Machine translation (MT) technologies advance day by day, reaching new markets. As this trend continues, companies and developers double their efforts to increase accuracy and contextual understanding. Human-in-the-loop (HITL) translation appears in this scenario as one of the emerging approaches to tackle this issue.

HITL translation is a collaborative approach that leverages the strengths of both machine translation and human expertise to improve translation quality. In simple terms, HITL translation integrates human translators’ feedback into the MT workflow, constantly tuning systems in. This method ensures accuracy, cultural appropriateness, and understanding of nuances in language (Source: Pangeanic Blog).

In this article we’ll discuss what is it, how it works, and its current use and developments.

How Does Human-in-the-Loop Translation Work?

HITL translation follows the developments in HITL machine learning. The process works by using an MT system to translate large volumes of text. However, instead of relying solely on the machine’s output, human translators post-edit the output to review and refine the translations. Language specialists must then make all the necessary adjustments to capture the intended meaning and tone, and correct grammar mistakes or inaccuracies.

This approach is different than annotation, where human translators must add comments on top of the edits. In comparison, HITL relies heavily on the interaction between the human and the machine. Translators’ feedback is used to fine-tune the MT system, enabling it to learn from its mistakes and improve its performance over time.

The HITL translation process typically involves the following steps:

1. Machine translation: An NMT system translates a large volume of text.
2. Human review: Professional human translators review the MT output. The focus is on ensuring accuracy, fluency, and cultural appropriateness.
3. Feedback integration: The post-edits are used to fine-tune the MT system, enabling it to learn from its mistakes and improve its performance over different iterations.
4. Continuous improvement: The process is repeated, and the MT system continuously learns and improves its output based on human feedback. Each new training will be over the tune-in system, narrowing the error margins down in each re-training.

Benefits of HITL Translation

Adopting a well-designed human-in-the-loop approach in a translation workflow can bring a wide range of benefits, including (source: Khan, 2024):

Improved translation quality

Integrating the expertise of human translators into MT workflows can leverage MT’s shortcomings in terms of accuracy, fluency, and cultural appropriateness. This is especially important in cases of ambiguity or context-depending texts.

Quality assurance

Constant assessment of MT outputs with the goal of post-editing to fine-tune systems means more attention to the overall quality. As feedback in “real-time” is fed into the engines, the errors decrease as well. Moreover, HITL approaches (see for instance Human-in-the-Loop by Translated) include automated metrics, quality checks, and PE evaluations. Thanks to this, HITL systems maintain translation standards.

Increased efficiency and adaptability

Human-in-the-loop translation leverages the speed and scalability of machine translation, while still maintaining high-quality standards. The key is ensuring translators’ feedback is easily integrated into the engines without the need for extra steps like downloading and uploading files into the cloud.

Fine-tuning systems as translation goes, guarantees that engines adapt their language models to fit language trends, terms, or dialects.

Cons of Human-in-the-Loop Translation

While human-in-the-loop translation offers many benefits, especially in terms of quality and language quality assessment (LQA), it also poses some challenges. Some of the potential cons include (source: Mosqueira-Rey, E. et al., 2023):

Dependence on human evaluation

Relying on the human factor has some drawbacks such as time constraints and quality differences. To implement a HITL workflow, it’s necessary to have linguists available at the drop of a hat (which for logistic reasons might not be feasible, even if the solution is implemented on a crowdsourcing platform).

Moreover, the quality of the final output depends heavily on the judgment and expertise of the human translators. Human judgment can vary in quality and consistency, so extra steps (review and LQA) must be taken to minimize this variation to keep translation standards.

Increased costs

Dependence on human evaluation also means an increase in cost. While still cost-efficient in comparison to human translation, human PE and feedback integration into systems in real-time or frequent iterations, add up to the overall MT costs.

Budget is always a sensitive issue in translation workflows for project managers and companies. Especially when it comes to implementing technological solutions.

Limited scalability

Although MT systems can handle large volumes of text, the human review and feedback integration process may not be scalable to the same extent. This can potentially limit the overall efficiency of the process if not handled properly.

Human-computer interaction concerns

Deep interaction between the human and the machine means that more efforts must be put into user-friendly platforms. Translators must be able to navigate the MTPE platform to reduce errors and time on task. This, again, implies more effort in designing the best environment for users to perform their tasks and feed that into the systems.

As we mentioned at the beginning of the year, the developments in terms of collaboration between the human and the machine will also redefine the role of linguists in the translation process. Concerns regarding ethics, remuneration, and translators’ status still need to be addressed so HITL translation is scalable and sustainable for all stakeholders.

HITL translation represents a game-changing approach to managing multilingual content with more precise MT outputs. As the language services industry continues to evolve, human-in-the-loop translation will likely continue increasing its role in the MT industry to reach language accuracy.