Unlike traditional translation, machine translation (MT), in theory, doesn’t require any human involvement.
Instead, a computer is enabled to translate text automatically from one language to another. The benefits are obvious – time and money are saved. What more could you wish for?
But not so fast. While MT is a great solution for many different scenarios and EQHO certainly does endorse it, it isn’t universally viable.
Here are some basic MT concepts to help you decide:
How does it work?
There are three main types of MT – Rules-based, Statistical-based and the latest technology, Neural MT. The rules-based system uses language and grammar rules and a dictionary filled with common words to carry out translations effectively. Specialist industry-specific dictionaries can also be developed and incorporated, typically delivering consistent translations with accurate terminology.
Statistical-based systems take a different approach, as they have no knowledge of language rules. Instead, this system analyzes large quantities of translation data (language corpus), which the computer ‘learns’ from. This data can be anything from thousands to millions of strings (sentences) of a chosen language pair (e.g. English > Thai, Chinese > Japanese etc.). This is referred to as ‘Training’ the engine. The translations produced by statistical engines are usually more fluent, but less consistent.
Neural machine translation (NMT) is an approach to machine translation that uses a large artificial neural network to predict the likelihood of a sequence of words, typically modeling entire sentences in a single integrated model. Neural MT requires only a fraction of the memory needed by traditional statistical machine translation (SMT) models. Furthermore, unlike conventional translation systems, all parts of the neural translation model are trained jointly (end-to-end) to maximize the translation performance.
Why would you use machine translation?
Return on investment (ROI) typically comes from savings made in the translation process. Machine translation can save companies money in two separate ways; by improving productivity and driving progressive improvements in profit margins as custom machine translation engines mature. The quality of machine translations can improve in a very short space of time as the system benefits from post-editing feedback. For example, a project by one major statistical-based MT engine developer was claimed to improve profit margins from 25 percent after one translation to between 60 percent and 70 percent after 20 translations.
Custom MT engines
If you want to produce translations that require the least amount of human editing in order to publish, you need to customize the translation engine being used. This means building on the lessons learned from previous projects. By leveraging a comprehensive range of automated tools and a human-guided process, a near-human quality finished version can be created. Using a granular custom engine means that every engine will be focused on an individual client, product and target audience. According to the same major MT engine developer, productivity gains of between 150 percent and 300 percent can be realized, while more than 50 percent of translations carried out by custom engines will not require any work. Translations are also much faster – human translators typically translate 2,000 to 3,000 words per day, whereas machine translation can translate hundreds of millions of words.
Control
In the past, control was limited when using machine translation, as the information was simply translated and then edited. However, most modern day MT systems show stakeholders every stage in the translation process so that it can be reviewed and fine-tuned. For example, if text is not translating correctly, the cause can be examined and a solution found to prevent the issue from recurring. Two issues that may affect a translation are the complexity of the language pair or the subject domain, and the quality of the source content. Garbage in – garbage out is an important concept to understand across localization, and even more so in the realm of MT. However, once this has been established, customized training plans and rapid quality improvement plans can be put in place to deal with the problem.
MT post-editing
There is no doubt, when looking to disseminate maximum amounts of content, machine translation can be a great option. It allows brands to translate high-volumes of content at reduced costs and in shorter timeframes than using only humans and can satisfy ‘low quality – high quantity’ project requirements. However, the use of trained linguists is vitally important if firms want to produce the best possible final output. By completing a full MT post-edit, brands can be sure they will have a translation that is not only accurate but also stylistically appropriate. At the highest level of post-editing, the finished product will be of a quality level that is indistinguishable from that of professional human translators.
‘The Human touch’ MT post-editing in more detail
Light post-editing or ‘MT 1’
Light post-editing requires minimal intervention by the post-editor, and is primarily intended to help the end-user or ‘MT 2’ Post-editor understand the meaning of the content. Typical purposes are for company internal usage, often when the text is needed urgently, where the project has a short timeframe, or as an intermediate step in the full MT process.
Full post-editing or ‘MT 2’
Full post-editing requires a higher level of interaction than MT 1 to achieve an elevated level of quality; the expectation is that the output will be a translation that is not only accurate but also stylistically appropriate. Typical uses for full post-editing MT include external company use, similar to if not the same as expectations of translations produced by human linguists. At the highest end of the post-editing scale, there is the expectation of a quality level which is indistinguishable from that of professional human translators.
Where and when to use MT?
There is a time and a place for MT. Not all content lends itself well to machine translation and some languages are simply not at all MT friendly. For example, such is the flowery and idiomatic nature of marketing materials and the need for creativity and brand consistency, MT is not recommended. Likewise, Japanese is so complex and has such a diverse readership, that it may be a while before Japanese MT can consistently satisfy a discerning Japanese audience. Just try to Google Translate some marketing copy from English into Japanese and you will start to see what we mean. Although Google isn’t at the absolute forefront of MT technology (just give them a few years), it is nevertheless a complex and impressive MT engine and can still give an indication of what types of content and language pairs may lend themselves better to MT usage than others.
On the flip-side, if your business involves frequently producing technical manuals written in controlled English and localizing into French, Spanish, German and Italian, MT may very well be for you.
It really comes down to what type of content you will translate, your language pairs, how much language corpus is available for your pairs (if you are considering statistical-based MT), who your audience is, and your quality requirements.
Do you really need the highest quality for all of your content?
The answer is probably ‘no’. It may be that you only require help systems to be translated with minimum quality requirements; after all it isn’t your actual product pages and functional ‘non-stylistic’ content wont be held against you by your customers. In some cases, you may not even expect a single hit for certain web-based content, but it is never
theless a legal requirement to have the pages in all European languages, so why spend all that money to translate using specialist human translators when MT will satisfy the requirement?
With all of this information taken into consideration, you can start to establish if MT may be a viable solution for you and your business.
