Universal Dependencies Version 2 for Japanese

Universal Dependencies Version 2 for Japanese Masayuki Asahara♡†, Hiroshi Kanayama♠, Takaaki Tanaka♣ Yusuke Miyao♢, Sumire Uematsu♢, Shinsuke Mori◦ Yuji Matsumoto•, Mai Omura♡, Yugo Murawaki◦ ♡ National Institute for Japanese Language and Linguistics, ♠ IBM Research AI ♣ NTT Communication Science Labolartories, ♢ National Institute of Informatics ◦ Kyoto University, • Nara Institute of Science and Technology masayu-a@ninjal.ac.jp† Abstract In this paper, we are presenting our work to build Universal Dependencies (UD) resources for Japanese. The UD Japanese resources are built based on automatic conversion from several treebanks. The word delimitation, POS, and syntactic relations of the existing treebanks are ported for the UD annotation scheme. We discuss the issues of the UD scheme found through porting of the Japanese language. 1. Introduction The Universal Dependencies (UD) project (McDonald et al., 2013) has defined a consistent, cross- linguistic target and syntactic structure representation format. In this presentation, we will show the work of the UD Japanese team. The UD Japanese team was organised by interested people who are developing their own treebanks or parsers. We developed and maintained several UD guidelines (version 2.0) compatible data for Japanese. Most of the data are made through automatic conversion from the existing treebank. The UD annotation guidelines were updated from version 1 to version 2 in early 2017. The automatic conversion enabled us to adapt the existing annotation based on traditional Japanese grammar conventions for the UD annotation guidelines changes. In this paper, we discuss the current issues of UD Japanese resources until today. These issues come from the difficulty to perform cross-linguistically consistent annotation for the different grammatical system from western European languages. The points at the issues related to the conversions are split into the delimitation (word, phrase and clause), undefined policies of UD guidelines, typological systems for UD, and copyright of Japanese language resources. 2. UD Japanese Resources We have the following UD Japanese resources: UD Japanese-BCCWJ, UD Japanese-KTC (Tanaka et al., 2016), UD Japanese-GSD, UD Japanese- PUD, and UD Japanese-Modern (Omura et al., 2017). Table 1 presents the current status of UD Japanese resources. Below, we describe these resources briefly. UD Japanese-BCCWJ is UD data based on the ‘Balanced Corpus of Contemporary Written Japanese’ (hereafter BC- CWJ) (Maekawa et al., 2014). The BCCWJ defines 1 million word-scale core data samples in which the mor- phological information is manually annotated with three layers of word delimitations: Short Unit Word (SUW), Long Unit Word (LUW), and bunsetsu. The BCCWJ has several syntactic annotations. The BCCWJ-DepPara (Asahara and Matsumoto, 2016) is a bunsetsu-based syn- tactic dependency and coordinate structure annotation. The BCCWJ-PAS (Ueda et al., 2015) is a predicate-argument relation annotation with the NAIST Text Corpus annotation schema (Iida et al., 2007). We maintain conversion rules based on these annotations. UD Japanese-KTC (Tanaka et al., 2016) is based on the NTT Japanese Phrase Structure Treebank (Tanaka and Nagata, 2013) which contains the same original text as the Kyoto Text Corpus (KTC) (Kurohashi and Nagao, 2003). KTC is a bunsetsu, namely base phrase, based …

Content truncated to 500 words