Cross-Lingual Named Entity Recognition via Wikification

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 219–228, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Cross-Lingual Named Entity Recognition via Wikification Chen-Tse Tsai, Stephen Mayhew, and Dan Roth University of Illinois, Urbana-Champaign 201 N. Goodwin, Urbana, Illinois, 61801 {ctsai12, mayhew2, danr}@illinois.edu Abstract Named Entity Recognition (NER) mod- els for language L are typically trained using annotated data in that language. We study cross-lingual NER, where a model for NER in L is trained on an- other, source, language (or multiple source languages). We introduce a language independent method for NER, building on cross-lingual wikification, a technique that grounds words and phrases in non- English text into English Wikipedia en- tries. Thus, mentions in any language can be described using a set of cate- gories and FreeBase types, yielding, as we show, strong language-independent fea- tures. With this insight, we propose an NER model that can be applied to all lan- guages in Wikipedia. When trained on English, our model outperforms compa- rable approaches on the standard CoNLL datasets (Spanish, German, and Dutch) and also performs very well on low- resource languages (e.g., Turkish, Taga- log, Yoruba, Bengali, and Tamil) that have significantly smaller Wikipedia. More- over, our method allows us to train on mul- tiple source languages, typically improv- ing NER results on the target languages. Finally, we show that our language- independent features can be used also to enhance monolingual NER systems, yield- ing improved results for all 9 languages. 1 Introduction Named Entity Recognition (NER) is the task of identifying and typing phrases that contain the names of persons, organizations, locations, and so on. It is an information extraction task that is im- portant for understanding large bodies of text and is considered an essential pre-processing stage in Natural Language Processing (NLP) and Informa- tion Retrieval systems. NER is successful for languages which have a large amount of annotated data, but for languages with little to no annotated data, this task becomes very challenging. There are two common ap- proaches to address the lack of training data prob- lem. The first approach is to automatically gener- ate annotated training data in the target language from Wikipedia articles or from parallel corpora. The performance of this method depends on the quality of the generated data and how well the language-specific features are explored. The sec- ond approach is to train a model on another lan- guage which has abundant training data, and then apply the model directly on test documents in the target language. This direct transfer technique re- lies on developing language-independent features. Note that these two approaches are orthogonal and can be used together. In this paper, we focus on the second, direct transfer setting. We propose a cross-lingual NER model which is trained on annotated documents in one or multiple source languages, and can be applied to all languages in Wikipedia. The model depends on a cross-lingual wikifier, which only requires multilingual Wikipedia, no sentence- aligned or …

Content truncated to 500 words