MT is here to stay: how do we make it work for you?

To answer that, apart from the Machine Translation (MT) engines, we should bring its invaluable colleague to the discussion — the Team. Now, we offer you to join the Machine Translation Post-Editing (MTPE) service journey with Alconost Team! Follow along!

Since technology is evolving rapidly, almost any content can be translated with Machine Translation. If not now, it'll probably be possible soon.

However, most often it’s not enough. Let’s see how the team makes it WORK.

1. First things first, we dive into the context of the client’s request.

We gain and benefit from as much information as the client can provide: briefs, glossaries, terminology. All of that helps us gain a better understanding of how to make the most cost-effective offer for the client.

2. Next stage is the evaluation of the MT engines.

At this stage, we consider all possible ways to achieve the best translation output. For instance, engine training, adaptive translation for Neural machine translation (NMT), and prompt engineering in the case of Large language models (LLMs). Let's dive in and explore two of the most popular methods we use in our practice.

2.1 Adaptive translation

Imagine you have a basic translation tool (or Neural machine translation (NMT)) that handles general text well. But you want your content to match specific criteria and a particular style. That’s the case when we offer adaptive MT — it tweaks the approach on the fly.

If you have any previous translations, we will prepare them for optimal use with the help of Translation Memory.

Adaptive MT responds to your unique needs, learning from your specific terminology and style to make translations more accurate and consistent. It’s like having a translation tool that understands your business as well as you do!

2.2 Prompt Engineering

Being a tech-savvy and trend-aware team we know that to receive high-quality content from LLMs you need to play by their rules and provide guidance. It takes effort, time, and a diverse approach. LLMs just love and respond to being told what to do with as many details as possible. We do it with proper attention to the context of the specific project and give them detailed instructions.

Why not just use a specialized AI tool or Chat GPT for that? — You may ask.

You definitely can, but it’s highly likely you won’t be able to tell if the quality is good enough. We’ll provide the service and quality achieved only with human and linguistic expertise. More about pricing and details you can find here.

Now, what’s so different between NMT and LLM?

3. Communication with the client on the upcoming evaluation can be of three types:

3.1 The default procedure — Express Evaluation.

It is a translation of a small sample (~2000 chars) of the content using a variety of MT engines. Additionally, a professional translator evaluates each MT variant based on quality and consistency. The highest-scoring MT engine is selected for the project.

3.2 Full Evaluation — for clients seeking deeper translation quality research.

What does it mean? The best result of MT translation according to the translator’s scoring undergoes the procedure of post-editing by a professional linguist to create a final result (reference translation). Next, we compare it with MT translations and bring numbers to the table. Metrics are calculated to select the best engine for the specific type of content and target language. The list includes, but is not limited to, the following metrics:

• TER — Translation Edit Rate is an automatic metric based on edit distance. It measures the number of edits (insertions, deletions, shifts, and substitutions) required to transform the MT output to the reference. It shows how much a human would have to edit a machine translation output to make it identical to a given reference translation.

• hLEPOR — is an automatic metric based on n-grams. It takes into account enhanced length penalty, n-gram position difference penalty, and recall. hLEPOR is an enhanced version of LEPOR metric. Basically, hLEPOR computes the similarity of n-grams in a machine translation and a reference translation of a text segment.

• COMET — is a neural framework for training multilingual machine translation evaluation models which obtains new state-of-the-art levels of correlation with human judgments. COMET predicts machine translation quality using information from both the source input and the reference translation.

3.3 Client-specific approach.

We value your time. Just let us know your preferences on the outcome, we’ll take it from here and deliver it without fail.

Summing the evaluation process up: choosing the best MT engine or translation approach for your project is simple and logical:

• First, we take a sample of your project's text and translate it using different methods, tailored to your specific context and needs.

• Then, we present the translation results to our translators for a thorough review.

• In case of receiving your agreement, we perform a test post-editing. We compare metrics and the translator's scoring (if ambiguous, communicate and clarify it with the translator) — to make sure we have chosen the right engine for translation.

4. The translation process itself.

We deliver the translations taking the results of the evaluation as our guide on the path to consistent, measured, cost-effective and high-quality MTPE services.

Can Machine Translation deliver high-quality content?

Yes, however, it still needs guidance and human expertise to accomplish that. Firstly, people are involved at every stage of Alconost MTPE service. Secondly, it's not about automation, but optimizing and increasing the efficiency of professional linguists’ work. And thirdly, the human has the last word in deciding on the quality of translation and adding both human touch and expertise to it.

Want to try MTPE? Write to us!

Related articles

How We Helped a Battery Testing Company Scale Across 6 European Languages through Machine Translation Post-Editing (MTPE)

Machine Translation

Multilingual marketing

Success stories

8 minutes read

Is there such a thing as too much innovation?

Machine Translation

12 minutes read