Attainable Text-to-Text Machine Translation vs. Translation: Issues Beyond Linguistic Processing

Attainable Text-to-Text Machine Translation vs. Translation: Issues Beyond Linguistic Processing Atsushi Fujita atsushi.fujita@nict.go.jp National Institute of Information and Communications Technology, 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289, Japan Abstract Existing approaches for machine translation (MT) mostly translate a given text in the source language into the target language, without explicitly referring to information indispensable for producing a proper translation. This includes not only information in the textual elements and non-textual modalities in the same document, but also extra-document and non-linguistic infor- mation, such as norms and skopos. To design better translation production workflows, we need to distinguish translation issues that could be resolved by the existing text-to-text approaches from those beyond them. To this end, we conducted an analytic assessment of MT outputs, tak- ing an English-to-Japanese news translation task as a case study. First, examples of translation issues and their revisions were collected by a two-stage post-edit (PE) method: performing a minimal PE to obtain a translation attainable based on the given textual information and further performing a full PE to obtain an acceptable translation referring to any necessary informa- tion. The collected revision examples were then manually analyzed. We revealed the dominant issues and information indispensable for resolving them, such as fine-grained style specifica- tions, terminology, domain-specific knowledge, and reference documents, delineating a clear distinction between translation and the translation that text-to-text MT can ultimately attain. 1 Introduction Translation is not a purely linguistic process (Vermeer, 1992) but also the process of producing a document in the target language that plays the same role (has the same effect) as the given source document written in the source language. When translating a given document, transla- tors refer not only to the textual elements in the document, but also to the role of each textual element (e.g., running text, section title, table element, and caption), other non-linguistic ele- ments (e.g., figures and formulae), and their structure. To produce a translation, we also need some extra-document and non-linguistic information, such as the norms specific to the register of the document and corresponding target sub-language (Toury, 1978), the objective and the intended usages of translation, i.e., skopos (Vermeer, 2004), and various specifications (Melby, 2012) designated by the translation client if any. Despite the requirements a (proper) translation must satisfy, techniques for machine trans- lation (MT) have been developed by regarding the task of translation as text-to-text transfer. Until very recently, most studies have performed a text-to-text MT for each text segment,1 even though a sequence of perfect segment-level text-to-text translations does not necessarily qual- ify as a proper translation. Recent studies on neural MT (NMT) have addressed issues beyond 1In this paper, we use “segment” for the unit of inputs for MT systems rather than “sentence,” because a segment is not necessarily composed of a single sentence, but can often be multiple sentences or non-sentential textual fragments. Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume 1: MT Research Track Page 215 this formulation, exploiting further …

Content truncated to 500 words