Entity Clustering Across Languages

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 60–69, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Entity Clustering Across Languages Spence Green*, Nicholas Andrews†, Matthew R. Gormley†, Mark Dredze†, and Christopher D. Manning* *Computer Science Department, Stanford University {spenceg,manning}@stanford.edu †Human Language Technology Center of Excellence, Johns Hopkins University {noa,mrg,mdredze}@cs.jhu.edu Abstract Standard entity clustering systems commonly rely on mention (string) matching, syntactic features, and linguistic resources like English WordNet. When co-referent text mentions ap- pear in different languages, these techniques cannot be easily applied. Consequently, we develop new methods for clustering text men- tions across documents and languages simulta- neously, producing cross-lingual entity clusters. Our approach extends standard clustering algo- rithms with cross-lingual mention and context similarity measures. Crucially, we do not as- sume a pre-existing entity list (knowledge base), so entity characteristics are unknown. On an Arabic-English corpus that contains seven dif- ferent text genres, our best model yields a 24.3% F1 gain over the baseline. 1 Introduction This paper introduces techniques for clustering co- referent text mentions across documents and lan- guages. On the web today, a breaking news item may instantly result in mentions to a real-world entity in multiple text formats: news articles, blog posts, tweets, etc. Much NLP work has focused on model adaptation to these diverse text genres. However, the diversity of languages in which the mentions appear is a more significant challenge. This was particularly evident during the 2011 popular uprisings in the Arab world, in which electronic media played a prominent role. A key issue for the outside world was the aggre- gation of information that appeared simultaneously in English, French, and various Arabic dialects. To our knowledge, we are the first to consider clus- tering entity mentions across languages without a pri- ori knowledge of the quantity or types of real-world entities (a knowledge base). The cross-lingual set- ting introduces several challenges. First, we cannot assume a prototypical name format. For example, the Anglo-centric first/middle/last prototype used in previous name modeling work (cf. (Charniak, 2001)) does not apply to Arabic names like Abdullah ibn Abd Al-Aziz Al-Saud or Chinese names like Hu Jin- tao (referred to as Mr. Hu, not Mr. Jintao). Sec- ond, organization names often require both translit- eration and translation. For example, the Arabic �������� ������ ������� ‘General Motors Corp’ contains transliterations of �������� ������ ‘General Motors’, but a translation of ������� ‘Corporation’. Our models are organized as a pipeline. First, for each document, we perform standard mention detec- tion and coreference resolution. Then, we use pair- wise cross-lingual similarity models to measure both mention and context similarity. Finally, we cluster the mentions based on similarity. Our work makes the following contributions: (1) introduction of the task, (2) novel models for cross- lingual entity clustering of person and organization en- tities, (3) cross-lingual annotation of the NIST Auto- matic Content Extraction (ACE) 2008 Arabic-English evaluation set, and (4) experimental results using both gold and automatic within-document processing. …

Content truncated to 500 words