DQF vs MQM: Which Translation Quality Model for Regulated US Content?
Direct answer. DQF and MQM are not competing standards. TAUS DQF is a subset of MQM, and that subset was developed further into MQM Core, which is now used in ISO 5060 and ASTM WK 46396. A US localization team should adopt MQM Core, add error types from the full MQM typology for high-risk documents, and record severity, scoring and root cause in an audit trail.
Table of Contents
What are DQF and MQM, and what is their scope?
MQM stands for Multidimensional Quality Metrics. The EU-funded QTLaunchPad project developed it as a flexible, analytic system for evaluating translation quality. It uses a hierarchical error typology. Users select the issue types that suit a specific task from a master catalog.
DQF stands for Dynamic Quality Framework. TAUS developed DQF in consultation with its members. It includes tools for evaluating translation quality and an error typology.
The scope of both is content quality. MQM cannot directly measure process aspects such as delivery timeliness or evaluator professionalism. Those belong to project management and vendor management controls.
MQM is a manual, reference-free framework. It is distinct from automatic metrics such as BLEU, which require reference translations. A human evaluator reads the output and classifies each issue.
How are DQF and MQM related today?
The two were harmonized. TAUS and DFKI, working with LTAC Global, created DQF:MQM, a harmonized error typology based on the first versions of MQM and DQF. The MQM Council now maintains the framework and works to keep industry standards consistent with the MQM typology.
MQM Core replaces the earlier MQM/DQF subset. According to the MQM Council typology page, it is a pre-established, widely used subset of error types at the two highest hierarchical levels. It can be used as a default.
The practical consequence is simple. A request for “DQF” in a vendor contract usually means the MQM Core subset. A team that still references DQF should state the typology version in its quality specification so that scores stay comparable over time.
How do the error categories, customization and evaluation methods differ?
Criterion | DQF (TAUS) | MQM (full typology) | MQM Core |
Origin | TAUS member consultation | QTLaunchPad, now MQM Council | Subset of MQM, successor to the DQF subset |
Error catalog | Defined error typology | Master catalog, user-selected types | Seven high-level dimensions with common subtypes |
Customization | Limited to the subset | Modular, select only needed types | Fixed default, extendable from full MQM |
Root cause analysis | Not part of the subset | Supported in MQM 2.0 | Supported through MQM 2.0 practice |
Current status | Folded into MQM | Maintained by MQM Council | Referenced in ISO 5060 and ASTM WK 46396 |
The seven high-level dimensions of the current typology are Terminology, Accuracy, Linguistic conventions, Style, Locale conventions, Audience appropriateness, and Design and markup.
MQM is designed to be modular. The MQM committee states that MQM is not meant to be used in its entirety. Selecting issue types is the core of a practical implementation.
MQM also provides a shared vocabulary for describing issues. Metrics can be compared where they overlap. The framework added a Verity dimension, which recognizes that an accurate translation may still be wrong for its intended context.
How are errors scored for severity and quality?
Under MQM, evaluators classify errors by type and severity. Weighted penalty points then produce a numerical quality score.
A scoring workflow has these elements:
Evaluation sample: a defined word count drawn from the deliverable.
Error annotation: each issue gets a type, a location and a severity.
Penalty weights: each severity level carries a weight set in the quality specification.
Pass threshold: a score that the content type must meet.
Reporting: the score and the error profile are stored with the job record.
The specific weights and thresholds are a policy decision for the buyer. The framework does not fix them for every content type. Record them in writing before the first evaluation.
MQM 2.0 adds root cause processing. Evaluators categorize errors by cause, for example pre-translation or production, to support continuous process improvement.
How do the frameworks apply to human, machine and post-edited output?
MQM is designed to evaluate human translation, post-edited machine translation and unedited machine translation on an equal, unbiased basis. The same typology applies in each case. A team can compare an AI+HUMAN hybrid workflow against a fully human workflow using one error profile.
MQM can also evaluate source texts. It diagnoses source problems and their effect on the translated text. For regulated documents this matters, because ambiguous source wording is a frequent origin of Accuracy and Terminology errors.
Automatic metrics may be used for triage. They do not replace a human MQM evaluation when a document carries regulatory or legal weight.
Which steps should a US localization team follow, and who owns each control?
The steps below assume a regulated document, such as a medical device instruction for use, a clinical document, a contract or a financial disclosure.
Classify the content by risk. Owner: Regulatory Affairs or Compliance. Control: a documented risk tier for each document type.
Select the typology. Owner: Localization Manager. Control: MQM Core as the default, with added types from the full typology for Tier 1 content. The MQM guidance recommends matching the subset to the risk level and quality requirements of the content.
Set severity weights and pass thresholds. Owner: Localization Manager with Compliance sign-off. Control: a written quality specification under change control.
Fix terminology before translation. Owner: Terminology lead. Control: an approved term base and translation memory, with access control on edits.
Translate and review. Owner: language vendor. Control: subject-matter expert review and QA against the agreed standard.
Run the MQM evaluation. Owner: an independent evaluator, not the translator. Control: sampled annotation with severity and error type.
Analyze root causes. Owner: Localization Manager and vendor. Control: root cause categories logged per error.
Store the evidence. Owner: Quality or Compliance. Control: an audit trail linking scores, annotations, reviewer identities and corrective actions to the document version.
What decision criteria separate the options?
Use these criteria to decide how much of MQM to adopt.
Risk of the content. Patient safety, legal liability and financial reporting content justify the full typology for added granularity. Marketing and support content can use MQM Core.
Diagnostic need. If the team needs to trace errors to process stages, enable root cause categories from MQM 2.0.
Output type. If the workflow mixes human, post-edited and machine output, MQM allows one scoring basis for all three.
Standards alignment. If the buyer needs to point to a recognized reference, MQM Core is used in ISO 5060 and ASTM WK 46396.
Legacy data. If historical scores use DQF categories, map them to MQM Core and document the mapping.
Data category. Protected health information and personal data require the evaluation environment to meet the same access control rules as the translation environment. Check the applicable US rules, such as HIPAA, and any EU GDPR duties for EU-origin data, with counsel.
The MQM committee recommends that most implementers use the DQF subset and draw on more types from the full typology when needed. Since that subset is now MQM Core, the recommendation translates to a Core default with targeted additions.
What do examples look like in practice?
Medical device labeling. A US manufacturer sells in the EU and translates an instruction for use. The team uses MQM Core plus extra Accuracy and Terminology subtypes. A dosage unit error is annotated as Accuracy with the highest severity. The document fails regardless of the overall score if a critical error is present. This rule must be written in the quality specification.
Financial disclosure. A bank localizes a prospectus summary for Spanish-speaking US customers. Locale conventions matter here: number formats, currency and date formats. The evaluator uses Locale conventions and Audience appropriateness types. The sample covers every section with figures.
Software user interface. A software company localizes strings. Design and markup errors, such as broken placeholders, carry the most weight. The team uses MQM Core and checks automated QA output first.
Post-edited machine output. A legal team compares two workflows on the same contract set. It annotates both with one typology. The error profiles show whether post-editing changes the share of Accuracy errors. The result feeds a vendor decision.
What failure modes occur, and how are they mitigated?
Failure mode | Effect | Mitigation control |
Typology version not recorded | Scores cannot be compared across periods | Version the quality specification |
Evaluator is also the translator | Biased scoring | Separate roles, independent evaluator |
Using the full typology without selection | Slow, inconsistent annotation | Select issue types per content tier |
Score used as the only gate | Critical errors pass on a high average | Add a critical error fail rule |
Expecting MQM to measure delivery or process | Gaps in vendor oversight | Use separate service level controls for timeliness |
Evaluation files shared without access control | Exposure of regulated data | Restrict access, log it, keep files in approved systems |
No root cause data | Same errors recur | Categorize causes and track corrective actions |
Treating machine translation metrics as MQM | False confidence | Use human MQM evaluation for regulated content |
The MQM Council notes that error and root cause data provide the evidence needed to suggest and implement process improvements, even though MQM evaluates the product.
Where AD VERBUM fits
AD VERBUM provides AI+HUMAN hybrid translation for regulated sectors. It operates a proprietary LangOps System hosted on EU servers and works with a network of 3,500+ subject-matter expert linguists. It holds ISO 17100, ISO 18587, ISO 13485, ISO 27001 and ISO 42001 certifications, independently audited by Bureau Veritas. It integrates client translation memories and term bases. Buyers can specify MQM Core as the evaluation basis in their quality terms.
Our perspective on the DQF and MQM question
Most debates about DQF versus MQM are settled by history. The subset became the core. The useful question for a regulated buyer is what evidence a quality score leaves behind. We see the same pattern in life sciences, legal and defense work. Teams adopt a typology, then fail to connect it to terminology governance, reviewer qualifications and data handling. A score without those controls does not survive an audit. This is why our AI+HUMAN hybrid workflow puts subject-matter experts on every deliverable and keeps processing on private EU-hosted infrastructure. The error categories from MQM give our clients and our linguists a shared language for what went wrong and why. We recommend that clients write the typology, weights and critical error rules into the contract before the first file moves. > Eric Brown
How can teams put this into production?
Teams that need a vendor to work against an MQM-based specification can review the translation and localization services AD VERBUM offers. The offer covers translation, localization, interpretation, multilingual SEO, voice over and multilingual documentation. Pricing is set by quote, based on the scope of the project.
For regulated content, share the document types, the target languages and the quality specification at the start. This lets the vendor align terminology, review roles and evaluation sampling with your risk tiers.
FAQ
Is DQF still used?
DQF exists mainly as a historical reference and through its harmonized form. The DQF subset was developed further into MQM Core. Most teams now cite MQM Core, and ISO 5060 and ASTM WK 46396 use it.
Is MQM the same as machine translation scoring?
No. MQM is a manual, reference-free evaluation by human evaluators. It can assess machine translation output, but it is distinct from automatic metrics such as BLEU.
Should a team use the full MQM typology?
Usually not. The MQM committee states that MQM is not intended to be used in its entirety. Start with MQM Core and add types for high-risk content.
Can MQM measure turnaround time or vendor professionalism?
No. MQM focuses on content quality. Track timeliness and conduct through separate service level and vendor review controls.
Sources