relatively static (it is not updated during the extraction process), expensive to construct, and doesn’t include any probabilistic information. Mention detection is by definition a semantic task: for example, a phrase is a person mention if it refers to a real-world person entity. We should thus expect a successful mention detection system to exploit world knowledge, in order to resolve hard cases. For example, if a reflexive pronoun (e.g. “himself”) is bound by a phrase in its governing category (Haegeman, 1994), then this phrase is likely to be a person mention (masculine or feminine). In addition, a person mention usually has a life and therefore is likely to be animate (Cobuild, 1995). Therefore, if we could automatically discover a large knowledge base of gender and animacy properties for all possible noun phrases, it will be a valuable resource for person mention detection. * This work was supported by National Science Foundation and Google, Inc. through the Johns Hopkins University 2009 Summer Workshop on Unsupervised Acquisition of Lexical Knowledge from N-Grams. The first author’s research was also sponsored by CUNY Research Enhancement Program and GRTI Program, and the U.S. Army Research Laboratory under Cooperative Agreement Number W911NF-09-2-0053. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the U.S. Government. Copyright 2009 by Heng Ji and Dekang Lin 220 23rd Pacific Asia Conference on Language, Information and Computation, pages …

Content truncated to 500 words