07 March 2025
How much can you save by using MTPE?
Or: how long is a piece of string?
In this blog, we’ll explore why quoting for MT and AI-assisted MTPE is so tricky, what factors determine the cost, and how Alpha works to ensure you get the best value for your investment.
MTPE stands for Machine Translation Post-Editing, a process that combines the speed and efficiency of machine translation (MT) with the expertise and finesse of human linguists. In essence, MTPE involves taking raw machine-generated translations and refining them to meet the desired quality standards for a specific purpose, audience, or context.
The goal of MTPE is to strike a balance between cost-effectiveness and quality, leveraging the strengths of machine translation while addressing its limitations. This process can range from light-touch edits to ensure basic accuracy and readability to in-depth revisions that align the translation with stylistic, cultural, and technical requirements.
Setting a general rate for post-editing is, quite frankly, impossible. There are just too many variables … Among them: the quality of the source text, content type, language pair(s), and other project-specific requirements, quite apart from the intended purpose, audience, and quality expectations.
Attempts at categorization such as Raw, Lite, Medium, and High also tend to confuse more than help.*
Rather than picking a figure from thin air, we recommend running a trial run on a few samples from your project. That allows us to determine which MT engine is best suited, and to experiment with AI prompting and other assistive tools (such as keyword extraction).
“Just” sending a project to Google Translate, DeepL, ChatGTP & Co. will rarely provide a satisfactory result (unless you have a very generic text with no specialist terminology, no particular stylistic requirements, and no concerns about consistency, or your corporate image).
If, on the other hand, you partner with Alpha to train the engine using your own approved TMs, glossaries, and stylistic instructions, we can significantly improve the quality of the output. Depending on the quality of the input (and the language pair, of course), you can expect to see considerable savings, which will only improve over time.
Alpha can provide you with further savings by harnessing AI. This involves evaluating the project, then formulating and testing the results before deciding on the best approach. This too, is open to further improvement over time, especially with Alpha’s expert linguists constantly assessing the output quality.
Initial prepping, whether for straight MT or for AI-enhanced output involves cleaning up existing glossaries and TMs. Depending on volume, this may take between 1 day to perhaps 2 weeks of preparatory work.
In some cases, it is worth improving and streamlining the source. This can start with something as basic as spellcheck and some homogenization of terms (particularly advisable when content has several authors making contributions, as in a knowledge base).
For effective post-editing of AI-generated or enhanced translations, significant effort and skill is required, particularly in crafting precise prompts at the start of the post-editing process to ensure the output is of sufficiently high quality to actually assist the human post-editor (i.e. to achieve the expected saving). For particular documents, prompts can be tweaked, for example, if a different target group is being addressed, or text types vary (white papers, marketing brochures, blogs, e-mails, etc.).
Another point worth making, which often gets swept under the carpet: If your projects are “bitty”, i.e. disconnected strings, alphabetically ordered, with no context, the gain you might make by using MT or AI could be minimal, or even negative. The engines, just like humans, need context if they are to do their job properly. Otherwise, they are, quite literally, in the dark, producing “random” translations. Just think of words such as “file”, “character”, “set”, “beam” or “command” that are polysemous – i.e. that can take on different meanings depending on context.
Not only do you need to decide on what type of MTPE you wish to go for, but another important decision is whether to restrict the use of MT-AI to all-new or low-match segments, while recycling existing TM segments, or to run MT-AI across the entire project. The latter may be sensible in cases where TMs are of questionable quality, or obsolete (say 15 years and older), or where terminology, style and register need a “refresher” – as when changing from a formal to informal register, and perhaps to a more modern style of writing altogether.
Over time, as we continue to train and refine the MT/AI process and improve the quality of the output for a client, we are typically able to achieve greater time and therefore, cost savings. These we pass on to our clients, ensuring you benefit from cost efficiencies as the workflow produces better and better results.
Get in touch for an initial discussion on how we can help or if you have any further questions. We look forward to working with you! … And the answer is: Anything from 10 to 75% …
*For more information on Lite, Medium and Heavy PE, watch this space.
