Cross-lingual Sentence Compression for Subtitles

Cross-lingual Sentence Compression for Subtitles Wilker Aziz and Sheila C. M. de Sousa University of Wolverhampton Stafford Street, WV1 1SB Wolverhampton, UK W.Aziz@wlv.ac.uk sheilacastilhoms@gmail.com Lucia Specia Department of Computer Science University of Sheffield 211 Portobello, S1 4DP Sheffield, UK L.Specia@sheffield.ac.uk Abstract We present an approach for translating subtitles where standard time and space constraints are modeled as part of the gen- eration of translations in a phrase-based statistical machine translation system (PB- SMT). We propose and experiment with two promising strategies for jointly trans- lating and compressing subtitles from En- glish into Portuguese. The quality of the automatic translations is measured via the human post-editing of such translations so that they become adequate, fluent and compliant with time and space constraints. Experiments show that carefully selecting the data to tune the model parameters in the PB-SMT system already improves over an unconstrained baseline and that adding specific model components to guide the translation process can further improve the final translations under certain conditions. 1 Introduction The increasing demand for fast and cheap genera- tion of audiovisual content is pushing research and development in the automatic translation of sub- titles. Several attempts have been made in recent years to translate subtitles automatically by using different Machine Translation (MT) approaches (see Section 2). Overall, it has been shown that translation tools can be very helpful in producing adequate and fluent translations of subtitles, yield- ing significant time (and cost) reductions when compared to manually translating subtitles. How- ever, subtitling has other important constraints in addition to translation quality: translations must fit c⃝ 2012 European Association for Machine Translation. the space available on the screen and time slot so that they can be read by viewers. None of the ex- isting approaches to translating subtitles considers these constraints. When generating or translating subtitles from audio transcripts, human subtitlers should follow several conventions. Especially due to the advent of the DVD and the increasing use of smaller and smaller screens, norms and conventions in subti- tling evolve quickly (Cintas and Remael, 2007). Currently, a norm of 40 characters per line, with two lines per screen, seems to be the most widely accepted for television screen, with common vari- ants reaching up to 50 characters per line. Regard- ing time, a subtitle should remain in the screen for at least 1 second and at most 6 seconds if it con- tains two full lines. It is important to make a distinction between translating directly from an audio transcript and translating from a subtitle in the source language. An audio transcript is likely to breach the time/s- pace constraints simply because of the differences between human listening and reading rates. There- fore, some compression is usually necessary when generating monolingual subtitles. Producing sub- titles in a second language however may require a second level of compression: even if the source language subtitle observes the time/space con- straints, depending on the language-pair, a trans- lation can be considerably longer than the source subtitle. This is particularly the …

Content truncated to 500 words