Deutsch
Français
EnglishWhat are translation memories (TMs) and termbases (TBs)?
A translation memory (TM) is a database that stores sentences or segments of text in one language alongside their translated equivalents in another language. In practice, whenever a translator works on new content, any sentence that has been translated before can be retrieved from the TM so it doesn’t have to be translated from scratch again. A termbase (TB), on the other hand, is an organized glossary of terminology – it contains important terms or phrases and their approved translations. Termbases help ensure that everyone uses the same consistent translations for key terms (like product names, legal phrases, or industry jargon).
Translation memories and termbases are fundamental tools for human translators. They improve consistency and efficiency by reusing past translations and enforcing standard terminology. Over time, a company builds up large TMs and TBs that reflect its preferred wording and style in multiple languages. These resources have long been valuable for speeding up translation projects and maintaining quality – and now they are becoming strategic assets for powering AI systems as well.
Custom multilingual AI models: are they affordable?
Not long ago, training a large language model from scratch was prohibitively expensive. Early advanced models (like the original GPT-3 in 2020) were estimated to cost several million dollars in computing power to train, often taking months on specialized hardware. Only tech giants with deep pockets could undertake such projects, while other organizations had to rely on generic pre-trained models.
Today, a clear paradigm shift is underway. Major cloud providers such as Amazon Web Services (AWS), Microsoft Azure, OpenAI, and Anthropic now offer mechanisms to fine-tune pre-trained language models at a fraction of previous costs. Rather than building an AI model from scratch, organizations can further train existing large models using their own data. For example, in December 2025, Amazon announced new features within its Bedrock and SageMaker AI platforms to simplify model customization, including reinforcement fine-tuning workflows capable of significantly improving model accuracy. These tools, together with comparable offerings from Azure and OpenAI, substantially reduce the cost and complexity associated with developing custom AI applications. Microsoft’s Azure Machine Learning service, for instance, provides a platform that supports language model fine-tuning and deployment without requiring extensive infrastructure investments. In summary, fine-tuning large language models on proprietary data is no longer a niche, multimillion-dollar endeavour, but is increasingly becoming a standard business practice.
For translation agency clients, this shift means that even medium-sized organizations can now consider training or refining AI models that are aligned with specific domains and languages. Rather than relying on a one-size-fits-all solution, an organization can deploy an AI model that reflects its terminology, product information, and communication style. Fine-tuning a model using organization-specific bilingual data, such as translation memories (TMs) and termbases (TBs), can significantly enhance AI performance in tasks such as multilingual customer correspondence or support interactions. In essence, an AI model tailored to organizational content is able to communicate more naturally and accurately, having been trained on the organization’s own linguistic assets.
Why is structured language data critical for AI training?
As AI model customization becomes more accessible, the fuel for these models is increasingly the organization’s own data. For language-focused AI, multilingual content like translation memories and termbases becomes critical training material. A translation memory is essentially a repository of paired sentences in different languages (source and target), and a termbase is a database of approved translations for important terms and phrases. Together, these resources represent a company’s accumulated linguistic knowledge – everything the company has already translated, along with how it prefers to phrase things.
This bilingual content is highly valuable for AI training. Fine-tuning a model using existing translation memories (TMs) and termbases (TBs) infuses the AI with organization-specific vocabulary and writing style across all supported languages. Studies have shown that fine-tuning large language models with in-house translation memory data significantly improves the use of correct domain-specific terminology and the ability to reproduce the desired style, resulting in higher-quality translations and responses. In one documented case, leveraging an organization’s own TM enabled a custom model to process highly specialized texts with appropriate jargon and nuanced phrasing far more effectively than a generic out-of-the-box model. In essence, the AI learns from past human translations and becomes increasingly adept at recognizing preferred terminology and phrasing established over time.
This trend also elevates the role of the translation provider. Translation agencies such as TTN are no longer limited to delivering translated documents, but increasingly act as custodians of multilingual knowledge assets. Rather than allowing translation archives to remain static, a provider such as TTN can maintain and curate linguistic databases in a form suitable for AI use. Many language service providers are already pursuing this approach, and TTN is well positioned to do the same. By leveraging large volumes of human-validated translations accumulated over time, a translation agency can support the development of custom AI models tailored to specific domains, thereby improving translation quality and maximizing the long-term value of translation data. In this context, past translations are not merely archived documents, but constitute the foundation for future multilingual AI systems capable of reflecting organizational terminology, style, and subject-matter expertise with a high degree of precision.
What is retrieval-augmented generation (RAG)?
One of the more recent developments in AI is retrieval-augmented generation (RAG). Even a fine-tuned language model can encounter limitations when handling very recent information or answering detailed, fact-based queries. RAG addresses this by connecting the AI model to an external knowledge source, such as a document repository or knowledge base, at the time an answer is generated. In this approach, the model retrieves relevant information from external sources in real time and incorporates it into the response generation process. The principal benefit of this method is known as grounding: AI responses are supported by concrete, verifiable information drawn from authoritative data sources rather than relying solely on knowledge acquired during training. Grounding responses in reference material significantly reduces the risk of hallucinations, as the model is guided by validated facts or approved content when formulating outputs.
Clean translation memories (TMs) and termbases (TBs) can form part of such external knowledge sources for AI systems. In multilingual chatbot or virtual assistant scenarios, the system can consult repositories of previously translated content or approved terminology when uncertainty arises regarding phrasing, translation, or factual accuracy. This represents a form of retrieval-augmented generation in which the AI not only produces responses, but also incorporates contextual information from relevant documents, such as curated bilingual resources. Over time, the model learns when and how to access these references, resulting in more reliable grounding in organization-approved information. The outcome is an AI assistant that consistently applies correct product names, legal notices, and technical terminology across languages by retrieving validated content from maintained translation memories and termbases whenever appropriate.
Why does translation data need to be clean for AI?
All AI-related advantages depend on the quality and maintenance of the underlying data. If a translation memory contains duplicate entries, outdated translations, or inconsistent wording, these deficiencies may be learned and reproduced by the AI model. Training or fine-tuning an AI system on poorly curated data can result in the propagation of errors or the generation of unreliable outputs. Similarly, if a termbase includes ambiguous or incorrect terminology, an AI-based assistant relying on that data may produce inaccurate or misleading information. In essence, the principle of “garbage in, garbage out” applies: AI performance is directly determined by data quality.
For this reason, sustained investment in clean and well-maintained translation memories (TMs) and termbases (TBs) is essential to ensure long-term AI readiness. Although data curation may appear operational rather than strategic, it has become a fundamental prerequisite for effective AI deployment. Inaccurate or obsolete data undermines trust in AI outputs, whereas consistently curated and up-to-date linguistic resources enable AI systems to deliver precise, reliable results. Multilingual content can be viewed as the training foundation for AI systems: high-quality, balanced data supports robust model behaviour, while inconsistent data leads to degraded performance.
For translation agencies and their clients, maintaining clean linguistic data entails several concrete operational practices. These include updating translation memories exclusively with validated and proofread translations, rejecting entries that do not meet quality standards, identifying and removing duplicate segments, and retiring obsolete content such as outdated slogans or superseded product descriptions. It also involves continuously expanding and maintaining termbases with clearly approved terminology and, where necessary, usage notes. Through these measures, training datasets used for AI fine-tuning or multilingual chatbot deployment remain accurate, consistent, and aligned with current linguistic standards. As a result, AI systems can reproduce the same level of quality, terminology consistency, and stylistic coherence that has been established through human translation processes.
Why are TMs and termbases an investment in the future of AI?
Looking ahead, the boundary between translation services and AI services is increasingly blurred. Translation providers are becoming key partners in the development of multilingual AI solutions, as they manage the linguistic data that underpins effective AI systems. Collaboration with a translation agency not only results in translated content, but also contributes to the continuous enrichment of bilingual knowledge repositories that can be leveraged to train future AI-based customer service agents or internal knowledge assistants. As cloud platforms such as AWS and Azure continue to expand their capabilities for custom AI development, and as providers like OpenAI further lower the barriers to model fine-tuning, competitive advantage will increasingly favour organizations with rich, clean, and domain-specific linguistic data readily available. Organizations maintaining well-structured translation memories (TMs) and termbases (TBs) are therefore best positioned to develop AI models that accurately reflect their terminology and content.
In this context, maintaining clean TMs and TBs extends beyond translation efficiency and represents a strategic investment in organizational AI readiness. A well-maintained translation memory effectively constitutes a parallel corpus of multilingual communication accumulated over time, providing a valuable foundation for training custom multilingual AI models. Similarly, a termbase functions as an authoritative multilingual reference of approved terminology, ensuring consistent and accurate term usage in AI-generated content, including product names and regulatory language. Treating these linguistic assets as strategic resources ensures that future AI deployments are grounded in reliable and validated data.
From a leadership perspective, this evolution highlights the long-term value of translation and localization activities. Multilingual content produced today directly supports future AI initiatives, reducing the need to build AI knowledge bases from scratch. Many organizations already possess extensive repositories of validated translations and terminology that can be reused for AI development, provided that these assets are properly maintained. As with financial data management supporting accurate business intelligence, disciplined linguistic data management is essential to ensure the reliability and accuracy of AI language models.
Ultimately, the role of translation memories and termbases extends well beyond supporting human translation workflows. These resources are becoming foundational components of multilingual AI capabilities. By maintaining comprehensive, consistent, and up-to-date linguistic data, organizations establish the conditions necessary for AI systems to understand content accurately and communicate effectively across languages. As AI increasingly supports customer interaction and content creation, a robust multilingual data foundation will represent a critical differentiator. Translation agencies, with their expertise in managing and curating linguistic resources, serve as natural partners in this transformation. Sustained investment in clean TMs and TBs today is therefore a key step in preparing for an AI-driven future, in which high-quality multilingual data underpins superior communication across all languages.