Artificial intelligence seems to be affecting all areas of our lives and businesses – and it’s here to stay. It’s certainly able to help us with many tasks and can be very useful. In the sphere of translation, these rapidly emerging changes have had a direct impact on us at LocalizationLab. We’re excited to be a part of it, and be able to pass on the advantages to our clients, because neural machine translation can help reduce cost and time.
Even so, while it’s true that neural machine translation (MT) engines are getting better and faster, they are far from perfect. And this is where the role of translators is key, as now they are increasingly working with post-editing rather than translating from scratch. In this article, you can get a closer look at the views of translation professionals on the use of machine translation and post-editing.
1- What is neural machine translation (NMT)?
Neural machine translation (NMT) is a machine translation method that uses artificial neural networks and it’s seen significant improvements in recent years. Before neural machine translation, machine translation engines were based on statistical methods with more limitations and errors.
When it comes to machine translation, it’s important to take security and confidentiality into account, because when we use free online platforms, there’s no guarantee that they are secure or confidential. The texts we put into them become part of the translation engines belonging to the companies offering them. To ensure the engines we use are secure and confidential it’s important to use paid versions that guarantee this.
2- Post-editing (PE): What is post-editing?
Machine translation post-editing (MTPE) is the review of an automatically translated text (using a neural machine translation engine (NMT)) by a professional translator or specialised linguist. A translator’s work in post-editing is different because they will have to take into account that the translation has been done by a machine, which is not always consistent and may have cultural and contextual errors.
This makes the translator’s work more creative and more similar to transcreation, giving it greater added value.
3- When can we use machine translation post-editing?
It is not always suitable to use machine translation and post-editing; this depends to a great extent on the type of text, the language combination and the function of the translation (commercial, informative, advertising, technical, etc.).
Each order is analysed by us to decide whether it’s appropriate to offer this service with a view to reducing cost and time, and if it is, we will always let the client know. Before coming to a decision, we test the machine translation engines and analyse the results.
Additionally, not all machine translation engines work the same way, and they change rapidly, so it is important to test them regularly.
We recently carried out two tests to check the quality of machine translations for two of our clients:
Test 1: Article on a customer’s success story from English into Japanese, Traditional Chinese and Simplified Chinese.
The results were fairly different depending on the language:
- In Simplified Chinese, we noticed that when the sentences were short, the result was pretty good. However, when it came to long and more complex sentences, the final text always had to be modified, and in some cases, where the translation engine divided the sentence into smaller segments, the translation was impossible to understand and had to be completely redone.
- Traditional Chinese surprised us. Although the engine offered this language, the result it came out with was Simplified Chinese rather than Traditional Chinese. Both languages are similar, but they are still different languages. The reason for this seems to be that most engines have been ‘trained’ to use Simplified Chinese. However, the Taiwanese publisher said it could still be used, and even with the language difference, encountered the same drawbacks as the Simplified Chinese publisher.
- In Japanese, the result was the worst, maybe because the Japanese are more demanding:
– Almost every sentence had to be modified by a linguist
– Some proper names had been translated while others were left in English, so the engine wasn’t consistent
– Most sentences had to be modified
– Sentence connectors had to be added or edited, probably because the engine didn’t have sufficient context.
However, all three editors of these Asian languages agreed that the result of the machine translation would help them reduce the writing time in the target language.
Test 2: An excerpt from a veterinary article and an excerpt from a marketing text from Spanish into Italian, English, French and Portuguese.
In this case, the results were more similar:
- All four reviewers agreed on two things:
1- The translation of the scientific text came up with the typical defects many machine translations make:
– Tendency to be too literal
– Some terminology errors
– Formatting problems (spaces between text and certain characters like ,:;…)
But it could be used perfectly well to work with, edit (to improve naturalness, fluency, and making it more idiomatic) and use in any context.
2- The translation of the marketing text could not be used because the machine translation had not been able to capture and convey the linguistic and idiomatic nuances used in the original.
In French, the reviewer was more critical and considered that using a machine translation as a starting point may be counterproductive, because it could influence the reviewer in a negative way and be detrimental to the final result.
4- Levels of post-editing:
Jeff Allen, in the chapter on ‘Post-editing’ in Computers and Translation (Harold, 2003) introduced the notion of PE levels. Depending on the result we are aiming for, more or less effort will need to be allocated to post-editing. He talks of ‘light post-editing’ or “hard post-editing’, depending on whether we require a lot of translator intervention or not. A lower quality text may be acceptable for an internal text, but if it’s for external communication, a commercial or marketing document for clients, or a medical report, we will probably have to spend more time and effort to ensure that the quality is at the expected level; it all depends on our objective and what we are looking to achieve.
Guidelines for post-editing can be consulted in ISO ISO 18587:2017.
Nowadays in the translation sector, it would seem strange not to use artificial intelligence, given the scope that new technologies have to help translation professionals and the increase in volumes we can translate (due to reduced costs and time).
However, it is essential to ensure that the quality of the final text meets our needs, so the role of post-editing is becoming increasingly relevant.
We work hard to ensure we’re constantly up to date with the latest technology, always placing the confidentiality of the content at the forefront. We always share our test results or findings with our clients in order to optimise financial resources.
Other related articles: