Los efectos de la temperatura se pueden observar en la Figura 5.12 , donde se muestra
5.3.3.5 Medidas del espectro de ganancia en diodos láser.
Different types of MT systems have been developed and tested for translating subtitles. Flanagan (2009) carried out research (as part of the MovRat project), which aimed to
apply MT to AVT and test the feasibility of seeding an example-based machine translation (EBMT) system with human-generated subtitles to machine translate new movie subtitles from English to German and English to Japanese. This study was based on previous research conducted by Armstrong et al. (2006), who focused on improving the quality of automated DVD subtitles by using a modular EBMT system, MaTrEx. Flanagan found that increasing corpus size and the number of source language repetitions while decreasing corpus homogeneity improved the readability of the EBMT- generated subtitles. However, these interventions did not have much impact on comprehensibility, style and well-formedness of the EBMT-generated subtitles. In addition, results showed that the viewing subjects’ judgements of the quality of EBMT- generated subtitles were not strongly affected by their linguistic background, while prior knowledge of the movie had an effect on how they rated the severity of observed errors in the subtitles and the style of subtitles, but this effect is training corpus-dependent.
Castilho et al. (2011) tested rule-based machine translation (RBMT) and statistical machine translation (SMT) systems for translating subtitles. They assessed post-editing for automatic and semi-automatic translation (Brazilian Portuguese and English) of DVD subtitles. She downloaded subtitles of the American TV series X files from three free subtitle websites, cleaned the corpus, used one RBMT system, two SMT systems and one TM (translation memory) system, and employed eleven participants to post-edit the automatic translations and to translate English subtitles from scratch. They found that post-editing was on average 70% faster than translating from scratch. More than 69% of the translation required little or no post-editing with their best system. The TM system also performed well compared to translating from scratch.
The EU-funded project SUMAT,32 which aimed at developing an online SMT service for
subtitles for nine European languages combined into 14 language pairs, also achieved promising results. The MT system was trained using corpora composed of combined professionally-created and crowd-sourced data (Fishel et al., 2012; Etchegoyhen et al., 2014). Etchegoyhen et al. (ibid.) carried out productivity measurement through SUMAT by comparing the time needed to translate a subtitled file from scratch with post-editing machine translated output. Results show that the productivity of post-editing MT output was nearly 40% higher than translating from scratch in terms of subtitle per minute. Meanwhile, their quality evaluation also demonstrated positive results about post- editing MT (see Section 3.3.2.1 for more information). Both Castilho’s study and the SUMAT project have confirmed that MT can be really helpful for translating subtitles, and they are faster to post-edit than to translate from scratch.
In addition to academic research, some scholars have explored the commercial possibilities of machine translated subtitles. Volk (2008) reports on a project that aimed to build an SMT system for translating film subtitles from Swedish to Danish and Norwegian in a commercial setting. The project had access to more than 14,000 subtitle files in each language and used 4 million subtitles for training the system. The MT systems built by Volk and his colleagues have been used in large scale subtitle production by a subtitle company who is satisfied with their work. Volk et al. (2010) report their experiences in working with the consumers of the subtitles and summarized lessons for SMT in subtitle production: a) TM-MT integration is not self-evidently beneficial due to the modest contribution of TM to the translation quality; b) there are
conflicting interests among different stakeholders, for whom quality or time might be more important; c) integrating SMT into the subtitling workflow is the key; d) ambiguous MT output with strange constructions may influence post-editors’ judgements; e) a complex SMT system requires maintenance and updates, but companies are reluctant to invest human resources in it; f) the idea of presenting alternative translations on the screen is rejected by post-editors due to the time cost; g) increasing the efficiency of post-editors and improving the MT system to filter out bad translations is important.
Georgakopoulou (2012) discusses Automatic Speech Recognition (ASR) technology in interlingual subtitling. She argues that there is an increasing demand for interlingual subtitling, and that intralingual subtitles are not only used by hard-of-hearing viewers, but also by hearing viewers as a means to retrieve information. She presents the problems that would prevent subtitling by MT from being implemented in business practice: firstly, a lack of corpora leads to unsatisfactory results, which makes post- editing a must; secondly, more technological progress in speech recognition such as speech-to-text conversion is needed; thirdly, responding to points made by Volk et al. (2010), Georgakopoulou emphasizes that it is difficult to integrate the technology into existing workflows, and for the success of MT systems, it is necessary to establish collaboration and build trust among subtitling companies, language experts and subtitlers. Georgakopoulou points out that the ideal scenario for the subtitling industry will be to create a system that can do automatic transcription of the audio streams and then machine translate the text. She also called for the training of linguists and interdisciplinarity in studies at university level to help people understand and learn subtitling by MT. The three-year project “ALST-Linguistic and Sensorial Accessibility:
Technologies for Voice-over And Audio Description” led by Matamala (2015) is also about the application of ASR and MT in video, although it is not directly related to subtitles, the results of the project also provide insights on integrating MT and ASR into subtitling.
Although there are research projects on subtitling by MT, no established subtitling group has publicly acknowledged their use of MT. On the contrary, the acceptance of MT for audio-visual products is still quite low in society. For example, OpenSubtitles.org is (see the end of Section 2.2.5.1) one of the best and largest subtitle sites. In its forum, there is a highlighted notice33 (last edited in May 2014) saying that “Opensubtitles has decided
to get rid of all ‘MACHINE TRANSLATED’ subtitles”, since this is “best for the future of OS as we can avoid mass trash being accumulated and saved here over the coming years”. The administrators of the forum believe that:
“Viewing a movie with a machine (google) translation, you are truly missing a lot of the real dialogue with these subtitles. Even though you can make basic sense out of them, the wording does not portray the true concept, belief and emotions depicted in a particular scene of a movie/TV series”.
To change this situation, apart from improving the translation quality, trustworthiness should be established among subtitlers, end users, and project managers when MT is deployed.