Severity Levels and Domain-Specific MQM

ai-machine-human # The Four Severity Levels

Q. How does MQM define severity? What are the four levels?

A. MQM uses four severity levels. The key principle is that severity reflects actual impact on the end user, not how 'obvious' the error looks.

Severity	Definition	Typical Triggers
Critical	Can cause extreme consequences — safety, legal, health, financial harm; damages brand reputation; corrupts product function; contains offensive content; or is designated critical by the project	Safety warning mistranslated; offensive expression generated; wrong instructions cause app crash; high-visibility UI element (button, menu) error
Major	Misleads or confuses users; prevents normal product/service use; meaning significantly altered or tone seriously inappropriate; error appears in high-visibility location	Accuracy error likely to mislead users; heading, TOC, or chapter title error; confirmed QA feedback not applied; GUI–document terminology mismatch
Minor	Does not affect comprehension but reduces style quality, fluency, clarity, or polish. Noticeable but unlikely to confuse the user.	Marginal meaning shift with no user impact; glossary deviation without confusion; scattered punctuation/capitalization errors; slightly stiff phrasing
Neutral	Not counted as an error. Does not affect score or pass/fail outcome.	No Error; Preferential Change (stylistic suggestion); Repeated Error (already counted); Exceptional Solution (positive annotation); Source Error

Q. What counts as a Neutral finding? I thought everything was either an error or not.

A. Neutral is a formal designation for five specific situations where no penalty is applied:

Neutral Tag	When to Use It
No Error	The segment has no errors. Used primarily in segment-level MT evaluation.
Preferential Change	The existing translation is not wrong, but the evaluator suggests an improvement. Incorporation is at the team's discretion.
Repeated Error	An error already counted elsewhere recurs. Record it but do not double-penalize.
Exceptional Solution	Positive annotation: the MT output elegantly solved a linguistic challenge — idiom, wordplay, cultural adaptation — at a level comparable to a skilled human translator.
Source Error	The problem originates in the source text, not the translation. Record for source correction; no score impact.

MQM Is Not One-Size-Fits-All: Domain Adaptation

Q. Does the same MQM framework apply to every domain? I work across gaming, legal, and medical.

A. This is one of MQM's most important — and most frequently misunderstood — features. MQM is not a fixed absolute standard. It is a flexible framework that can be configured per domain by:

Adjustment Type	Example
Promoting a category's severity threshold	In legal translation, even a Minor Accuracy error may be elevated to Major
Downgrading a category's importance	In gaming, Locale Convention formatting errors may be Minor or excluded
Excluding categories not applicable to the domain	Source Error annotations excluded from a marketing copy evaluation
Adding domain-specific sub-types	Medical translation may add a sub-type for drug interaction warnings

Let's walk through four major domains and see how MQM looks in each.

Domain 1: Game Localization

Gaming prioritizes player experience (UX) and cultural resonance. The question isn't just 'Is this accurate?' but 'Does this feel native to a player in this market?'

Category	Priority in Gaming	Notes
Fluency / Naturalness	Very High — Major	Stiff dialogue breaks immersion immediately
Tone / Register	Very High — Major to Critical	Each character must have a distinct, consistent voice
Cultural Adaptation	High — Major	Western memes, humor, cultural refs must be localized
Tags & Placeholders	Critical	Translating {hero_name} will break the game engine
Glossary Compliance	High	Item names and skill names must match approved term lists
Numeric Formats	Low / Excluded	Game numbers usually kept as-is per project specs

Example errors in gaming:

Source (EN)	AI-MT Output	Error & Severity
"The {hero_name} has slain the dragon!"	"The hero has slain the dragon!" (variable dropped)	Tags & Placeholders / Critical
"Use the Flaming Sword of Doom"	"Use the Sword of Flame and Doom" (item name changed)	Glossary Compliance / Major

Domain 2: Legal Translation

Legal translation operates under zero-tolerance accuracy requirements. A single word can change the meaning of a contract, trigger a dispute, or create liability. MQM severity thresholds are significantly elevated across the board.

Category	Priority in Legal	Notes
Mistranslation	Critical (expanded scope)	Any meaning distortion in legal clauses is Critical
Omission	Critical	Missing clauses, conditions, or obligations
Numbers & Symbols	Critical	Wrong amounts, dates, or periods
Glossary Compliance	Major to Critical	Legal terms carry precise judicial interpretations
Fluency / Naturalness	Minor or excluded	Accuracy takes absolute priority over style
Exceptional Solution	Rarely used	Creative translation discouraged; fidelity required

Example errors in legal translation:

Source (EN)	AI-MT Output	Error & Severity
"The contract shall be void and of no effect."	"The contract may become invalid." (shall mistranslated as may)	Mistranslation / Critical — modal verb changes binding nature
"within 30 days of execution"	"within 3 days of execution" (digit dropped)	Numbers & Symbols / Critical
"PRN medication"	"Regular medication" (PRN = as needed)	Industry-Standard Terminology / Critical

Domain 3: Medical / Healthcare Translation

Medical translation combines the precision requirements of legal translation with direct patient safety implications. The margin for error is extremely narrow.

Category	Priority in Medical	Notes
Mistranslation	Critical (broad scope)	Dosage, symptom, or procedure errors endanger patients
Numbers & Symbols	Critical	Unit confusion (g vs mg) can cause fatal dosing errors
Industry-Standard Terminology	Critical to Major	ICD codes, drug generic names, WHO-standard terminology required
Cultural Adaptation	High	Healthcare systems and norms differ across markets
Addition (Hallucination)	Critical	Fabricated medical information is life-threatening
Fluency	Minor	Patient comprehension matters, but accuracy comes first

Example errors in medical translation:

Source (EN)	AI-MT Output	Error & Severity
"Do not exceed 4g per day."	"Do not exceed 4mg per day." (g confused with mg)	Numbers / Critical — 1000x dose difference
"PRN medication"	"Take medication regularly." (meaning reversed)	Mistranslation / Critical
"Contraindicated in patients with renal impairment"	"Use with caution in patients with kidney problems." (contraindicated softened)	Mistranslation / Critical

Domain 4: Marketing & Creative Localization

Marketing localization moves in the opposite direction from legal and medical. Here, persuasiveness, emotional resonance, and brand voice take precedence. MQM is configured to reflect that shift.

Category	Priority in Marketing	Notes
Fluency / Naturalness	Major — elevated from Minor	Stiff copy kills conversions; unnatural = brand damage
Tone / Register	Very High — Major	Brand voice consistency is non-negotiable
Cultural Adaptation	High — near Critical	Humor, idioms, cultural references must land locally
Omission / 1.3 Addition	Relaxed	Transcreation may legitimately expand or compress source
Capitalization	Low / excluded	Brand stylistic conventions override standard rules
Glossary Compliance	High	Brand terminology and product names must be consistent

Example errors in marketing translation:

Source (EN)	AI-MT Output	Error & Severity
"Just Do It."	"Please proceed with the action." (brand impact lost)	Fluency + Tone / Major
"Our product gives you wings."	"This product provides flight capability." (metaphor taken literally)	Mistranslation + Fluency / Major
"Limited time offer — grab yours now!"	"A time-limited promotional opportunity is available." (urgency and energy gone)	Tone + Fluency / Major

Domain Priority Matrix

Q. Can you summarize how category priority shifts across the four domains?

Category	Gaming	Legal	Medical	Marketing
Accuracy	★★★	★★★★★	★★★★★	★★★
Terminology	★★★	★★★★★	★★★★★	★★★
Linguistic Conv.	★★	★★★★	★★★★	★★
Style / Fluency	★★★★★	★★	★★	★★★★★
Consistency	★★★	★★★★	★★★★	★★★★
Locale Conv.	★★	★★★★	★★★★	★★★★
Format / Tags	★★★★★	★★★	★★★	★★