Entity Linking Leveraging Automatically Generated Annotation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1290–1298, Beijing, August 2010 Entity Linking Leveraging Automatically Generated Annotation Wei Zhang† Jian Su‡ Chew Lim Tan† Wen Ting Wang‡ †School of Computing National University of Singapore {z-wei, tancl} @comp.nus.edu.sg ‡ Institute for Infocomm Research {sujian, wwang} @i2r.a-star.edu.sg Abstract Entity linking refers entity mentions in a document to their representations in a knowledge base (KB). In this paper, we propose to use additional information sources from Wikipedia to find more name variations for entity linking task. In addition, as manually creating a training corpus for entity linking is labor- intensive and costly, we present a novel method to automatically generate a large scale corpus annotation for ambiguous mentions leveraging on their unambi- guous synonyms in the document collec- tion. Then, a binary classifier is trained to filter out KB entities that are not simi- lar to current mentions. This classifier not only can effectively reduce the am- biguities to the existing entities in KB, but also be very useful to highlight the new entities to KB for the further popu- lation. Furthermore, we also leverage on the Wikipedia documents to provide ad- ditional information which is not availa- ble in our generated corpus through a domain adaption approach which pro- vides further performance improve- ments. The experiment results show that our proposed method outperforms the state-of-the-art approaches. 1 Introduction The named entity (NE) ambiguation has raised serious problems in many areas, including web people search, knowledge base population (KBP), and information extraction, because an entity (such as Abbott Laboratories, a diversified pharmaceuticals health care company) can be referred to by multiple mentions (e.g. “ABT” and “Abbott”), and a mention (e.g. “Abbott”) can be shared by different entities (e.g. Abbott Texas: a city in United States; Bud Abbott, an American actor; and Abbott Laboratories, a diversified pharmaceutical health care company). Both Web People Search (WePS) task (Artiles et al. 2007) and Global Entity Detection & Rec- ognition task (GEDR) in Automatic Content Ex- traction 2008 (ACE08) disambiguate entity men- tions by clustering documents with these men- tions. Each cluster then represents a unique enti- ty. Recently entity linking has been proposed in this field. However, it is quite different from the previous tasks. Given a knowledge base, a document collec- tion, entity linking task as defined by KBP-091 (McNamee and Dang, 2009) is to determine for each name string and the document it appears, which knowledge base entity is being referred to, or if the entity is a new entity which is not present in the reference KB. Compared with GEDR and WePS, entity link- ing has a given entity list (i.e. the reference KB) to which we disambiguate the entity mentions. Moreover, in document collection, there are new entities which are not present in KB and can be used for further population. In fact, new entities with or without the names in KB cover more than half of testing instances. 1 http://apl.jhu.edu/~paulmac/kbp.html 1290 Entity linking has been explored …

Content truncated to 500 words