Transforming Wikipedia into Named Entity Training Data

Transforming Wikipedia into Named Entity Training Data Joel Nothman and James R. Curran and Tara Murphy School of Information Technologies University of Sydney NSW 2006, Australia {jnot4610,james,tm}@it.usyd.edu.au Abstract Statistical named entity recognisers require costly hand-labelled training data and, as a result, most existing corpora are small. We exploit Wikipedia to create a massive corpus of named entity annotated text. We transform Wikipedia’s links into named en- tity annotations by classifying the target ar- ticles into common entity types (e.g. per- son, organisation and location). Compar- ing to MUC, CONLL and BBN corpora, Wikipedia generally performs better than other cross-corpus train/test pairs. 1 Introduction Named Entity Recognition (NER), the task of identifying and classifying the names of people, organisations, locations and other entities within text, is central to many NLP tasks. The task devel- oped from information extraction in the Message Understanding Conferences (MUC) of the 1990s. By the final two MUC evaluations, NER had be- come a distinct task: tagging the aforementioned proper names and some temporal and numerical expressions (Chinchor, 1998). The CONLL NER evaluations of 2002 and 2003 (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003) focused on determining superior machine learning algorithms and feature models for multilingual NER, marking tags for person (PER), organisation (ORG), location (LOC) and miscellaneous (MISC; broadly including e.g. events, artworks and nationalities). Brunstein (2002) and Sekine et al. (2002) expanded this into fine-grained categorical hierarchies; others have utilised the WordNet noun hierarchy (Miller, 1998) in a similar manner (e.g. Toral et al. (2008)). For some applications, such as bio- textmining (Kim et al., 2003) or astroinformat- ics (Murphy et al., 2006), domain-specific entity classification schemes are more appropriate. Statistical machine learning systems have proved successful for NER. These learn terms and patterns commonly associated with particu- lar entity classes, making use of many contextual, orthographic, linguistic and external knowledge features. They rely on annotated training corpora of newswire text, each typically smaller than a million words. The need for costly, low-yield, expert annotation therefore hinders the creation of more task-adaptable, high-performance named entity (NE) taggers. This paper presents the use of Wikipedia1— an enormous and growing, multilingual, free resource—to create NE-annotated corpora. We transform links between encyclopaedia articles into named entity annotations (see Figure 1). Each new term or name mentioned in a Wikipedia article is often linked to an appropriate article. A sentence introducing Ian Fleming’s novel Thun- derball about the character James Bond may thus have links to separate articles about each entity. Cues in the linked article about Ian Fleming in- dicate that it is about a person, and the article on Thunderball states that it is a novel. The original sentence can then be automatically annotated with these facts. Millions of sentences may similarly be extracted from Wikipedia to form an enormous corpus for NER training. 1http://www.wikipedia.org Having produced annotated text in this man- ner, it can be used to train an existing NER sys- tem. By training the C&C tagger (Curran and Clark, 2003) …

Content truncated to 500 words