A Parallel Corpus of Arabic-Japanese News Articles

A Parallel Corpus of Arabic–Japanese News Articles Go Inoue,1,2 Nizar Habash,2 Yuji Matsumoto,1 Hiroyuki Aoyama3 1Computational Linguistics Laboratory, Nara Institute of Science and Technology, Japan 2Computational Approaches to Modeling Language Lab, New York University Abu Dhabi, UAE 3Graduate School of Global Studies, Tokyo University of Foreign Studies, Japan {inoue.go.ib4, matsu}@is.naist.jp, nizar.habash@nyu.edu, aljabal@tufs.ac.jp Abstract Much work has been done on machine translation between major language pairs including Arabic–English and English–Japanese thanks to the availability of large-scale parallel corpora with manually verified subsets of parallel sentences. However, there has been little research conducted on the Arabic–Japanese language pair due to its parallel-data scarcity, despite being a good example of interestingly contrasting differences in typology. In this paper, we describe the creation process and statistics of the Arabic–Japanese portion of the TUFS Media Corpus, a parallel corpus of translated news articles collected at Tokyo University of Foreign Studies (TUFS). Part of the corpus is manually aligned at the sentence level for development and testing. The corpus is provided in two formats: A document-level parallel corpus in XML format, and a sentence-level parallel corpus in plain text format. We also report the first results of Arabic– Japanese phrase-based machine translation trained on our corpus. Keywords: Arabic, Japanese, Parallel Corpus, Sentence Alignment, Machine Translation 1. Introduction Machine translation (MT) has been a very active research area in natural language processing. Whether its paradigm is statistical or neural, the availability of parallel data is essential for building high-quality systems. In particular, manually verified data sets for development and testing are of great importance for improving and evaluating MT systems. Much work has been done on MT between ma- jor language pairs including Arabic–English and Japanese– English, thanks to the availability of large-scale parallel corpora across various domains with manually aligned sub- sets. However, there has been little research conducted on the Arabic–Japanese language pair due to its parallel-data scarcity, despite being a good example of interestingly con- trasting differences in typology. For instance, Arabic is a verb-initial language, while Japanese is a verb-final lan- guage, where the position of verb is completely opposite as shown in Figure 1. An Arabic token can be highly ambigu- ous in morphological, syntactical, and lexical levels due the absence of optional diacritics for short vowels and conso- nant doubling. In addition, Arabic has a complex system of derivation, inflection, and cliticization. In contrast, a Japanese token can be highly ambiguous due to the absence of spaces between tokens. For more details in linguistic is- sues, see Habash (2010) for Arabic and Bond and Baldwin (2016) for Japanese. In this paper, we present a parallel corpus of Arabic– Japanese news articles, part of which is manually aligned at the sentence level for tuning and evaluation. We also pro- vide the first results of Arabic–Japanese phrase-based MT trained on our corpus. The corpus represents an ongoing project carried out at Tokyo University of Foreign Studies (TUFS) entitled TUFS Media Project,1 which produces translated news ar- ticles in eight languages (Arabic, Bengali, Burmese, …

Content truncated to 500 words