Creating an Extended Named Entity Dictionary from Wikipedia

Proceedings of COLING 2012: Technical Papers, pages 1163–1178, COLING 2012, Mumbai, December 2012. Creating an Extended Named Entity Dictionary from Wikipedia Ryuichiro Higashinaka Kugatsu Sadamitsu Kuniko Saito Toshiro Makino Yoshihiro Matsuo NTT Media Intelligence Laboratories, NTT Corporation 1-1 Hikarinooka Yokosuka-Shi Kanagawa 239-0847 Japan {higashinaka.ryuichiro, sadamistu.kugatsu, saito.kuniko, makino.toshiro, matsuo.yoshihiro}@lab.ntt.co.jp ABSTRACT Automatic methods to create entity dictionaries or gazetteers have used only a small number of entity types (18 at maximum), which could pose a limitation for fine-grained information extraction. This paper aims to create a dictionary of 200 extended named entity (ENE) types. Using Wikipedia as a basic resource, we classify Wikipedia titles into ENE types to create an ENE dictionary. In our method, we derive a large number of features for Wikipedia titles and train a multiclass classifier by supervised learning. We devise an extensive list of features for the accurate classification into the ENE types, such as those related to the surface string of a title, the content of the article, and the meta data provided with Wikipedia. By experiments, we successfully show that it is possible to classify Wikipedia titles into ENE types with 79.63% accuracy. We applied our classifier to all Wikipedia titles and, by discarding low-confidence classification results, created an ENE dictionary of over one million entities covering 182 ENE types with an estimated accuracy of 89.48%. This is the first large scale ENE dictionary. TITLE AND ABSTRACT IN ANOTHER LANGUAGE (JAPANESE) Wikipedia を用いた拡張固有表現辞書の構築 従来の固有表現辞書では，少ない数（最大で 18）の固有表現タイプが用いられてきたため， ピンポイントな情報抽出に適用することが難しいという問題があった．そこで，本稿では， 200 の拡張固有表現タイプを用いた固有表現辞書の構築を目指す．具体的には，教師あり学 習による多クラス分類器を用い，Wikipedia の見出し語を拡張固有表現タイプに分類するこ とで辞書を構築する．特徴量として，見出し語そのもの，本文，そして，カテゴリ等のメタ データに関するものを数多く列挙し用いた．結果として，見出し語を，79.63%の精度で，拡 張固有表現タイプに分類できることが分かった．学習された多クラス分類器を，Wikipedia の すべての見出し語に適用し，また，信頼度の低い分類結果については除外するようにしたと ころ，推定分類精度が 89.48%で，また，182 の拡張固有表現タイプをカバーする，百万以 上のエントリを持つ拡張固有表現辞書を構築することができた．この辞書は，初の大規模な 拡張固有表現辞書である． KEYWORDS: Extended Named Entity, Dictionary, Wikipedia. KEYWORDS IN JAPANESE: 拡張固有表現，辞書，Wikipedia. 1163 1 Introduction For information extraction, it is important to recognize named entities (NEs) in texts. NEs are typically recognized by such techniques as support vector machines (SVMs) (Isozaki and Kazawa, 2002) and conditional random fields (Suzuki et al., 2006), using words surrounding a target entity as cues to determine if that entity belongs to a certain NE type. The limitation is that it is difficult to recognize NEs when there are few contextual cues, such as in search queries and snippets of web search results. In such cases, an NE dictionary, or a gazetteer, is particularly useful. Here, an NE dictionary means a list of entities associated with their NE types (e.g., Tokyo → LOCATION, Barack Obama → PERSON). Such a dictionary is also useful for deriving gazetteer features for training an NE recognizer (Kazama and Torisawa, 2008). A number of studies have focused on automatically creating NE dictionaries; e.g., (Toral and Muoz, 2006; Saleh et al., 2010). Such studies generally use a small number of entity types, mostly adopting those defined in the Message Understanding Conference (MUC) (Grishman and Sundheim, 1996) or Information Retrieval and Extraction Exercise (IREX) (Sekine and Isahara, 2000). To enable more fine-grained information extraction, some attempted to cover more NE types: (Chang et al., 2009), (Watanabe et al., 2007), and (Tkatchenko et al., 2011), using up to 18 NE types. Still, we consider the current granularity of the NE …

Content truncated to 500 words