A Deep Dive Into the State-of-the-Art AI Solution for Translation

Translated Human in the Loop

While many companies are looking to large language models (LLMs) for their translation and localization needs, LLMs are not the state-of-the-art solution. Neither is generic neural machine translation (MT). The modern enterprise requires a dynamic, customizable solution that can quickly adapt to its ever-changing needs. The answer is adaptive neural MT, which, according to a recent comparative MT evaluation by Achim Ruopp, founder of Polyglot Technology, outperforms leading public MT systems and the best LLM for translation, GPT-4. Let’s explore why adaptive neural MT is the cutting-edge solution for businesses.

To improve any artificial intelligence (AI), you must find errors, fix them, and retrain the system. The strength of adaptive neural MT is its ability to quickly and continuously learn from customers’ data, ongoing corrections, and document context, helping it improve as MT usage expands within an enterprise. This constant corrective feedback ensures that translations match the intended tone and terminology for the target audience and increases overall productivity.

While LLMs can sometimes provide high-quality output with much coaxing and effort, they can also be unreliable and inconsistent, much more prone to hallucinations, slow, and difficult to embed in highly automated translation workflows. LLMs are also currently limited in their language coverage, and these limitations will take time to resolve. 

In addition, the most successful AI deployments need continuous learning and improvement capabilities to meet evolving business and use case requirements. This activity demands rapid error detection capabilities, the ability to incorporate closely integrated corrective feedback, and automated and dynamic updates to the translation model to keep it in step with the latest requirements.

Translated’s ModernMT offers all of these crucial benefits and more. Now enhanced with Human-in-the-Loop, ModernMT is the first dynamic, adaptive MT solution that combines advances in AI with human reviews in the background to deliver continuous, real-time quality improvements. Every day, based on user requirements across languages and projects, a percentage of the delivered machine-translated content is automatically sent to professional translators for review and revision, providing feedback for immediate improvement of the MT system. Everything happens in the background, without any need for intervention by the user after they have configured their approach in the dashboard. You can learn more from Translated’s VP of AI Solutions, John Tinsley, in the video above.

This cutting-edge solution is powered by an adaptive Machine Translation Quality Estimator (MTQE), which rapidly analyzes all translated content and ranks segments by the likelihood of errors. This way, human reviews are directed where they’re needed most. By addressing challenging errors and retraining the MT system in real time, we can more efficiently address problematic segments in subsequent translations, leading to continuous improvement of the MT system. Translated’s adaptive MTQE has another distinct advantage over generic MTQE models. Similar to adaptive MT, it learns from customer data to make better decisions about quality. This approach allows users to solve a large proportion of significant problems in a corpus by focusing revisions on a small percentage of the worst-scoring translations

This crucial difference is critical for companies localizing large volumes of data, as human oversight of these volumes is near-impossible. It’s a prudent approach to managing risk when companies want to go multilingual at scale, enabling faster turnaround times and increased efficiency.

While adaptive neural MT is the state-of-the-art solution for enterprises, that may change in the coming months or years. LLMs hold great promise for the future of automated translation, but right now, they’re not ready to replace MT for enterprises. However, according to recent experiments conducted by Translated, they’re likely to reach this stage very soon. As with any AI, implementing LLMs for enterprise translation will require real-time adaptivity to meet business needs and provide continuous, automated quality improvements.