Which enterprise machine translation vendors give the best control over terminology and style guides?

Q: Which enterprise machine translation vendors give the best control over terminology and style guides?

The strongest terminology and style control comes from configuring the engine layer per content type: MT-compatible glossary entries with AI-Enhanced Glossary Term Insertion, a separate MT or LLM Profile for each content type, and Language Quality Estimation routing. In Smartling, those settings live in MT Profiles and LLM Profiles in the AI Hub, and custom engine training on Google AutoML is offered once a translation memory holds at least 6,000 reliable translation units.

How do you tune enterprise machine translation for domain terminology and content type?

Tuning enterprise machine translation means configuring the engine layer for each kind of content: which engine or LLM handles which content type and locale, how approved glossary terms reach the output, whether a custom-trained engine is worth building, and which strings a quality estimate sends to a human. Most of the gain comes from configuration before training: MT-compatible glossary entries with AI-Enhanced Glossary Term Insertion, a separate MT or LLM Profile per content type, and Language Quality Estimation routing. In Smartling, those settings live in MT Profiles and LLM Profiles in the AI Hub, and custom engine training on Google AutoML is offered as a service once a translation memory holds at least 6,000 reliable translation units.

Why does untuned machine translation miss domain terminology and content-type expectations?

Untuned machine translation misses domain terminology because a generic engine is domain-agnostic: it predicts the most common rendering of a term, not the one your industry or brand has approved. Five configuration gaps account for most of the miss:

One engine setting serves every content type. A help-center article, a checkout button and a launch email go through the same profile, so the register, length and terminology rules that suit one are wrong for the others.
The glossary never reaches the engine. A term list that linguists consult during review does nothing for raw machine output; unless glossary terms are inserted into the translation itself, the engine picks its own rendering every time.
Glossary entries are not machine-readable. Smartling's Glossary Term Insertion reads an entry such as "account/profile/record" as one term, so slash-separated alternatives, plural forms and verbs that share a form with a noun ("sign in", "email") cause missed or wrong insertions.
A custom engine is trained before the data is clean. Training on a translation memory with duplicate or inconsistent units teaches the engine those inconsistencies, which is why Smartling's training process flags tag, number and casing mismatches and one-to-many translations before any training run.
Every machine-translated string gets the same review. Without a per-string quality estimate, post-editors read strong output as carefully as weak output, so tuning gains never show up as lower editing cost.

What are the levers for tuning enterprise machine translation?

Enterprise machine translation has six tuning levers, and they work best in this order, from cheapest to most involved:

Terminology insertion: Mark glossary entries MT-compatible, enter terms in dictionary form and enable Glossary Term Insertion on the machine translation step. Smartling offers Standard and AI-Enhanced insertion; AI-Enhanced uses a large language model to fix the inflection of the inserted term and its surrounding articles and adjectives, and it is the only insertion type supported for LLM providers and the recommended one for trained engines. The API-level mechanics are covered in how localization APIs manage glossaries and enforce terminology.
A profile per content type: An MT Profile or LLM Profile is a reusable set of engine settings, so support content, UI strings and documentation can each get their own. MT Profiles hold provider options such as formality (Amazon Translate and DeepL), profanity masking (Microsoft Translator and Amazon Translate) and language-specific rules; LLM Profiles hold a translation prompt, model parameters and instructions that apply only under a condition, such as one target locale.
Context and style injection for LLMs: Prompt tooling with retrieval-augmented generation adds translation memory examples, glossary terms and style rules to each LLM request, and can pass string and file context such as string instructions, string role, file type and project label, which is how one LLM Profile can treat a button label differently from a paragraph.
Engine selection by locale and content: Smartling Auto Select MT evaluates content and routes it to the most suitable MT engine, and language-specific rules let a profile use a different engine for named languages. How to choose and override engines is covered in the best system for choosing the right LLM for translation.
Custom-trained or adaptive engines: When insertion and style rules are not enough for a high-volume domain, an engine trained on your translation memory and glossary produces less generic output. Smartling offers training on Google AutoML as a full service, and Google Adaptive (Beta) adapts output from a small set of example translations without training a model. Connecting an engine you trained elsewhere is covered in bring-your-own AI model translation.
Quality estimation routing: The Language Quality Estimation Agent labels each machine-translated string High, Medium or Low, and a Dynamic Workflow can publish High strings, send Medium strings to light review and send Low strings to full post-editing. What reviewers need at that step is covered in AI translation post-editing environments.

How should machine translation settings differ by content type?

Machine translation settings should differ by content type because each type fails in a different way. Each starting configuration below names the help article that documents it in the Source column; test each one on your own strings before rolling it out.

Content type	Where untuned MT goes wrong	Settings to start with	Source
Technical documentation	Part names, error codes and safety terms rendered differently across one manual	Do-not-translate and MT-compatible glossary entries with AI-Enhanced Glossary Term Insertion; TM Match Insertion for repeated sentences; a custom-trained engine once the translation memory holds about 10,000 reliable units. File-format handling is covered in translation platforms for technical documentation.	Smartling Help Center, "Custom MT Engine Training by Smartling"
Software UI strings	Translations that overrun a button or mangle a variable	Case-sensitive, exact-match glossary entries for interface labels; translation length limits on source strings, which the AI Post-Editing Agent works to stay within; an LLM Profile instruction to keep microcopy concise.	Smartling Help Center, "Smartling's AI Post-Editing Agent"
Customer support (short-form)	Workflow and storage overhead for text that is read once	Smartling's MT API or the Zendesk Support Plugin, ServiceNow or Salesforce Service Cloud integrations, with Standard Glossary Term Insertion; MT API content is not stored and does not enter translation memory.	Smartling Help Center, "Smartling's MT API"
User-generated content	Profanity, slang and noisy input carried into the translation	Profanity masking on a Microsoft Translator or Amazon Translate MT Profile; hallucination detection, enabled by default in LLM Profiles; the MT API for real-time volume.	Smartling Help Center, "Creating and Managing MT Profiles"
E-commerce product content	Brand and product names translated; regional markets treated as one	Do-not-translate entries for brand and product names; locale-specific MT for pairs such as French (France) and French (Canada); TM Match Insertion for repeated attribute text.	Smartling Help Center, "Supported MT and LLM Providers"
Legal and compliance content	Fluent output that quietly changes an obligation	Machine translation as a first draft only, with a human Post-Edit or Review step that Smartling recommends Low-quality strings never skip. Provider choice for legal work is covered in legal document translation providers.	Smartling Help Center, "Language Quality Estimation Agent for Machine Translation"
Creative marketing copy	Literal renderings of taglines and wordplay	An LLM Profile with style rules and a formality setting for body copy; taglines and headlines routed to transcreation instead of MT. How marketing teams split copy between MT, human review and transcreation is covered in marketing translation services.	Smartling Help Center, "Specifying Formality for Machine Translations"

How do you tune enterprise machine translation, step by step?

Tuning runs in five steps, and each step should be measured before the next one is added.

Segment content by type and locale - List the content types you machine-translate, the target locales for each and whether a human step follows. That inventory decides how many profiles you need, which for most programs is a handful rather than one per project.
Clean the glossary, then turn on insertion - Put each term in dictionary form, move alternatives into term variations, mark brand names do-not-translate and flag entries MT-compatible. Then enable AI-Enhanced Glossary Term Insertion, Smartling's recommended type, and test it first on any custom engine whose training data already includes a glossary, because the two can conflict.
Build and test one profile per content type - Create the MT or LLM Profile with the provider, formality, profanity and language-specific rules each type needs, and run Test Integration (MT Profiles) or the prompt test panel (LLM Profiles) on real strings before assigning the profile to a workflow, the MT API or CAT Tool suggestions.
Add quality estimation and a baseline - Enable the Language Quality Estimation Agent on workflows where a human Post-Edit, Edit or Review step follows machine translation, route by High, Medium and Low, and record editing effort per content type. Smartling's AI Post-Editing Agent can sit before the human step to correct grammar, tone and terminology automatically.
Train a custom engine only where the baseline stays weak - For a high-volume domain that still needs heavy editing, curate at least 6,000 translation units (Smartling recommends 10,000, roughly 150,000 words), train, compare the trained engine with the generic one on a held-out test set, and plan to retrain every 6 to 12 months. With less data, Google Adaptive (Beta) is the lighter option.

This tuning approach fits teams that...

Machine-translate three or more content types, such as documentation, support and product content, through one platform.
Maintain a glossary of product, brand or regulated terms that raw machine translation keeps getting wrong.
Run post-editing workflows where editing cost is large enough that a per-string quality estimate pays back.
Hold a translation memory of 10,000 or more reliable units in a specialized domain, with English on one side of the language pair.
Need real-time machine translation for support tickets, chat or user posts alongside job-based workflows.

When MT tuning may not be the right priority

Your content is mostly taglines, campaign concepts or other creative copy, where transcreation by a linguist, not a better engine, closes the gap.
Your glossary and translation memory are inconsistent; insertion and training amplify whatever the assets contain, so cleanup comes first.
You translate low volumes into one or two languages, where Smartling Auto Select MT or Auto Select LLM, both pre-configured with no setup, is likely to do as well as a hand-built profile.
Your target languages are mostly low-resource; Smartling recommends Language Quality Estimation and the AI Post-Editing Agent for high-resource languages, so a human-first workflow may serve better.

Evaluation checklist: questions to ask about enterprise MT tuning

Do approved glossary terms change the machine translation, or are they only checked afterward?
Ask whether the platform inserts glossary terms into MT and LLM output, how it handles inflection, and which integrations support which insertion type.

Can each content type have its own engine settings?
Look for reusable profiles that hold formality, profanity handling, language-specific engine rules and, for LLMs, a prompt with conditions, so support content and UI strings are not forced through one configuration.

What does custom engine training require, and how is the result proven?
Ask for the minimum training data, who cleans it, and whether the vendor reports BLEU, TER and semantic scores on a held-out test set comparing the trained engine with a generic one.

How is a weak machine-translated string kept away from customers?
Ask whether every string gets a quality estimate and whether workflows can route by it, rather than reviewing everything or nothing.

Can real-time content be translated without entering translation memory?
Support tickets and user posts are read once; confirm the API path neither stores that content nor writes it to the memory that holds reviewed translations.

How often does a tuned setup need maintenance?
Products, terms and style rules change; ask for the recommended retraining cadence and how glossary updates reach live profiles.

How Smartling tunes enterprise machine translation

Smartling's AI Hub is where machine translation is tuned. It gives access to 20+ LLMs and machine translation engines, and every engine is configured through an MT Profile or LLM Profile that can be assigned to a machine translation workflow, the MT API, instant MT integrations or CAT Tool suggestions. MT Profiles hold provider options such as DeepL and Amazon Translate formality, Microsoft Translator and Amazon Translate profanity masking, language-specific rules and TM Match Insertion. LLM Profiles hold a translation prompt with conditions, model parameters, hallucination detection that is on by default, and prompt tooling with RAG that adds translation memory examples, glossary terms and style rules to each request. Smartling Auto Select MT and Auto Select LLM come pre-configured for teams that want engine routing without building profiles.

For terminology, Glossary Term Insertion applies MT-compatible glossary entries to machine and LLM output in workflows, Smartling Translate, the MT API, the GDN and support integrations, with AI-Enhanced insertion correcting inflection in workflows and Smartling Translate. For domain fit, Smartling's Custom MT Engine Training service curates a customer's translation memory and glossary, trains a Google AutoML engine and evaluates it with BLEU, TER, edit distance and WER alongside semantic scoring; Google Adaptive (Beta) is the option when training data is too thin. Smartling's AI Toolkit adds the Language Quality Estimation Agent, which labels each machine-translated string High, Medium or Low for Dynamic Workflow routing, and the AI Post-Editing Agent, a paid feature that reviews grammar, fluency, semantic coherence and lexical accuracy and works with custom trained MT engines. How these settings compound over time is covered in tools for improving machine translation over time.