Technological progress – Machine translation!

Table of contents

Human translation, machine translation or digitally managed translation? These days, there are as many ways to translate your language project as there are roads to Rome. In this blog post, we’ll focus on machine translation to give you a peek behind the digital syntax and semantics.

The history of machine translation

Let’s start by taking a brief look at its origins. Before the development of artificial intelligence and computational linguistics , human translation was the only option available. It was the measure of all things.

The first developments in machine translation date back some 60 years. The earliest known project was initiated by the US military using a translation program that could translate from Russian to English without an interpreter. But by 1966, all research into machine translation was abandoned as it was deemed unfeasible by the U.S. Department of Defence.

Over the next 20 years, research efforts stagnated. It was not until 1980 that a handful of companies in the electrical industry decided to resume research, including Siemens AG, which conducted special research in cooperation with the University of Saarland.

This collaboration resulted in two different systems – the SUSY system, which could translate from and into German, and the ASCOF system, which provided additional semantic information for the translation process. Later, research achievements included the development of a statistical translation system in 2006 and, 10 years later, the incorporation of neural networks  into translation programs.

Limitations of machine learning

Today, the quality of a machine translation (also known as MT) is almost comparable to that of a linguist. However, comparable does not mean equal, as machine translations have their limits. They include the following problems:

• The author’s style of writing can become lost
• Mistakes in dictionaries (the collection of already translated words in the translation program)
• Technical problems in computational linguistics
• The editor knows the target language too well and changes the language too much (additional time and cost consideration)
• Idioms, puns, expressions and colloquial language cannot be translated 1:1

Despite these drawbacks, the demand for machine translation continues to grow. One of the main reasons is the advance of digitalisation. Many organisations are now digitising their legacy documents, while others are going paperless from the outset. This includes the digital processing of translations into every required language.

Companies that operate in East Asian languages, which make up a large part of the international economic market, tend to prefer AI-powered translations, for example. Another major demand for machine translations is within the military domain, where they need to communicate quickly in conflict zones and provide direct assistance to people in foreign countries.

DeepL, Google Translate & co. as a quick solution

There are several compelling arguments in favour of machine translation. Particularly for non-commercial use where errors aren’t an issue. Even beyond private use, we know how tempting it can be. It’s fast and free, after all. And having medium to large documents professionally translated is a project of its own in terms of organisation and cost.

These days, there are a number of free translation tools available. However, like everything else in this world, they have their pros and cons. As a translation agency, it might be obvious that we’d quickly advise against using a free service. But there’s more to it than that.

When free becomes expensive

We’ve all seen MT translation fails when they go viral. And they’re often very funny – for anyone not involved with your brand, campaign or company. In reality, dealing with the aftermath of a mistranslation is far more expensive and problematic than using a professional in the first place.

One of the biggest issues with machine translation is the lack of quality control. Actually, the list of issues is long, so we’ll make it as brief as possible:

In addition to there being no quality control or consistency checks, MT can’t interpret meaning, can’t recognise social or cultural oversights, can’t detect errors in the original text, or use an organisation’s style guide, tone of voice, or terminology rules.

The translations are not unique, certain phrases can’t be translated directly, the accuracy doesn’t always cut it and all of this can damage your credibility.

If the organisation publishing the translation doesn’t speak the second language, its customers will be the first to spot the errors. Mistakes or a bad translation can undo all your hard work in an instant. The wrong message will reach an audience quickly but stay with them for a long time.

Go hybrid

The good news is that machine and human translation are not mutually exclusive. At intercontact, we close the gap between language engineers and linguists by augmenting machine translation with professional post-editing.

Even the best, most well-known translation websites cannot compare to trained linguists and editors. These professionals have far more knowledge and experience in editing machine translations and can therefore provide you with much better support.

For any marketing or promotional content, especially high-visibility materials such as slogans, banners, newsletters or your home page, you definitely need to work with human translators. Here, you want your content to be smart and to stand out from the crowd. MT technology can’t do that. Not least because it uses content found on the web that has already been translated.

A good agency, on the other hand, will analyse the text you need translated and assign the best and most suitable translators and editors to your project.

Post-editing is key

Professional machine translation methods

Translation technology is constantly evolving and requires different approaches. An automatic translation can be produced instantly by first translating the text word for word in the original order and then checking the sentence structure and grammar. It can also analyse the grammatical structure first and draw the necessary information for the target text from an intermediate language index (interlingua).

Other methods that are particularly popular include example-based and statistical machine translation. Example-based machine translation (EBMT) manages a translation memory by storing frequently used words and phrases. The system can accurately predict the work and costs involved in each project by calculating the matches already saved in the stored terms. A more advanced method is statistics-based machine translation (SBMT), which stores a large volume and variety of pre-analysed multilingual text in its memory.

Ultimately, both methods take an original  and translate it into another language. Depending on the type of text, topic, terminology and context, machine translation technology gets better every year, but it cannot replace a creative and experienced linguist.

by Sabrina Baumgardt