Pre-reordering for Statistical Machine Translation of Non-fictional Subtitles

Pre-reordering for Statistical Machine Translation of Non-fictional Subtitles Magdalena Plamad˘a1 Gion Linder2 1Institute of Computational Linguistics University of Zurich Binzm¨uhlestrasse 14 CH-8050 Zurich {plamada, volk}@cl.uzh.ch phillip.stroebel@uzh.ch Phillip Str¨obel1 Martin Volk1 2SWISS TXT Schweizerische Teletext AG Alexander-Sch¨oni-Strasse 40 CH-2501 Biel gion.linder@swisstxt.ch Abstract This paper describes the challenges of building a Statistical Machine Translation (SMT) system for non-fictional subtitles. Since our experiments focus on a “dif- ficult“ translation direction (i.e. French- German), we investigate several meth- ods to improve the translation perfor- mance. We also compare our in-house SMT systems (including domain adap- tation and pre-reordering techniques) to other SMT services and show that pre- reordering alone significantly improves the baseline systems. 1 Introduction The recent advances in Statistical Machine Trans- lation (SMT) have drawn the interest of the lan- guage industry towards it. The main advantages of integrating automatic translations are both cost and time savings, since the translation efforts can be reduced to post-editing activities. Experiments for different topical domains (such as software lo- calization, film subtitling or automobile marketing texts) reported time savings between 20% and 30% (Volk, 2008; Plitt and Masselot, 2010; L¨aubli et al., 2013). These success stories strengthen our moti- vation to build a SMT system specialized on non- fictional content (e.g. documentaries, informative broadcasts). The challenge of this task lies in the desired translation direction, namely from French into German. As the target language is morpholog- ically richer than the source language, we ex- c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. pect difficulties in generating grammatically cor- rect output. This drawback can be overcome by means of hierarchical models (Huck et al., 2013), improved morphological processing (Cap et al., 2014) or models enriched with part-of-speech (POS) information (St¨uker et al., 2011). Another known issue with translations into German is the word order (e.g. the long-range disposal of separa- ble prefix verbs or composed tenses), which can re- sult in missing verbs or verb particles in the trans- lated output. A general solution when translating between languages with different word order is to reorder the source texts according to the word or- der in the target language, as suggested by Niehues and Kolls (2009). In this paper we investigate how well these tech- niques can be applied for subtitles and we par- ticularly focus on the problem of missing verbs. We show that handling this aspect alone improves the SMT performance. We furthermore discuss whether the SMT performance is good enough to be incorporated in the translation workflow of a subtitling company. 2 The proposed solution 2.1 Domain description SWISS TXT provides multimedia solutions for the Swiss National Radio and Television association. The company includes a subtitling division, which is responsible for producing subtitles for the broad- casted TV shows in the Swiss national languages: German, French, Italian and Rumansh. The sub- titles are localized for the region where the TV show is broadcast (e.g. in the German-speaking part of Switzerland subtitles are …

Content truncated to 500 words