Today, everybody in the localization industry is talking about machine translation. The most heated topics in this debate are the ethics and efficiency of technological solutions. We decided to tell you about our experience in implementing machine translation. Spoiler: no machine revolt on the horizon… yet.
Prerequisites
One day, we realized that most of our competitors started using machine translation and writing about it. This coincided with the disruptive arrival of ChatGPT and led to many requests from our customers to use machine translation. Well, our clients asked for it, so we looked at how we could implement artificial intelligence into our daily workflow!
Implementation
The first step was testing various tools. We studied about ten solutions and ranked them according to our needs, namely:
- Translation memory support
- Glossary support
- Engine trainability
- Wide range of supported languages
- Reasonable costs.
We found four of the most suitable tools and took them all for a spin!
A variety of game texts were used for tests, such as dialogs, narrative texts, and UI texts. We chose two test projects with different settings and diverse contents. The robots got to work, and the texts were translated from English into French, German, Russian, and Brazilian Portuguese.
The quality of the results was evaluated by a bunch of linguists that we know and trust and who have proven to have excellent command of their respective language pairs. We made a list of criteria and scored each translation, much in the same way we test new linguists for our team.
The final statistics allowed us to rank the engines. Some turned out to be completely unusable—the quality of the translated text did not satisfy us in any language. Bad robot! Others showed a good result in one language and a completely unusable result in another. To be used in a game, even a good translation might require post-editing, but unusable ones would need to be translated again from scratch.
In the end, the winner was DeepL. It demonstrated adequate quality in all languages and had additional advantages in the use of translation memories and glossaries. At the time of testing, DeepL supported glossaries in seven languages, and since then, four more have been added—some of them at our request! Its another advantage was the ability to integrate with memoQ, the CAT tool of our choice.
Having decided on the engine, we needed to put the machine through its paces and started testing it on real projects. Of course, our clients first gave their consent! We began with small texts, which were evaluated by our linguists. Gradually, the projects became more numerous, although machine translation is still not used everywhere.
Machine translation is indispensable with larger projects and tighter budgets. In many cases, it also helps reduce job completion times, as post-editing is faster than conventional translation. However, since we want to be sure of the quality of the final product, texts have to undergo additional checking. With machine translation, it also takes more time to compile a glossary than it normally does with human translators. In the long run, machine translation speeds up the localization workflow, but for now, we are not bragging about the timeframes—they differ only slightly.
Performance Review
Long-standing localization projects accumulate extensive translation memories over time. A linguist may miss something or use a different sentence structure than in previous batches—for example, they may use an infinitive instead of an imperative. DeepL helps in this respect, as it memorizes sentence structures and reduces the risk of inconsistencies.
Why Artificial Intelligence Won’t Overtake Natural Intelligence
Pre-translation analysis plays a significant role in machine translation. It helps to assess whether the text is even suitable for processing by a machine. It often happens that one portion of a text can be machine translated and post-edited, while another—perhaps a more creative one—requires a human touch throughout and is given to a human translator.
Many texts are not suitable for machine translation. These include:
- Anything related to poetry: A robot can write poetry, but it’s unlikely to put the right meaning into it.
- Riddles: They require imagination, which machines do not have.
- Marketing and advertising copies: They must be bright and attractive, and human reaction is too important to leave it to machines.
- Texts with a lot of complex tags: A translation engine can understand simple tags, but more complex ones are almost guaranteed to break its silicon brain!
- Any texts with random generation: The situation in this case is the same as with tags—you need to take care of too many variables and add a drop of creativity.
- Humor: As we already mentioned in our last article about AI, robot jokes are never funny!
The Verdict
Machine translation is a great tool that can speed up the localization workflow and reduce costs. However, a complete switch to machine translation is not feasible in the near future. First, texts need to be reviewed—both before and after translation—and secondly, there are many types of texts that a machine simply cannot handle.
So fear not. Artificial intelligence can perform routine tasks, but we still need humans to manage it and handle anything that requires creativity!
