Finding Alternative Translations in a Large Corpus of Movie Subtitle

Finding Alternative Translations in a Large Corpus of Movie Subtitles J¨org Tiedemann Department of Modern Languages University of Helsinki firstname.lastname at helsinki.fi Abstract OpenSubtitles.org provides a large collection of user contributed subtitles in various languages for movies and TV programs. Subtitle translations are valuable resources for cross-lingual studies and machine translation research. A less explored feature of the collection is the inclusion of alternative translations, which can be very useful for training paraphrase systems or collecting multi-reference test suites for machine translation. However, differences in translation may also be due to misspellings, incomplete or corrupt data files, or wrongly aligned subtitles. This paper reports our efforts in recognising and classifying alternative subtitle translations with language independent techniques. We use time-based alignment with lexical re-synchronisation techniques and BLEU score filters and sort alternative translations into categories using edit distance metrics and heuristic rules. Our approach produces large numbers of sentence-aligned translation alternatives for over 50 languages provided via the OPUS corpus collection. Keywords: parallel corpora, alignment, machine translation 1. Introduction Multilingual parallel corpora are useful resources for many purposes such as machine translation, cross-lingual stud- ies, bilingual lexicography but also monolingual tasks such as word sense disambiguation and discovery (Dyvik, 2002) and the detection of idiomatic expressions (Villada Moir´on and Tiedemann, 2006). In most cases, they contain exactly one version per language, referring to either the source text or its translation. However, there are cases in which one would like to consider alternative translations, for example, when evaluating machine translation using metrics such as BLEU (Papineni et al., 2002). In general it is not satisfac- tory (and reliable) to use these automatic evaluation meth- ods with one reference translation only and scores obtained in this way may be quite misleading as many acceptable translation alternatives are not considered. Furthermore, reference-based parameter tuning is also effected by this limitation. Nevertheless, most research is reported with single-reference test/development sets as multiple transla- tions are difficult to obtain. In this paper, we investigate the use of user-contributed movie subtitles as a source of alternative translations. We base our study on the OpenSubtitles corpus included in OPUS (Tiedemann, 2012), which provides a large number of subtitles in various languages including alternative up- loads for many movies and TV shows. In the following, we first discuss the issues of user-uploaded subtitles that can lead to translation differences and after that we propose our method for filtering the data set that makes it possible to identify various categories of subtitle alternatives. As a side product we also obtain cross-lingual links be- tween all subtitle files via the monolingually aligned sub- title alternatives that make it possible to create truly multi- lingual parallel corpora across all languages, which is not available in the previous version of the corpus. Yet another result is the identification of possible errors (spelling mis- takes, encoding problems, etc.) in the collection. This in- formation will be very useful to clean-up the corpus in fu- ture releases. More details and examples are given below. …

Content truncated to 500 words