Machine translation (MT) has become a common aspect of many people’s daily lives. Some websites offer automatically translated content. Thousands of videos are posted each day, and video platforms auto-extract subtitles and translate them on demand. If you are well familiar with the process of localization, you are aware of MT but weary of its shortcomings. Here, we will suggest safe and useful ways to include MT in your localization projects.
Quality of machine translation
Machine translation engines have become astonishingly good. Consumers of multilingual content find themselves querying an online translation tool rather than reaching for the good old dictionary. Speech-enabled mobile or home devices allow us to ask for quick translations without typing anything. However, examples of shoddy translations abound, and “this looks like machine translation” is one of the most common complaints from translation buyers.
There is also no denying that machine translation engines rarely reach the quality of professional human translators. When quality translations are the desired result, it is thus vital to at least involve human editors. Furthermore, all translation engines show varying strengths with different language pairs. Thus, several engines may come into play in multilingual projects and you are well-advised to spend some time selecting the best engine for each language pair.
Using machine translation in your translation projects
You may be familiar with the concept of translation memories – databases of source language sentences paired with the corresponding expressions in the target language. When translators work on your content in a translation management system, such as Phrase, they get existing translations offered for every source phrase found in the database. This saves time and effort for the translator and increases consistency.
As soon as the translator encounters new text, machine translation can be a useful productivity tool. The translator would start with a translation provided by an MT engine via Phrase’s pre-translation. If only minor edits are required, this can be more efficient than translating from scratch. Notice the hedges “if” and “can” – in many instances, translators will insist that they can translate faster than they can edit questionable output from machine translation engines. Thus, if you want to try using MT for actual translations, involve the translators in test runs and let them choose the engine that produces the most promising results for them. And then – be sure to leave further decisions on actual translation issues up to them.
History of machine translation systems
The earliest machine translation engines were so-called rule-based machine translation (RBMT) systems. These used hand-crafted language patterns (linguistic “rules”) with carefully compiled dictionaries (“lexicons”) to convert source-language sentences into target-language equivalents. The output quality was underwhelming overall because any grammatical construction that did not neatly fit a rule in the system would predictably lead to failure. And since languages are ever-evolving, they are a moving target for makers of RBMTs who need to define an explicit language model to succeed.
Statistical machine translation (SMT) systems aimed to overcome these difficulties by analyzing large compilations of bilingual text (“corpora”) and automatically extracting correspondences. Instead of using linguistically informed algorithms to generate translations, SMTs gave their best guess on snippets of text.
Since pure statistical approaches often resulted in awkward or ungrammatical formulations, machine translation vendors created hybrid machine translation (HMTs) engines. These used linguistic rules to form grammatical output sentences from the pieces supplied as best guesses from the statistical module. While such hybrid systems may have improved on the quality somewhat, they inherited the cost from both underlying types of systems: large corpora for the statistical component plus sophisticated linguistic models for the rule-based one.
The best translation systems today are based on neural network technology (a.k.a. “artificial intelligence“). These neural machine translation (NMT) systems maintain the benefit of statistical systems because neural networks find statistical regularities in text corpora (the “training data”). But instead of relying on manually created language rules, these systems also use neural networks to discover grammatical patterns. This means that new neural machine translation engines can be created relatively quickly and cheaply – you compile a useful collection of texts (NMT engines often do not even need bilingual corpora) and “train” the networks.
Machine translation engines in Phrase
Phrase recognizes the potential for time and cost savings through machine translations without forgetting about the pitfalls. Therefore, it offers multiple engines that you can select in your Account Settings. The best-of-breed machine translation services available in Phrase are:
- Amazon
- DeepL
- Microsoft
When you test these engines, you may find that DeepL outperforms the others, but it is only available for a relatively limited number of languages. And you may determine that Google is best for German, Microsoft is strongest for Arabic, while Amazon can handle Portuguese best. Therefore, you have the opportunity to select a specific engine for each language in Phrase, unless you are happy to stick with the default Microsoft Translate.
Apart from helping out the translator, machine translation also has other uses in Phrase. With pre-translation, you can generate quick automatic translations for your whole translation project. Thus, you can immediately get a very good idea of what your content may look like when it is translated. Even if a human translator will later provide some modifications, you will be able to assess beforehand whether
- all needed text is properly exposed to translation,
- translations are likely to fit in your layout,
- your software still functions as expected.
Pseudotranslation is sometimes used for this purpose, where text is replaced with foreign-looking gobbledigook. At Phrase, we believe that it is more useful to work with realistic text.
If translators prefer to look at sentences without pre-translation, the feature can be easily turned off. In that case, translators can still benefit from machine translation by using autocomplete – machine translation on demand. With a quick glance, the translator can here decide to use or discard a machine-translated item.
Is machine translation a security problem?
On the web, you can find many articles warning of confidentiality issues posed by using online translation tools. Such problems may well exist with freely available web translators. However, Phrase uses cloud-based engines that have security built-in – encrypted data transfer and data separation per customer – to make sure that none of your data would leak out into the web.
Does machine translation make humans superfluous?
Clearly, no. The best machine translation engines may come close to human translators under ideal circumstances, but even such tools cannot possibly match a translator who thinks along. A good translator does not only provide fluent output but can also determine whether a piece of text makes sense in its context. In a functioning translation team, you may receive queries from translators that prompt you to rewrite your English source text. When a translator has trouble understanding what your text says, it could be poorly written or altogether wrong. No machine translation engine can give you feedback on style, logic, or appropriateness.
Efficient translation workflows with machine translation
In Phrase, your localization project could go like this:
- You internationalize your software,
- You push text resources to Phrase,
- You receive machine translations from Phrase’s pre-translation at minimal to no cost without delay,
- You evaluate the effects of translating on your software (go back to step two if needed),
- Humans translate with machine translation as a productivity aid,
- You receive human quality translations.
Thus, machine translation gives you valuable information before a translator touches the content. And while human translators work on your text, machine translation speeds up the process with suggestions. Thus, machine translation delivers crucial efficiencies that help you stay within your budget and meet your deadlines. All without foregoing the quality that you need to serve your foreign target markets.
