Internalising AI translations, how the company entrusts the correction of machine translation to employees

The performance of machine translation has evolved considerably thanks to advances in artificial intelligence, especially in Deep Learning techniques. This breakthrough has led companies of all sizes to internalise translation and mobilise their workforce to perform AI correction. This strategic reorientation of translation needs, quickly perceived as an exceptional cost killer, leads to a loss of employee productivity and unexpected costs. A solution to repositioning the professional translator in the internalised translation workflow is emerging.

The growing performance of machine translation

About ten years ago, we wrote a blog post, comparing human translation and machine translation, and it accepted the idea that these two methodologies were comparable.

However, since the two processes were completely different in terms of process, cost and result, there was no point in comparing them. We now see that it was useful ten years ago, but is no longer relevant today. Machine translation has reached a level that even translation agencies and multilingual service agencies use (Voice Cloning, time-coded transcription) and have even created specific services based on machine translation such as Post-editing or Machine Translation Checking (MTC). Why? Simply because with certain language combinations, certain types of texts in certain areas of translation, machine translation is very effective.

Grammatical accuracy allows machine translation tools to avoid common errors, fluency significantly reduces the number of awkward sentences, the increase in contextual understanding allows us to avoid translation errors due to ambiguities or homonyms and finally, it is getting better at managing idioms and complex phrases.

The performance of machine translation has continued to grow and the measurement indicators BLEU (Bilingual Evaluation Understudy), METEOR (Metric for Evaluation of Translation with Explicit Ordering) or BERT (Bidirectional Encoder Representations from Transformers) indicate an increasing correspondence with human translations.

These indicators measure the quality of machine translations by comparing, a same source text, the translation generated by one machine and another one performed by a human. To do this, they compare segments of words of equivalent length in the machine translation and human version and count the number of exact matches between the segments automatically translated and human-translated segments. The score is given as a percentage of similarity. Theoretically, a score of 100% would indicate a perfect match between a machine translation and a human translation. Today we are at 30 or 40% in the best cases.

The widespread internalisation of translations

Perceived as a game changer, the machine translation, which has become effective, significantly changes the dynamics of the economic and cultural sectors. It has logically led to a change in the companies’ habits that have gradually decided to cut their translation budget from translation agencies and other linguistic service providers (language service providers) and head to, exclusively for some, to machine translation. Since the service is almost free, it appears that there is an immediate financial gain.

Can we go from a 100% human translation to a 100% automated translation without any consequences?

The necessity to check and correct machine translation

In fact, the scores of the BLEU, METEOR and BERT tools vary considerably depending on the language combinations, the technicality of the documents, the field of application; The European Commission has been working on this subject since the beginning of the year. Progress is uneven across languages, with performance still insufficient for less documented languages or having few resources available to train models.

Among the notable performances, Google Translate achieved an average BLUE score of 31.07 for translations from non-English languages to English and 27,54 for translations from English into non-English languages. These scores, among the highest reported for machine translation systems show that 60% to 70% of a machine translation requires careful checking and correction.

In concrete terms, if the automatic translations from or into English give a satisfactory result, languages with very different grammatical structures from those of the majority languages and languages with ideograms (Chinese, Japanese) still present challenges. The BLUE scores for translations from and into Chinese can vary between 10% and 40%. At the 6th edition of the Workshop on Open-Source Arabic Corpus and Processing Tools (OSACT6), various machine translation models are developed to handle translation between Arabic dialects and Modern Standard Arabic and have achieved a BLUE score of 21,0 on the development game and 9.57 on the test game, a very moderate performance.

While modern tools are better at understanding context, they may still lack cultural sensitivity, producing translations that are technically correct but culturally inappropriate. Due to the presence of misunderstandings, poor terminology and syntactic heaviness, the machine translation remains very imperfect and requires correction when the expected objective is higher than a correct understanding of the content by the reader.

Loss of productivity due to the correction of machine translation in the company

Companies operating internationally have gradually generalised the use of automatic translation tools Google Translate, DeepL or ChatGPT in order to quickly translate large volumes of text. The employees are responsible for proofreading and correcting the translations generated. Their mastery of business jargon and industry terminology ensures accuracy and consistency of the final product.

With an average match of 20-30%, between machine translation and human translation, the proofreading phase takes much more time for a non-native employee and longer for a non-professional translator. In concrete terms, where a professional translator of native language can proofread and correct about 5000 words in one day, a non-native employee will need double, triple or more. And this means having long time slots to be able to concentrate on the text, thus, proofreading cannot be done on the go, between two phone calls or two emails.

In addition, the proofreading and correction of machine translations are perceived by employees as repetitive and monotonous tasks; the result is a loss of motivation and job satisfaction. Moreover, it is not an easy exercise and requires constant vigilance, the long-term effect of this can be to increase cognitive fatigue and reduce employee efficiency.

Time-consuming activity is difficult to reconcile with the daily schedules of companies in optimised or sub-workforce. The expected financial gain is significantly reduced, as it is often the case in times of crisis and economic uncertainty.

The gain is even reduced when employees must undergo specific training to be effective in post-editing. This training is an investment of time and resources for the company, and not all employees adapt easily to this new task.

Repositioning the professional translator into machine translation

The repositioning of the professional translator in the internalised translation workflow links the financial gain expected by the company and the maintenance of internal productivity.

Professional translators easily detect and correct subtle errors that machine translation tools and employees can overlook. They are able to ensure that the translated text meets the cultural and stylistic expectations of the target audience, which employees can hardly achieve. They can interpret the context more accurately, which is essential to avoid misunderstandings and misinterpretations. They can adapt the tone and style of the text to match the expectations of the target reader, something that machines and amateur translators struggle to do effectively.

Through its two low-cost translation services integrating the AI that are the correction of machine translation (machine translation checking) and the revision of automatic translation (Post-Edition), Atenao responds to the logic of optimising business costs while maintaining productivity. Translations are outsourced, but at lower prices. This segmentation, with 4 levels of translation and 2 levels of proofreading, offers a specific answer for each need.