Annotation of ‘Word List by Semantic Principles’ Labels for the Balanced Corpus of Contemporary Written Japanese

Annotation of ‘Word List by Semantic Principles’ Labels for the Balanced Corpus of Contemporary Written Japanese Sachi Kato Masayuki Asahara National Institute for Japanese Language and Linguistics, Japan Tokyo, Japan {yasuda-s, masayu-a, yamazaki} AT ninjal DOT ac DOT jp Makoto Yamazaki Abstract This article presents a word-sense annotation for the Balanced Corpus of Contemporary Written Japanese: a mashed-up Japanese lexi- con based on the ‘Word List by Semantic Prin- ciples’ (WLSP). The WLSP is a large-scale Japanese thesaurus which includes 98,241 en- tries with syntactic and hierarchical seman- tic categories. We utilized a morpheme-word sense alignment table to extract all possible word sense candidates for each word appear- ing in the target corpus. Then, we manually disambiguated the word senses for 182,166 content words in the texts. 1 Introduction Semantic information annotation is an important lin- guistic resource to explore synonyms by semantic category or figurative expressions by discrepancies in the co-occurrence of semantic categories. These annotations can also be used as training and eval- uation data for word-sense disambiguation tasks. Among Japanese language resources, the EDR cor- pus (Yokoi, 1995) and the RWCP corpus (Toyoura et al., 1996) include word-sense information based on the gloss of the dictionaries. The Balanced Corpus of Contemporary Written Japanese (hereafter BC- CWJ) (Maekawa et al., 2014) was compiled using a sampling method that preserved representativeness of actual usage. Word-sense information from the Iwanami dictionary is annotated in a subset of the BCCWJ, and the data are utilized in the SemEval- 2010 Japanese WSD Task (Okumura et al., 2010). Several all-word sense disambiguation methods are proposed in the benchmark data. A thesaurus-based word-sense tagged corpus has also been proposed. (Bond et al., 2012) developed annotation data for Japanese WordNet, which is a translation of English data. In this study, we develop a new semantic infor- mation annotation of BCCWJ. The semantic infor- mation is based on the thesaurus ’Word List by Se- mantic Principles, Revised and Enlarged Version’ (hereafter WLSP) (Kokuritsu Kokugo Kenkyusho, 2004). This article presents the design of the an- notation and basic statistics of the data. We use a subset of the core data of BCCWJ as the target cor- pus for word-sense annotations. The data include books (BCCWJ sample ID: PB), magazines (PM), and newspapers (PN). The BCCWJ has morpheme information annotations such as word boundary and part-of-speech annotations, based on a morphologi- cal analyzer dictionary, UniDic. (Kondo et al., 2018) developed an alignment table between UniDic mor- pheme information and WLSP word-sense labels. We extract all possible word-sense candidates for the text by the alignment table. We manually dis- ambiguate the word senses of all content words by their contextual information. When the word-sense candidates are not appropriate in the context, we newly assign the WLSP word-sense label manually. The data include 347,094 morphemes making up 182,166 content words and 164,928 function words. The rest of the paper is organized as follows. Section 2 presents how we annotated the WLSP word senses on BCCWJ, employing the language resources of …

Content truncated to 500 words