Only a few years ago using machine translation for subtitles, especially in the entertainment area, seemed completely unrealistic. The very nature of subtitles and oral speech seemed to make it impossible: sentences split into subtitle lines, interjections, character count and reading speed limitations are only some of the challenges that a typical MT engine wouldn't be able to handle, and in the best-case scenario produce an unusable result. 

Today, neural MT changed the way we think about it, and many companies, including very famous streaming platforms, are successfully applying MT for subtitles. But are all these problems solved?

In this talk we will present a study we carried out with the aim to identify the current limitations of NMT engines when applied to subtitles. Using content from several different TV shows we analyzed the types of errors produced by NMT engines. Our goal was to understand the level of quality produced by state-of-the-art MT and whether it is actually helps increase translation efficiency. In addition, this analysis helped us identify areas on which we need to focus our improvement efforts. Finally, by covering three different language combinations we could reflect on which languages pose more challenges in this context. 

This presentation will be of interest to anyone who works in the field of media translations or is interested in MT and its current limitations. The attendees will gain insight into how much MT can really drive efficiency gains in media translation workflows, main factors it depends on, such as language pair, type of content, among others, and find out what is needed to make such setup successful.