Readability and Translatability Judgments for “Controlled Japanese”

Readability and Translatability Judgments for ‘Controlled Japanese’ Anthony Hartley Toyohashi University of Technology a.hartley@imc.tut.ac.jp Midori Tatsumi Toyohashi University of Technology midori.tatsumi2@mail.dcu.ie Hitoshi Isahara Toyohashi U of Technology isahara@tut.jp Kyo Kageura University of Tokyo kyo@p.u-tokyo.ac.jp Rei Miyata University of Tokyo rei@p.u-tokyo.ac.jp Abstract We report on an experiment to test the ef- ficacy of ‘controlled language’ authoring of technical documents in Japanese, with respect both to the readability of the Jap- anese source and the quality of the Eng- lish machine-translated output. Using four MT systems, we tested two sets of writing rules designed for two document types written by authors with contrasting professional profiles. We elicited judg- ments from native speakers to establish the positive or negative impact of each rule on readability and translation quality. 1 Introduction It is widely acknowledged that the typological ‘distance’ between Japanese and English (the most common European target language for MT from Japanese) hampers the achievement of high-quality translation. We seek to address this challenge by investigating the feasibility of de- veloping a ‘controlled Japanese’ with explicit restrictions on vocabulary, syntax and style ade- quate for authoring technical documentation. Our starting point is sentences extracted from two types of document: consumer user manuals (UM) and company-internal documents articulat- ing the know-how of key employees (KH). UM are produced by professional technical authors, while KH are written as ‘one-offs’ by the em- ployees themselves, capturing their own know- how. Thus, there is a sharp difference in the ef- © 2012 European Association for Machine Translation. fort the two groups of writers can be expected to invest and the linguistic knowledge they bring to a controlled authoring task. In outline, our experiment entailed formulat- ing a set of writing rules (‘authoring guidelines’) for each document type. Sentences violating the rules were extracted from the original data and rewritten (‘pre-edited’ in this experimental set- ting) in accordance with the respective rule. The original and rewritten sentences were then trans- lated by different MT systems; finally, the inputs and outputs were submitted to human evaluation. Since the readers of the original Japanese and the readers of the translated English are equally important, we devised protocols to assess what we termed the ‘readability’ of the Japanese source sentences and their ‘translatability’ as gauged by the perceived quality of the English target sentences. In interpreting the results, we try to identify the most promising avenues for further develop- ment. 2 Controlled Language and MT The general principles of controlled language (CL) and the challenges posed by its deployment are clearly summarised by (Kittredge, 2003; Nyberg et al., 2003). Evidence of the effective- ness of CL in cutting translation costs has been in the public domain for some 30 years, from (Pym, 1990) in the automotive domain to (Roturier, 2009) in the software domain. More specific studies have been undertaken to identify those rules which have the greatest im- pact on the usability of MT output (e.g., O’Brien and Roturier, 2007). Proceedings of the 16th EAMT Conference, 28-30 May 2012, Trento, …

Content truncated to 500 words