Expert Knowledge & News

Many companies feel secure once they have introduced a "concept-oriented" terminology database. The logic sounds simple: a concept receives a unique ID, and definitions, images, permitted and prohibited synonyms and translations are stored beneath it. But those who believe the problem is solved often face a rude awakening in practice. Misunderstandings in technical documentation and errors in translation memory usually arise precisely here — from too vague a definition of what a "concept" actually encompasses.

The hidden trap: The problem of quasi-synonyms

A false assumption prevails: that all words captured under the same ID are 100% identical in meaning. With the two German words for "screwdriver", this may hold true. But what about "spacer ring" and "spacer washer"? In many databases, they are listed as synonyms, even though they describe geometrically and functionally different components.

The problem becomes even more critical during internationalisation:

• The German "Abluftsystem" becomes either "exhaust air system" (when dealing with pure exhaust gases) or "ventilation system" (when dealing with indoor air) in English.
• Ignoring these linguistic realities and blindly applying 1:1 equivalents provokes serious errors in translation and the subsequent post-editing process.

Practical tips for your terminology management

1. Use semantic component analysis: Break down critical terms into hard features (e.g. function, material, control type). Only when all features match can we speak of true synonyms.
2. Courage to leave a gap (fuzzy equivalence): Accept that languages divide reality differently. If the differences are process-critical for translations, it is better to create two cleanly separated concepts.
3. Enforce technical validation mechanisms: Configure your quality assurance tools (e.g. ErrorSpy) to actively issue warnings to editors for known, context-dependent words.

The search for information within companies has been fundamentally transformed by Large Language Models (LLMs) and modern AI assistants. Where employees once had to spend hours sifting through rigid databases for exact terms, semantic search functions and generative AI now often deliver contextual answers within seconds.

But the new technology brings a hidden, costly challenge: when the underlying corporate terminology is imprecise or ambiguous, even the best AI begins to "hallucinate" or delivers factually incorrect results. The costly search for knowledge may have been replaced by prompting, but the risk of poor decisions remains high when the AI lacks the exact semantic anchor it needs. The solution lies in intelligent terminology that works hand in hand with LLMs.

The new reality: LLMs understand context, but need guardrails

Modern AI models excel at analysing the context of a sentence to resolve ambiguities.

• The word "container" can be a shipping box or a liquid tank in a coffee machine. When the LLM reads "As soon as the container is empty, the brewing process stops", the AI automatically recognises the context and selects the correct translation or explanation.

The problem arises when the context is too small or imprecise — for example, in short software strings, table overviews, bill-of-materials entries, or fragmented support requests. This is where LLMs reach their limits. Structured concept relations (such as "Is_Part_Of" or "Assigned_To") then become the decisive complementary measure, stepping in when the text itself does not provide the AI with enough context to decode the correct meaning.

The paradox: AI optimises AI

The key insight in modern terminology work is that these relational knowledge networks (ontologies) no longer need to be painstakingly assembled by hand over years. We use the strengths of AI to mitigate the weaknesses of AI search. LLMs are themselves the tool for automatically extracting relational links between concepts from existing data and building structured knowledge bases.

Practical tips for your company

1. Use RAG systems with terminology injection: When building internal AI assistants (Retrieval-Augmented Generation), feed the model not only with running text. Supply your validated corporate terminology as a structured knowledge base in the prompt or vector database to dramatically improve AI output quality.
2. Build AI-supported concept relations: Have your existing legacy documents analysed by LLMs to automatically identify semantic relationships between your core concepts (e.g. "Extract all functional dependencies for component X"). This saves up to 80% of manual maintenance effort.
3. Context flanking for short texts: When translating or processing short, ambiguous text segments (such as UI elements), routinely pass the relational metadata of the concept to the AI. This way, the model knows the exact domain and meaning even without a long accompanying text.

Companies often define a highly standardised internal technical language. Yet the market, customers in support forums, or buyers on the customer side speak entirely differently. A "potentiometer" quickly becomes a "dial" and a "flow regulator" becomes a "thingamajig".

Ignoring the language of your target audience means losing touch: products are not found on the web, tenders are lost, and technical support is overwhelmed with avoidable queries. The challenge is to systematically capture the real language of the market and embed it in your own terminology system.

The path to target-group-appropriate language

Simply consulting a standardised synonym dictionary does not help, as it does not reflect actual market usage. You need to gather data where your customers are speaking. A structured approach conserves resources.

Practical tips for implementation

1. Tap internal goldmines: Analyse existing, unstructured data sources. Support tickets, customer correspondence, complaint reports, and your website's search logs tell you exactly which terms customers use to search for your solutions.
2. Use the bootstrapping method: Enter core terms in combination in search tools (e.g. "transformer – winding – magnetic field"). The search contexts automatically deliver the real market synonyms straight to you.
3. Control via usage attributes: Link customer expressions as permitted synonyms with the official term, but attach strict attributes (e.g. "Permitted in SEO/Marketing, prohibited in technical documentation").

In the start-up phase of a terminology project, everything usually runs to plan: the rules for spelling are fresh, the participants highly motivated. But over the years, as the data stock grows and new departments become involved, it inevitably happens: the internal consistency of the database erodes. It is like a forest — with so many individual trees, no one can see the overall state of the woodland any more.

Unrecognised duplicates, contradictory attributes, and outdated designators creep in. Since content in authoring systems and translation memories is now reused in large quantities automatically, these errors multiply rapidly and drive up correction costs in post-editing enormously.

Typical pitfalls in the system

• Hidden synonyms: While spelling variants are easy to spot, corporate jargon is harder. Does your system or a new employee know that "first-aid kit" and "medical bag" mean exactly the same thing within the company?
• Contradictory metadata: Importing new lists often leads to logical errors. A term must not simultaneously have the status "unreviewed" while its usage is set to "permitted".

Practical tips for sustainable database maintenance

1. Sort by word endings: Export your term lists and sort them alphabetically by word ending. This brings conceptually identical compounds directly alongside each other for quick clean-up.
2. Strict validation of status attributes: Carry out regular logic checks. Consolidate spelling variants in the metadata (e.g. unify "forbidden", "Forbidden", "deprecated" into a single system value).
3. Set up role-based views: Use a terminology management system (such as LookUp) that allows target-group-specific filters. A translator needs different metadata than a sales employee.