One for All: Towards Language Independent Named Entity Linking

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2255–2264, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics One for All: Towards Language Independent Named Entity Linking Avirup Sil and Radu Florian IBM T. J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598 avi@us.ibm.com, raduf@us.ibm.com Abstract Entity linking (EL) is the task of dis- ambiguating mentions in text by associ- ating them with entries in a predefined database of mentions (persons, organiza- tions, etc). Most previous EL research has focused mainly on one language, English, with less attention being paid to other lan- guages, such as Spanish or Chinese. In this paper, we introduce LIEL, a Lan- guage Independent Entity Linking system, which provides an EL framework which, once trained on one language, works re- markably well on a number of different languages without change. LIEL makes a joint global prediction over the entire document, employing a discriminative re- ranking framework with many domain and language-independent feature func- tions. Experiments on numerous bench- mark datasets, show that the proposed sys- tem, once trained on one language, En- glish, outperforms several state-of-the-art systems in English (by 4 points) and the trained model also works very well on Spanish (14 points better than a competi- tor system), demonstrating the viability of the approach. 1 Introduction We live in a golden age of information, where we have access to vast amount of data in various forms: text, video and audio. Being able to ana- lyze this data automatically, usually involves fill- ing a relational database, which, in turn, requires the processing system to be able to identify actors across documents by assigning unique identifiers to them. Entity Linking (EL) is the task of map- ping specific textual mentions of entities in a text document to an entry in a large catalog of entities, often called a knowledge base or KB, and is one of the major tasks in the Knowledge-Base Popula- tion track at the Text Analysis Conference (TAC) (Ji et al., 2014). The task also involves grouping together (clustering) NIL entities which do not have any target referents in the KB. Previous work, pioneered by (Bunescu and Pasca, 2006; Cucerzan, 2007; Sil et al., 2012; Ratinov et al., 2011; Guo et al., 2013), have used Wikipedia as this target catalog of entities be- cause of its wide coverage and its frequent updates made by the community. As with many NLP ap- proaches, most of the previous EL research have focused on English, mainly because it has many NLP resources available, it is the most prevalent language on the web, and the fact that the English Wikipedia is the largest among all the Wikipedia datasets. However, there are plenty of web docu- ments in other languages, such as Spanish (Fahrni et al., 2013; Ji et al., 2014), and Chinese (Cao et al., 2014; Shi et al., 2014), with a large num- ber of speakers, and there is a need to be able to develop EL …

Content truncated to 500 words