Cross-Language Entity Linking

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 255–263, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Cross-Language Entity Linking Paul McNamee and James Mayfield HLTCOE & Applied Physics Laboratory Johns Hopkins University {paul.mcnamee,james.mayfield}@jhuapl.edu Dawn Lawrie Loyola University Maryland lawrie@cs.loyola.edu Douglas W. Oard iSchool & UMIACS University of Maryland, College Park oard@umd.edu David Doermann UMIACS University of Maryland, College Park doermann@umiacs.umd.edu Abstract There has been substantial recent interest in aligning mentions of named entities in unstructured texts to knowledge base de- scriptors, a task commonly called entity linking. This technology is crucial for applications in knowledge discovery and text data mining. This paper presents ex- periments in the new problem of cross- language entity linking, where documents and named entities are in a different lan- guage than that used for the content of the reference knowledge base. We have created a new test collection to evaluate cross-language entity linking performance in twenty-one languages. We present ex- periments that examine issues such as: the importance of transliteration; the utility of cross-language information retrieval; and, the potential benefit of multilingual named entity recognition. Our best model achieves performance which is 94% of a strong monolingual baseline. 1 Introduction Entity Linking involves aligning a textual mention of a named entity to the entry in a knowledge base (KB) that represents the mentioned entity, if it is present. The problem has two main complicating features: entities can be referred to using multi- ple name variants (e.g., aliases or misspellings); and several entities can share the same name (e.g., many people are named Mar´ıa S´anchez). Ap- plications of entity linking include linking pa- tient health records from separate hospitalizations, maintaining personal credit files, preventing iden- tity crimes, and supporting law enforcement. Starting in 2009 the NIST Text Analysis Conference (TAC) began conducting evaluations of technologies for knowledge base population (KBP). Systems addressing the entity linking sub- task take as input a name string from a document and produce as output the knowledge base node, if any, corresponding to the mentioned entity. This capability is vital for knowledge discovery; with- out it, extracted information cannot be properly in- serted in the correct KB node. We follow the TAC- KBP problem formulation and use its reference knowledge base, which was derived from a 2008 snapshot of English Wikipedia. In the present work our focus is on person entities. We seek to develop and evaluate technologies for matching foreign language names to the appropriate knowl- edge base descriptor (or kbid) in the English KB. To support this research we created what we believe to be the first cross-language entity link- ing test collection. Our dataset includes twenty- one languages in addition to English, and covers five writing systems. Compared to the problem of monolingual (English) entity linking, a solution to the cross-language variant requires both a method to match foreign names to their English equiva- lents, and a way to compare contextual features from the non-English source document with con- textual information …

Content truncated to 500 words