Focus

The Ever-Evolving Technology of MT

Evolving from being considered a threat and a hindrance to translators, the latest machine translation technology has become an asset to the industry and the entire business world

Arturs Vasilevskis

Arturs Vasilevskis is the head of machine translation at Tilde, a European language technology and service provider.

arturs-vasilevskis
arturs-vasilevskis

Arturs Vasilevskis

Arturs Vasilevskis is the head of machine translation at Tilde, a European language technology and service provider.

M

achine translation (MT) has come a long way from its humble beginnings in the post-war era to becoming an integral part of the 21st century business environment. So what are the current and future trends as we head into 2021? And how is dynamic learning — adaptive neural machine translation— transforming the way professional translators work? We have asked these questions to professional translators to find out how MT has developed over the years and where to apply it best.

Currently, MT is an integral part of all our lives. It is central to future plans of the world’s biggest and most powerful corporations and nations, and will play a significant role in helping to shape our multilingual world of more than 7,000 languages.

A.D. Booth, an Englishman, and the American Warren Weaver, both of whom worked for the Rockefeller Foundation, conceptualized the idea to use computers to translate from various languages in 1946. Several research teams began working on this problem, and in 1954 the first MT system was demonstrated in Georgetown. In the following decades, MT continued to develop, but was hindered by computing limitations. It was perhaps thanks to the perceived success of applying MT for military purposes during the Vietnam War that large-scale MT projects and research budgets continued well into the 1970s. With major advancements in computational power in the 1980s and the advancement of the internet in the 1990s, MT took a huge leap forward. By 1996, the world’s first free online translation tool for short texts became available. And the rest, as we say, is history.

Current and future MT trends

Two major events have shaped current MT trends. Of course, the coronavirus pandemic is one of them — it created the supply chain and market disruption, changed the demand landscape, and had a substantial financial impact on the industry. Uncertainty brought about by COVID-19 will extend well into 2021. However, the life goes on and research and development of machine translation progresses rapidly partly due to the US-China trade war. The ongoing trade war has, for example, resulted in restricted export lists that include language technologies. And security concerns include allegations that Chinese MT services are used to collect data on users beyond China.

With these and other issues at play, the leading technology players continue to develop new technologies on both sides of the political and economic divide. Regardless of how these situations play out, it is clear that MT will become more and more advanced in the coming years. It is also evident that MT is used well beyond the translation industry.

MT is present in many business areas, as well as in the public sector. It helps in building digital bridges between public administrations, facilitates international multilingual communication with other nations, and enables everyone to instantly and securely access and exchange multilingual information. In the business environment, MT helps in entering new markets much faster, speeds up business operations, reduces translations costs, and helps in improving internal and external communication. Businesses and translators who take advantage of the current advancements in the MT technology such as the dynamic learning technology will definitely benefit from their competitive advantage.

Over the decades, MT has progressed from dictionarybased translation of phrases to a more general rule-based MT. More recently, the MT paradigm has shifted from statistical MT to neural MT, and later to adaptive neural MT (dynamic learning). MT is a very powerful tool that has become available to translators. But translators are also aware of a substantial quality gap between general and custom systems, primarily due to the overall accuracy and the very specific terminology required for domain-specific translations. Many translation agencies recognize that it is not always affordable and time-efficient for them to develop custom MT engines for each and every domain. In this case, the dynamic learning technology offers a solution.

What is dynamic learning, and how does it help translators?

Neural MT is far more advanced and useful than previous generations of MT systems. However, it still struggles to learn from translators’ feedback, and often requires correction of the same mistake in the MT output over and over again. Dynamic learning is customized to overcome this problem by allowing the MT to instantly learn from translators’ post-edits. Dynamic learning analyses every correction made by translators and adapts the translation engine to meet lexical, terminological, and stylistic requirements.
It means that the post-editing outputs are applied to future translations to save time and effort.

Seamless integration of dynamic learning MT into computer-aided translation (CAT) software tools used by most professional translators has a clear advantage. CAT tools save previous translations in translation memories and reuse them when similar sentences are being translated. In addition, if no similar sentences are found in the translation memory, CAT tools offer translations generated by the MT. However, many translators are still relying on MT tools that keep mistranslating the same word or phrase over and over again. With dynamic learning, this problem is easily solved, as MT is able to make use of translation memories and adapt. This means that unlike before, when translation memories were useful only for sentences similar to those that had already been translated, the current MT reuses them to improve the translation of specific phrases and expressions in novel sentences.

Another significant advantage of the new approach is the ability to use the same MT engine with various domainspecific translation memories to serve as a separate custom engine for each domain. While dynamic learning MT is more expensive than a standard MT engine, it is considerably cheaper than building several custom engines. Dynamic learning allows MT engines to deliver close to
custom domain-specific MT engine quality for medium-sized translation projects for which custom MT engines might be unfeasible.

Finally, dynamic learning complements other MT technologies. For example, professional translators often work with bilingual glossaries that provide translation domain or customer-specific terminology. Getting the terminology right is essential for high-quality translation in many technical domains with bilingual glossaries that provide translation domain or customer-specific terminology. Getting the terminology right is essential for high-quality translation in many technical domains.

What do translators think of dynamic learning?

Diāna Breita is an experienced translator working in a medium-sized LSP, whose customers range from governmental bodies to IT companies, financial institutions, and banks.

Breita’s first experience with MT was five years ago when statistical machine translation became available to the translation industry. It was helpful in specific situations such as translation of technical documents with highly technical terminology and textual information. However, she says that very often, it actually made her job more difficult. Breita notes that sometimes the text was merged, and in some cases she and her colleagues had to start documents afresh. Breita also remembers the early days of MT, when she and her colleagues collected lists of funny translations produced by the MT. Every day, new words or sentences were added to the list — but now these lists are hardly ever updated, as there are rarely any bad translations.

“With the development of neural MT, we could translate longer and more complicated texts. But currently I am working with dynamic MT, which is much better. It is certainly more coherent. Dynamic learning analyses every correction I make and adapts the MT model we use,” Breita explains. “Consequently, the translations require far less translator engagement. They are more human-like, and I can translate larger volumes in less time.” When asked as to whether she fears that one day AI will replace her, she explains that by being ahead of the game and using MT, her position in the market is stronger rather than weaker.

Sandra Zdanovska has worked as a translator for 18 years. She specializes in technical and creative translations. “As a translator, I always notice mistakes in texts. In the early days of MT, the translated texts were often clumsy and the mistakes were obvious. Now, the MT systems have evolved. On the one hand, it is good, as they require less editing. On the other hand, it is bad, as mistakes are also much more difficult to notice.”

MT is not only widespread in the translation industry, but also used in many other areas of our lives. Individuals, businesses, governments, and international organizations will rely on the MT in 2021. Considering the technological and usability advancements of the MT, it has evolved to a stage when it benefits all of us. Translators also recognize that the MT has become standard in LSPs. And by adapting and learning how to use the AI to their advantage, the job of translators has become more straightforward, enabling them to become more productive in the workplace.