Japanese controlled language rules to improve machine translatability of municipal documents

Japanese Controlled Language Rules to Improve Machine Translatability of Municipal Documents Rei Miyata rei@p.u-tokyo.ac.jp The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-0033, Japan Anthony Hartley a.hartley@rikkyo.ac.jp Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima-ku, Tokyo, 171-8501, Japan C´ecile Paris Cecile.Paris@csiro.au CSIRO, Crn Vimiera and Pembroke Roads, Marsfield NSW, 2122, Australia Midori Tatsumi tatsumi@rikkyo.ac.jp Rikkyo University, 3-34-1 Nishi-Ikebukuro, Toshima-ku, Tokyo, 171-8501, Japan Kyo Kageura kyo@p.u-tokyo.ac.jp The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-0033, Japan Abstract We report on experiments to test the effectiveness of controlled language (CL) rules on texts from Japanese municipal websites. We compiled a set of rules by trial and error, systematically rewriting Japanese source texts and analysing the machine translation (MT) outputs. We then employed native English speakers with little knowledge of Japanese as human evaluators and tested the understandability and accuracy of the English machine translated text. Comparing the results of four MT systems showed that the effectiveness of CL rules varies depending on the particular MT systems. A preliminary selection of optimal rules for each system showed more than 15% increase in MT performance. We also assessed the readability of the Japanese source texts and discuss the way to make them compatible with the quality of the MT outputs. 1 Introduction It has become increasingly important for Japanese municipalities to provide information in mul- tiple languages. Although combining machine translation (MT) and human post-editing has increasingly proven to be an effective MT workflow, it may not be an option for the websites of municipalities, which must offer a large volume of information, such as daily announce- ments, procedural guidelines, emergency information and policy white papers, with the need for frequent updates. Having all these documents translated into a number of languages using human translators or post-editors is simply too expensive and time consuming. In such cases, raw machine translation output is often provided on-demand. Translations between languages that have completely different structures such as Japanese and English are, however, still difficult compared to those between European languages or be- tween Japanese and Korean, and in many cases the quality of raw outputs of MT is of low or even not understandable quality. When post-editing is not a standard option, and the machine translation system is used as a ‘black box’, controlled authoring remains the key to the improve- ment of translation quality. Controlled authoring involves a suite of technologies and environ- ments that provide document templates, glossary management, grammar and style checkers. In the current study, we focus on controlled language (CL) rules as a starting point. We chose En- Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 90 glish as the target language since English is still an overwhelming choice in translating Japanese municipal texts, followed by Chinese, Korean and Portuguese (Carroll, 2010). In our scenario, we assume that effective CL rules should (i) help to raise the quality of MT output, and (ii) not degrade the quality of the Japanese source texts. To achieve …

Content truncated to 500 words