Splitting Noun Compounds via Monolingual and Bilingual Paraphrasing: A Study on Japanese Katakana Words

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 959–969, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Splitting Noun Compounds via Monolingual and Bilingual Paraphrasing: A Study on Japanese Katakana Words Nobuhiro Kaji Institute of Industrial Science University of Tokyo, Tokyo, Japan kaji@tkl.iis.u-tokyo.ac.jp Masaru Kitsuregawa Institute of Industrial Science University of Tokyo, Tokyo, Japan kitsure@tkl.iis.u-tokyo.ac.jp Abstract Word boundaries within noun compounds are not marked by white spaces in a number of languages, unlike in English, and it is benefi- cial for various NLP applications to split such noun compounds. In the case of Japanese, noun compounds made up of katakana words (i.e., transliterated foreign words) are par- ticularly difficult to split, because katakana words are highly productive and are often out- of-vocabulary. To overcome this difficulty, we propose using monolingual and bilingual paraphrases of katakana noun compounds for identifying word boundaries. Experiments demonstrated that splitting accuracy is sub- stantially improved by extracting such para- phrases from unlabeled textual data, the Web in our case, and then using that information for constructing splitting models. 1 Introduction 1.1 Japanese katakana words and noun compound splitting Borrowing is a major type of word formation in Japanese, and numerous foreign words (proper names or neologisms etc.) are continuously being imported from other languages (Tsujimura, 2006). Most borrowed words in modern Japanese are transliterations1 from English and they are referred to as katakana words because transliterated foreign words are primarily spelled by using katakana char- acters in the Japanese writing system.2 Compound- 1Some researchers use the term transcription rather than transliteration (Breen, 2009). Our terminology is based on stud- ies on machine transliteration (Knight and Graehl, 1998). 2The Japanese writing system has four character types: hi- ragana, katakana, kanji, and Latin alphabet. ing is another type of word formation that is com- mon in Japanese (Tsujimura, 2006). In particu- lar, noun compounds are frequently produced by merging two or more nouns together. These two types of word formation yield a significant amount of katakana noun compounds, making Japanese a highly productive language. In Japanese as well as some European and Asian languages (e.g., German, Dutch and Korean), con- stituent words of compounds are not separated by white spaces, unlike in English. In those languages, it is beneficial for various NLP applications to split such compounds. For example, compound splitting enables SMT systems to translate a compound on a word-by-word basis, even if the compound itself is not found in the translation table (Koehn and Knight, 2003; Dyer, 2009). In the context of IR, decom- pounding has an analogous effect to stemming, and it significantly improves retrieval results (Braschler and Ripplinger, 2004). In abbreviation recognition, the definition of an abbreviation is often in the form of a noun compound, and most abbreviation recogni- tion algorithms assume that the definition is properly segmented; see e.g., (Schwartz and Hearst, 2003; Okazaki et al., 2008). This has led NLP researchers to explore meth- ods for splitting compounds, especially noun com- pounds, …

Content truncated to 500 words