Machine translation (MT) has become a part of life; Google Translate, for example, boasts more than 200 million users. The quality of machine translation also has improved over the years. Of course, MT can't do everything any more than electronic assistants such as Siri or Alexa can make decisions for you. The keys to leveraging machine translation at your company are knowing when to use MT, when to use a highly skilled human translator, and which type of MT makes the most sense.
History of MT
The idea to use computers to translate is as old as computers themselves and the development of MT follows closely that of computer technology. The first efforts at MT began in the 1940s, however, this effort failed.
Rule-based MT (RbMT)
During the 1950s, MT consisted of rules and dictionary-based machine translation systems. These systems parsed a source sentence to determine its structure and identify words, then converted it into the target language based upon rules and lexicons compiled by linguists. While RbMT has the advantage of being able to translate a large amount of material quickly, developing a good system can require a lot of time and money. Also, rules can't adequately cover all possible situations. If existing rules conflict, the machine will deliver poor translations.
RbMT was followed by the reintroduction of statistical MT in the late 1990s. In statistical MT, massive amounts of mono and bilingual data is fragmented and reassembled for translation. The system learns how to translate by analyzing human translations. Its translations are phrase-based, rather than word-based as in RbMT. It does not have rules programmed into it, but it “learns” these rules by analyzing human translations and using statistical analysis.
SMT is still popular today and was the core of most search engine translations, such as Google Translate. One advantage is that humans don't need to handcraft engines for each language. A downside, however, is that it requires a large bilingual body of text to translate accurately.
Neural MT (NMT) is the newest approach to MT and uses neural networks composed of nodes that model the human brain. The first neural systems became available in 2016. In many cases, NMT is able to deliver more accurate guesses at translations and to continually learn and adjust. NMT requires a lot of computing power and so has been possible only recently. Now many translation engines use NMT. While NMT can translate complex sentence structures, it isn't as good at translating shorter phrases. Another downside is that it requires more data than an SMT engine does to train.
Hybrid approaches use a combination of RbMT, SMT, and NMT depending upon which method will be most efficient and accurate for any given passage. They provide seamless integration among the different engines.
When MT Works Best
Of course, for MT to work it must support the languages you're interested in as well as regional variations, such as European Spanish or Latin American Spanish. MT also works well in some situations but less well in others. The more structured the text is the more likely it is to be a candidate for MT. Also, text that presents a low risk if not translated 100 percent correctly can use MT instead of human translation. For example, these types of content are best suited for MT:
- Structured content
- Long sentences
- Technical language
- Professionally written
This might include blogs and social media, frequently asked questions, after-sales care, and reviews.
Texts requiring professional human translators would include:
- Legal or medical texts
- User interfaces
- Advertising and marketing materials
- Informal content
- Non-structured content
It's important to note that many translation engines and providers are not confidential and using them can run awry of compliance issues, in particular, General Data Protection Regulations. For example, the two major online players, Google Cloud Translate and AWS (Amazon Web Services) send queries across the internet and therefore in the context of corporations translating company and customer information, are not GDPR compliant. If compliance is an issue for your business, check with companies that offer translation engines implemented on-premises, with data protection. Contact Andovar to explore on-premises, secure enterprise machine translation solutions.
The Most Efficient Translations
The best way to leverage MT for many companies is to use a hybrid translation engine. on-premises that is GDPR compliant and to use it in combination with human translators and post-editors. Ideally, a translation management system (TMS) will automatically route the content to a machine or human translator and track each part of the translation project.
We have a variety of smart translation solutions including machine translation engines, traditional tools (including CAT tools, Translation Memory and Termbase), human translators, and a translation management system to provide efficient and accurate translations. Our solutions are fully GDPR compliant. We can help you determine what solutions work best for your needs. Contact us today to learn how we can help you leverage machine translation.