ModernMT

ModernMT is an adaptive neural machine translation system developed by Translated, an Italian language services company headquartered in Rome. First launched in 2014 as a research project supported by the European Commission's Horizon 2020 program, ModernMT has evolved into one of the most sophisticated machine translation engines available for professional translators and language service providers (LSPs). What distinguishes ModernMT from generic translation services like Google Translate is its ability to learn and adapt in real time from a user's translation memory and corrections, delivering increasingly accurate results that match the specific terminology and style requirements of each project.

The fundamental philosophy behind ModernMT is that machine translation should work alongside human translators rather than trying to replace them. By continuously learning from corrections and translation memories, ModernMT produces output that becomes more aligned with the translator's preferences over time. This adaptive approach bridges the gap between raw machine translation and the nuanced, context-aware translations that professional work demands. Independent evaluations have consistently shown that ModernMT outperforms leading generic MT systems for enterprise translation tasks, with the latest version (V7) showing quality improvements of 45 to 60 percent over its predecessor in human evaluations.

Adaptive Translation Technology

At the heart of ModernMT lies a two-component architecture: a Background Model and a Foreground Model. The Background Model is trained on billions of sentence pairs of general-domain data, providing broad linguistic knowledge across 60+ language pairs. The Foreground Model is created dynamically for each specific translation context, capturing and applying real-time adaptations based on the user's translation memory, glossaries, and corrections. This dual-model approach means that ModernMT starts with strong general translation quality and then refines its output based on the specific requirements of each project or client.

The real-time learning capability is particularly powerful. When a translator corrects a machine translation suggestion, ModernMT processes that feedback instantly and adjusts its subsequent suggestions within milliseconds. This creates a virtuous cycle where the system gets progressively better as the translator works, reducing the post-editing effort required with each subsequent segment. For LSPs handling large volumes of content for repeat clients, this means that ModernMT effectively learns each client's preferred terminology and style, delivering increasingly consistent and accurate first-pass translations.

Document-Level Context Understanding

Unlike most machine translation systems that process sentences in isolation, ModernMT considers the entire document's content when generating translations. This document-level context awareness ensures that terminology choices are consistent throughout a text and that ambiguous terms are translated based on the broader context rather than just the immediate sentence. For technical documentation, legal texts, and marketing materials where consistency is paramount, this capability significantly reduces the post-editing effort and improves the overall quality of the translated output.

The context-aware approach also helps with proper handling of pronouns, gender agreement, and register (formal vs. informal tone), which are notoriously difficult challenges in machine translation, especially for European languages with grammatical gender systems. By understanding the document as a whole, ModernMT can make more informed choices about these linguistic features, producing translations that read more naturally and require fewer corrections.

ModernMT V7: Latest Advancements

The latest major release, ModernMT V7, introduced several breakthrough features. Trust Attention is a novel neural architecture technique that increases the accuracy of terminology matching from translation memories and glossaries. Rather than simply suggesting terms, Trust Attention ensures that the model reliably uses the correct terminology from the user's resources, reducing the "hallucination" problem where MT systems generate plausible but incorrect terms. Advanced Terminology Control gives companies self-managed glossary control to ensure brand-specific and context-specific terminology consistency across all translations.

DataClean AI is another significant V7 innovation -- a sanitization algorithm that automatically identifies and removes poor-quality data from training material. This is crucial for organizations that maintain large translation memories that may contain inconsistencies or errors accumulated over years of work. By cleaning the data before it influences the model, DataClean AI ensures that ModernMT learns from the best available examples, reducing the likelihood of errors and hallucinations in output.

CAT Tool Integration

ModernMT integrates seamlessly with the most popular computer-assisted translation (CAT) tools used by professional translators. Native plugins are available for SDL Trados Studio, memoQ, Memsource (now Phrase), MateCat, and other major platforms. These integrations allow translators to access ModernMT's adaptive translation capabilities directly within their familiar working environment, without switching between tools or disrupting their workflow. The integration is bidirectional -- not only does ModernMT provide translation suggestions within the CAT tool, but the translator's corrections are fed back to ModernMT in real time, continuously improving the model's accuracy.

For LSPs managing multiple projects and clients, the CAT tool integrations support project-level configuration, meaning each project can use a different translation memory and glossary to customize ModernMT's behavior. This ensures that automotive terminology is used for automotive clients, legal terminology for law firms, and medical terminology for healthcare organizations, all without manual switching or configuration changes.

API and Developer Integration

ModernMT provides a comprehensive REST API that enables developers to integrate adaptive translation capabilities into any application. The API supports text translation with optional translation memory context, language detection, quality estimation, and glossary management. Authentication is handled through API keys, and the documentation includes code examples for major programming languages. For organizations building custom translation workflows, content management systems, or localization platforms, the API provides the flexibility to incorporate ModernMT's adaptive capabilities wherever they are needed.

The API also supports batch processing for large translation jobs, with asynchronous endpoints that can handle thousands of segments efficiently. Rate limiting and usage quotas are configurable, and detailed analytics provide insights into translation volume, quality metrics, and adaptation performance over time.

Language Coverage and Quality

ModernMT supports over 60 language pairs, covering all major European languages as well as many Asian and other global languages. Quality varies by language pair, with European language combinations generally achieving the highest quality due to the abundance of high-quality training data. The system performs particularly well on language pairs involving Italian, French, German, Spanish, Portuguese, and English, reflecting both the European training data emphasis and the company's European client base.

Independent evaluations by translation industry analysts and academic researchers have consistently ranked ModernMT among the top machine translation systems for professional use cases. The adaptive capability is the key differentiator -- while generic MT systems may produce adequate translations for casual use, ModernMT's ability to learn from professional-grade translation memories produces output that is significantly closer to production quality, reducing post-editing effort by up to 50 percent compared to non-adaptive alternatives.

Open-Source Foundation

ModernMT's core engine is available as open-source software on GitHub, based on the Fairseq Transformer architecture. This transparency allows organizations to inspect the code, understand the model architecture, and contribute to improvements. For enterprise customers who require on-premises deployment for data security or regulatory compliance, the open-source option enables fully self-hosted installations. The open-source version includes the adaptive learning engine, meaning organizations can benefit from the real-time learning capability even without using the cloud-hosted commercial service.

Pricing and Plans

ModernMT's commercial cloud service starts from EUR 25 per month for individual translators, with tiered plans that scale based on volume and features. Enterprise plans include dedicated infrastructure, priority support, custom model training, and advanced analytics. The pricing model is designed to be accessible for freelance translators while scaling to meet the needs of large LSPs processing millions of words per month. Compared to generic MT APIs that charge per character, ModernMT's subscription-based pricing can be significantly more cost-effective for high-volume professional translation work.

GDPR Compliance and Data Security

As an Italian company operating under EU law, ModernMT is fully GDPR compliant. Translation data processed through the cloud service is handled within European data centers, and the company provides data processing agreements for enterprise customers. For organizations with the most stringent data security requirements, the open-source option enables fully on-premises deployment where no data leaves the organization's network. This combination of cloud convenience and self-hosted security makes ModernMT suitable for even the most sensitive translation work, including legal, medical, and government content.

Limitations and Considerations

While ModernMT excels in professional translation scenarios, it is not designed as a general-purpose translation service for casual users. The adaptive capabilities are most valuable when paired with high-quality translation memories, meaning the system performs best for organizations with established translation workflows. For languages with limited training data, quality may not match that of more widely supported language pairs. Additionally, the learning curve for optimal configuration -- selecting the right translation memories, building glossaries, and fine-tuning project settings -- requires some translation technology expertise, though the documentation and support team provide helpful guidance.