Why the industry will survive

When we talk about text creation, translation, AI editing, machine translation post-editing, and so on, in every case, we compare the text created by such tools with the text created by a human, i.e., with a benchmark, and we provide not only a technical (automatic) assessment, but also a human (manual) assessment.

Let’s take a closer look at machine translation (MT) and the creation of MT models. Training the model requires data (a large corpus of text). There is no limit to the amount of data, but high-quality data that has been verified by linguists and other subject matter experts is worth its weight in gold. After that, an MT expert customizes the model. Once the model is finished, it produces a translation of some text; however, human evaluation is the only way to determine the quality of the text. We can evaluate MT using a benchmark, which refers to a correct and proofread translation. Additionally, we can analyze MT output without relying on a benchmark by checking grammar, style, and identifying all kinds of errors using both automatic and manual methods. However, the final decision always rests with a human expert. Even with a high percentage of similarity between the machine translation or some other text generated by similar means and the benchmark, only humans can detect important nuances or catch gross errors. Any kind of machine translation (or MT model) can be improved indefinitely, but not without human involvement.

Now, let’s break down machine translation post-editing (MTPE). MTPE can be performed by humans, artificial intelligence (AI), and other AI-based tools that follow specific criteria. Just like with machine translation, any AI tool that performs MTPE needs a model, which is handled by profile specialists. Similarly to MT, MTPE requires data (high-quality text samples), processing algorithms, prompts, etc. The quality of MTPE should not be evaluated without human involvement either. The editing done by AI may include, for example, improvements to style and tone, correction of various types of errors, and more. However, there are always subtleties, exceptions, complex or non-standard cases that require the expertise of human professionals.

The same is true for editing any text and for editing a translation with or without an original reference while using AI. Human involvement is indispensable for these types of work. There are instances where a certain model might be well-suited for certain subjects, but there is no model that can be universally applicable. Professionals create multiple models, combine and improve them, conduct experiments, and utilize additional features related to context and other details, which are usually very challenging to work with. All of these efforts are aimed at achieving optimal outcomes.

You are unlikely to find unanimity on the quality of a certain text. Even if terminological consistency, other style requirements, and other aspects are observed, opinions, evaluations, and points of view may still differ regarding sentence structure, specific text fragments, wording, and many other nuances. When there are two or more experts, whose points of view differ at least in some ways, each might have their own valid opinion. This applies not only to literary translation, but also concerns technical, legal, and other domains. And it makes perfect sense. In my opinion, that is the main reason why such technologies will not take over human jobs. Only a human will choose the “right” text. Besides, training and further refining a linguistic model that will generate a particular text require a good team of professionals, such as mathematicians, linguists, and experts in various other fields. They study people’s opinions, not just “machines,” to improve the model.