Multidimensional Quality Metrics (MQM) Definition

1.1. Scope

This document applies primarily to quality assessment of translated content (and thus to the output of translation systems). It does not apply to assessment of translation processes or projects. Here “translated content” is to be understood broadly to include text, graphics, and any other content which may be translated or adapted for multiple locales (i.e., a combination of language and geographical region). MQM applies to the translation industry, interpreted broadly to include localization (of software and other technical content) and “transcreation” (creative adaptation of content for target audiences and purposes, including but not limited to, adaptation of marketing materials and multi-media content), as well to various types of purely textual translation.

MQM is useful for assessing verifiable qualities of translations. It is not intended to address purely subjective criteria (such as “artistry” or “elegance”) that may be of key importance in some circumstances. Rather, it provides a functional approach to quality that seeks to see whether a translation meets specifications and to identify aspects that may fall short of expectations.

MQM is designed to apply to (monolingual) source texts as well as translated target texts. MQM’s , , , , and branches apply equally to source texts and target texts (although some specific issues within them might apply more to one or the other). The dimension is specific to translated texts (or, more properly, to the relationship between source and target text). While is more likely to apply to target texts, it can apply to source texts. And, finally, the dimension applies solely to source content (many of its issues correspond to specific faults in the target text that can be identified under ).

1.2. Quality assessment, quality assurance, and quality control

Within the translation industry, three terms are used somewhat interchangeably to refer to quality activities: quality assessment, quality assurance, and quality control. However, within broader literature on quality these terms have distinct meanings and should be distinguished:

• Quality assessment (sometimes considered synonymous with “quality evaluation”) is the measurement of the extent to which a product complies with quality specifications (see the definition of “quality” below).
• Quality assurance refers to ways of preventing mistakes or defects in manufactured products and avoiding problems when delivering solutions or services to customers. Quality assurance relies on continual assessment of quality.
• Quality control is the process of checking whether manufactured products meet stated quality specifications. Quality control is a decision process based on the results of quality assessment.

The focus of MQM is on quality assessment, which is essential to quality assurance and quality control. This document does not, however, specify or recommend particular quality assurance or quality control processes. (Note that within the translation industry there is widespread confusion between “quality assessment” and “quality assurance” within the localization industry, partially due to the adoption of the LISA Quality Assurance Model, which actually provided a model for quality assessment.)

3.1. Fairness

As noted above, MQM applies to both source and target texts (with different dimensions applying to them). The default MQM scoring method accordingly allows for users to assess source texts to obtain a quality score for source texts and, if both source and target are assessed, issues found in the source may be counted against penalties for issues in the target text, resulting in higher scores. While not all implementations or usages scenarios will examine the source or count problems in the source in favor of translators, this principle is intended to help ensure that translators are recognized and credited when they have to translate inferior source texts rather than being blamed for all problems, even those beyond their control.

3.2. Flexibility

There are a number of ways to assess the quality of translations. Two primary methods are used in industry and academia:

• Analytic. Analytic quality assessment methods focus on the identification of precise issues within the object being assessed, such as (for a translation) identification of specific mistranslations, spelling errors, etc.
• Holistic. Holistic assessment methods focus on overall characteristics of the object being assessed, such as (in the case of translated texts) reader impression, sentiment, clarity, accuracy, style, whether it enables a task to be completed, and so forth.

MQM is ideally suited for implementation as an analytic metric. It is also easily adapted to serve as the basis for holistic assessments.

Rather than proposing a single metric for assessing all translations, MQM provides a flexible method for defining and declaring metrics that can be adapted to specific requirements. These requirements are generally stated in terms of a set of 12 “parameters” (see Section 9.2. MQM parameters), a subset of the translation parameters described in ASTM F2575:2014 that focuses primarily on aspects of the translation product (rather than the project or process). Using these parameters to define requirements and expectations before translation allows users to create appropriate metrics before translation begins and provides translators with a clear view of the criteria for assessing their work.

In addition, metrics must support both simple and sophisticated requirements. Rather than proposing yet another metric with more detail, MQM provides a flexible catalog of defined issue types that can support any level of sophistication, from a simple metric with two categories to a complex one with thirty or forty. It also supports both holistic assessment (for quick acceptance testing) and error markup/counts for cases where detailed analysis is required.

5.1. MQM issues

5.1.1. List of MQM issues

The full list of MQM issues is maintained in a separate document at .

5.1.2. High-level structure

At the top level, MQM is defined into major dimensions:

• : Accuracy issues address the relationship of the target text to the source text and can be assessed only by considering this relationship. Changes in intended meaning, addition and omission of content, and similar issues are considered in it.
• : Design includes issues related to the physical presentation of text, typically in a “rich text” or “markup” environment.
• : Fluency includes those issues about the linguistic “well-formedness” of the text that can be assessed without regard to whether the text is a translation or not. Most Fluency issues apply equally to source and target texts.
• : Internationalization covers areas related to the preparation of the source content for subsequent translation or localization. Internationalization issues may be detected through problems found in the target (particularly from those included in , but an Internationalization audit is generally conducted separately from a general assessment of translation quality.
• : Issues in Locale convention relate to the formal compliance of content with locale-specific conventions, such as use of proper number formats. If content is otherwise correctly translated and fluent but violates specific locale expectations (as defined in the translation specifications), it is addressed in this dimension. This dimension does not cover issues related to whether the content itself is appropriate for the locale (these issues are covered under .
• : Style issues relate to what is commonly known as “Style”, defined both formally (in style guides) and informally (e.g., a “light style” or an “engaging style”). These issues are closely related to , but are often treated separately by tools and quality processes and so are grouped as a separate dimension in MQM.
• : Terminology issues relate to the use of domain- or organization-specific terminology (i.e., the use of words to relate to specific concepts not considered part of general language). Adherence to specified terminology is widely considered an issue of central concern in both translation and content authoring. Issues in this branch should not be used for general language mistranslation (e.g., translations that would not be considered correct under reasonable circumstances), and should be reserved for issues related to terminology (e.g., a translation is reasonable but incorrect in the context of specific technical domain or for a particular organization).
• : Verity issues relate to the suitability of content for the target locale and audience. They do not relate to fluency or accuracy since content may be fluently written and accurately translated and still be inappropriate for the target locale or audience. For example, if a text translated for Germans in Germany refers to options available only in the UK, these portions will likely be problematic. For more details on Verity, see the discussion below.
• : The Compatibility dimension includes issues taken from legacy metrics that are not considered appropriate for general use in MQM (because they are related to areas not covered by MQM, such as deadlines, software functionality, or physical production). They are included only for compatibility with these older metrics and should not be used for new MQM metrics.
• : This dimension is used for issues which cannot be otherwise classified into a dimension of MQM. In cases where an unforeseen issue can be classified as belonging to a dimension, it should be classified in that dimension under the top level or using a custom issue type. In practice Other should be used extremely rarely.

More information on the dimensions and their content can be found in the full list of MQM issues at

5.2. MQM Core

In order to simplify the application of MQM, MQM defines a smaller “Core” consisting of 20 issue types that represent the most common issues arising in quality assessment of translated texts. The Core represents a relatively high level of granularity suitable for many tasks. Where possible, users of MQM are encouraged to use issues from the Core to promote greater interoperability between systems.

The MQM Core can be graphically represented as follows (branches in gray italics represent major branches not included in the MQM core) (available here in SVG format):

The 19 issues are defined in the MQM core as follows:

Definitions for these issues can be found in the list of MQM issue types.

Even the 20 issues of the Core represent more issues than are likely to be checked in any given quality application and users may define subsets of the core for their needs. It is recommended for translation quality assessment tasks that issues contain at least the issue types and if no other more granular types are included.

5.3. User extension

While users are strongly encouraged to limit issue types to pre-defined MQM issues, they may add additional issue types to MQM to meet additional requirements. User-defined issue types MUST include the following information:

• Name: A human-readable name for the issue type.
• ID: An unprefixed QName that serves as the XML identified of the issue type. The ID MUST begin with x- to indicate that the issue is a user-defined issue. E.g., x-respeaking-error would be a valid ID for a user-defined “respeaking error” but respeaking-error would be invalid.
• Parent: The ID value of the parent issue type. The parent may by a predefined MQM issue type or a user-defined issue type.
• Definition: A human-readable explanation of the issue type.

User extensions do not provide interoperability between systems and impede the exchange of data. Nevertheless they may be needed to support requirements not anticipated in MQM. Users should tie extensions into the predefined hierarchy using the parent value as much as possible since doing so provides consumers of MQM data with the best guidance in interpreting unknown categories and mapping them to other systems. As with other aspects of MQM, users should limit granularity to the least granular level that meets requirements.

Users who encounter frequent need for custom extensions are encouraged to communicate their requirements to the MQM project for possible inclusion of these types in future versions of MQM.

5.4. Integration with other metrics

In addition to MQM, it may be desirable to use other metrics that cannot be converted to a native MQM representation for various purposes. The key principle in integrating metrics is that they must be scoped to indicate to what MQM content they apply. For example, if a metric assesses only readability, it would be scoped to provide a score for MQM , while a metric that provides a score for “Adequacy” would provide a score for MQM . A metric that provides an undifferentiated “quality” score would take all of an MQM metric as its scope and thus provide an overall score.

Non-MQM scores may be indicated in an MQM report by using the nodeScore and scoreType attributes, which may be appended to any node in the score report.

As the interpretation of any particular metric’s result/score is likely to depend on the specifics of the assessment, MQM can provide no guidance on how to utilize the result/score of non-MQM metrics. Results may be appended to MQM reports at the appropriate nodes in the MQM hierarchy and users may wish to combine these results with the results of MQM-based evaluation, (e.g., through averaging MQM and non-MQM scores normalized on a 1-100 scale). Such combinations are outside the scope of MQM.

As an example, the BLEU metric, an automatic metric for assessing machine translation (MT) quality with respect to human reference translation(s), is widely used in MT research. In the case of BLEU, the scope is global because BLEU provides a single, undifferentiated quality score. A BLEU score would thus be provided as parallel to the overall MQM score (see Section 8. Scoring for a recommended method for generating an MQM score). An implementer could utilize the BLEU score in various ways in conjunction with MQM: e.g., only assessing those translations that obtain a BLEU score over a specific threshold, averaging the BLEU and MQM scores, or using both scores for thresholds.

While the specific use of other scores cannot be mandated, their usage should not conflict with the MQM principles. For example, a metric’s results should not be stated to apply to the branch of MQM if they include the results of an evaluation of whether or not terms have been translated correctly.

6.1. MQM metrics description

MQM provides an XML mechanism for exchanging descriptions of MQM-compliant metrics. MQM metrics description files use the .mqm file name extension. An .mqm file contains a hierarchical list of MQM issue types. This listing MUST conform to the hierarchy of issue types.

The following is an example of a small metric description file with issue names in both English and German. It includes a user-defined extension (x-respeaking) used to identify errors caused when a vocal text being respoken without background noise based on a live audio feed is incorrectly repeated by the person doing the respeaking, leading to a mistranscription.

<?xml version="1.0" encoding="UTF-8"?>
<mqm version="0.9">
  <head>
    <name>Small metric</name>
    <descrip>A small metric intended for human consumption</descrip>
    <version>1.5</version>
    <src>http://www.example.com/example.mqm</src>
  </head>
  <issues>
    <issue type="accuracy" display="no">
      </issue>
      <issue type="omission" weight="0.7"/>
      <issue type="addition"/>
    </issue>
    <issue type="terminology" weight="1.5"/>
        <issue type="style" weight="0.5"/>
    <issue type="fluency" display="no">
      <issue type="spelling"/>
      <issue type="grammar"/>
      <issue type="unintelligible" weight="1.5"/>
    </issue>
    <issue type="x-respeaking" weight="1.5"/>
  </issues>
  <displayNames>
    <displaNameSet lang="en">
      <displayName typeRef="accuracy">Adequacy</displayName>
      <displayName typeRef="terminology">Terminology</displayName>
      <displayName typeRef="omission">Omission</displayName>
      <displayName typeRef="addition">Addition</displayName>
      <displayName typeRef="fluency">Fluency</displayName>
      <displayName typeRef="style">Style</displayName>
      <displayName typeRef="spelling">Spelling</displayName>
      <displayName typeRef="grammar">Grammar</displayName>
      <displayName typeRef="unintelligible">Unintelligible</displayName>
      <displayName typeRef="x-respeaking">Respeaking</displayName>
    </displaNameSet>
    <displayNameSet lang="de">
      <displayName typeRef="accuracy">Genauigkeit</displayName>
      <displayName typeRef="terminology">Terminologie</displayName>
      <displayName typeRef="omission">Auslassung</displayName>
      <displayName typeRef="addition">Ergänzung</displayName>
      <displayName typeRef="fluency">Sprachkompetenz</displayName>
      <displayName typeRef="style">Stil</displayName>
      <displayName typeRef="spelling">Rechtschreibung</displayName>
      <displayName typeRef="grammar">Grammatik</displayName>
      <displayName typeRef="unintelligible">Unverständlich</displayName>
      <displayName typeRef="x-respeaking">Sprecherfehler</displayName>
    </displayNameSet>
  </displayNames>
  <severities>
    <severity id="minor" multiplier="1"/>
    <severity id="major" multiplier="10"/>
    <severity id="critical" multiplier="100"/>
  </severities>
</mqm>

6.2. MQM inline attributes

MQM implements the following attributes in the mqm namespace:

• issueType. Contains the MQM issue type, listed by ID. Note: MQM implementations MUST use the ID and MUST NOT use a localized name.
• issueSeverity. Contains the MQM issue severity using the name defined in the metric. Note that the default severity level names are minor, major, and critical. While other values MAY be used, if they are used they MUST be defined in the metric definition for proper interpretation.

MQM is designed to be used in conjunction with the following ITS 2.0 attributes from the localization quality issue data category:

• locQualityIssueType: Contains the issue type as defined by ITS 2.0. Mapping the native MQM value to the appropriate ITS issue type helps ensure compatibility with ITS 2.0-aware implementations, even if they do not implement MQM.
• locQualityIssueComment: Contains a human-readable comment about the issue.
• locQualityIssueSeverity: Contains a rating of severity from 0 to 100. Mapping from the name contained in the MQM issueSeverity attribute to this attribute enables ITS 2.0-aware tools to interpret the severity of the issue.

To ensure compatibility with ITS 2.0 markup, implementers SHOULD use ITS 2.0 markup where possible. All of the ITS 2.0 localization quality annotation may be used. MQM markup adds capability to the ITS 2.0 quality markup.

			
<?xml version="1.0"?>
<doc xmlns:its="http://www.w3.org/2005/11/its" its:version="2.0">
<doc xmlns:mqm="[XXXXXXXXXXX]" mqm:version="1.0">
  <para><span 
      mqm:issueType="spelling"
      mqm:issueSeverity="major"
      its:locQualityIssueType="misspelling"
      its:locQualityIssueComment="Should be Roquefort"
      its:locQualityIssueSeverity="50">Roqfort</span> is an cheese</para>
</doc>

To create this markup the following process is followed:

1. The MQM issue type () is mapped to the corresponding ITS 2.0 type (ITS 2.0 is less fine-grained than MQM in many cases) as described in 8. Relationship to ITS and added as the value of its:locQualityIssueType.
2. The MQM issue type and severity are declared in the mqm: namespace
3. The value of the severity multiplier is declared on a scale from 0 to 100 and inserted as the value of the its:locQualityIssueSeverity attribute. In this case the multiplier value was 5 (out of 10), so it is represented as 50 in ITS markup.
4. A comment is added using the its:locQualityIssueComment attribute.
5. Globally, the relevant profile (specifications and metric definition) are linked using the its:locQualityProfile attribute.

6.3. MQM inline elements

In general, MQM XML implementations should use existing span-level elements in the native XML format that MQM is being added to where possible. This use can be done using any of the ITS 2.0 methods with the addition of the MQM-specific attributes. However, such elements may not be available. In such cases, MQM defines two elements that can be used to add inline markup:

• <mqm:startIssue />. This element defines the starting position of an MQM span.
• <mqm:endIssue />. This element defines the end position of an MQM span.

Two empty elements are used so as to prevent any interference between MQM tags and existing XML structure, such as those that could be caused by improperly nested elements. To pair these tags the id attribute is used. ID values MUST be unique within the document to prevent confusion.

An example of an MQM annotation is seen in the following XML snippet:

    <para>“Instead of strengthening
        <mqm:startIssue type="function-words" id="1f59a2" severity="minor" agent="f-deluz" comment="article unneeded here" active="yes"/>
        the<mqm:endIssue idref="1f59a2"/>
        civil society, the president cancels
        <mqm:startIssue type="agreement" severity="major" comment="should be “it”" agent="f-deluz" id="3c469d" active="yes"/>them<mqm:endIssue idref="3c469d"/>
        de facto”, deplores Saeda.
    </para>

The mqm:startIssue element MUST take the following mandatory attributes:

• id. Used to match the corresponding mqm:startIssue and mqm:endIssue tags within text
• type. Provides the MQM issue type.

The mqm:startIssue element CAN take the following optional attributes:

• severity. Permissible values defined by the MQM metric in use. Provides the severity of the issue. Default value is undefined.
• agent. Text string identifying the agent that supplied the annotation. Default value is undefined.
• comment. Text string containing a human-readable comment attached to an issue. Default value is undefined
• active. One of yes or no. Indicates whether the issue is considered active (yes) or inactive (no). Default value is yes. If an issue is marked as inactive, this means that it has either been resolved or determined not to be an actual error.

In addition, ITS 2.0 attributes MAY be added to these elements to promote greater interoperability.

The mqm:endIssue element MUST take the following mandatory attribute:

• idref. A value corresponding to the id of the mqm:startIssue tag that begins the identified span.

Use of these inline elements also requires that the mqm namespace be declared in the document. The method for declaring this namespace needs to be determined.

7.1. MQM-to-ITS mapping

MQM issue types are mapped to ITS issue types according to the following table. Note that this mapping is unambiguous and MUST be followed to ensure consistency between applications.

MQM issue type	ITS 2.0 issue type
	mistranslation
	addition
	mistranslation
	mistranslation
	inconsistent-entities
	mistranslation
	numbers
	mistranslation
	omission
	mistranslation
	untranslated
	other (for all children)
	formatting
	length
	formatting
	markup
	length
	formatting
	other
	other
	characters
	other
	other
	non-conformance
	duplication
	grammar
	register (ITS register covers both and )
	inconsistency
	other
	other
	characters
	other
	pattern-problem
	other
	misspelling
	typographical
	uncategorized
	internationalization (for all subtypes)
	locale-violation (for all subtypes)
	style
	register (ITS register covers both and )
	style
	terminology (for all subtypes)
	other
	legal
	locale-specific-content

Note that the entire Internationalization branch of MQM maps to the ITS internationalization type. It is anticipated that this mapping will apply to all children of the MQM Internationalization issue type that may be added in the future.

7.2. ITS-to-MQM mapping

Mapping from ITS to MQM is less likely to be used and presents particular problems since MQM metrics typically contain only a small subset of the full MQM issue set. As a result MQM issues to which ITS localization quality issue type values are mapped may not exist in a particular MQM metric. In such cases processes MUST map the ITS value to the closest higher-level issue type in MQM if one exists in the target MQM metric. If no higher-level issue type exists in the target MQM metric, the process MUST skip the ITS 2.0 issue type (but MAY preserve the ITS 2.0 markup).

For example, if a process encounters the ITS 2.0 type and the target MQM metric does not contain but does contain , the ITS omission value would be mapped to MQM . However, if the MQM metric does not contain , the higher node in the MQM hierarchy, the ITS omission issue type would be ignored/omitted by the conversion process.

Note that the above requirements mean that in some cases there may be a many-to-one mapping from ITS to MQM. For example, if a document contains ITS annotations for omission, untranslated, and addition, but the target MQM metric contains and no daughter categories, all of these categories would be mapped to MQM . In other words, there is no universal mapping from ITS to all MQM metrics since MQM metrics do not all contain the same issues.

Processes encountering issues such as those described in the previous paragraphs SHOULD alert the user about the information loss or remapping if user interaction is expected by the process.

In most cases the table shows that the ITS issue types map to MQM issue types with identical (except for casing) or similar names, highlighting the evolutionary relationship between ITS and MQM. Those items where names are different in a non-trivial manner are marked with an asterisk (*) to help draw attention to the fact that the names do not match.

Note that the ITS uncategorized category maps to MQM even though MQM maps to ITS uncategorized. In other words, the mapping is asymmetric because the semantics of uncategorized are broader than .

8.1. Default severity levels for error-count metrics

Version 0.3.0 made major changes with respect to severity multipliers. These changes render the default scoring for versions 0.3.0 and later incompatible with earlier versions. Version 0.9.1 introduced a new severity level, none, that always has a penalty of 0, i.e., it does not count against the transaltion, it is used to mark items that should be changed, but which are not considered errors for scoring purposes (see below).

For the purposes of calculating quality scores, the following default values apply:

Weight

All issues have a default weight of 1.0. This weight can be updated on a per-issue basis to reflect specific requirements.

Severity

The default severity levels are defined as follows:

• none: 0. Issues with the severity level none are items that need to be noted for further attention or fixing but which should not count against the translation. This severity level can be conceived of as a flag for attention that does not impose a penalty. It should be used for “preferential errors” (i.e, items that are not wrong, per se, but where the reviewer or requester would like to see a different solution), systematic repeated errors that can be easily fixed (e.g., a translator has systematically used an incorrect domain term but it is a simple matter of search and replace to correct them all). Because no penalty is assessed for this level, it is not discussed in the scoring formulae.
• minor: 1. Minor issues are issues that do not impact usability or understandability of the content. For example, if an extra space appears after a full stop, this may be considered an error, but does not render the text difficult to use or problematic (even if it should be corrected). If the typical reader/user is able to correct the error reliably and it does not impact the usability of the content, it SHOULD be classified as minor. Since minor errors do not impact the usability of the content, resolution of them is at the discretion of those responsible for the content.
• major: 10. Major issues are issues that impact usability or understandability of the content but which do not render it unusable. For example, a misspelled word may require extra effort for the reader to understand the intended meaning, but do not make it impossible. If an error cannot be reliably corrected by the reader/user (e.g., the intended meaning is not clear) but it does not render the content unfit for purpose, it SHOULD be categorized as major. While it is generally advisable to fix major errors prior to use of the content, the inclusion of major errors may not, by themselves, render the text unfit for purpose.
• critical: 100. Critical issues are issues that render the content unfit for use. For example, a particularly bad grammatical error that changes the meaning of the text would be considered critical. If the error prevents the reader/user from using the content as intended or if it presents incorrect information that could result in harm to the user it MUST be categorized as critical. In general, critical errors have to be fixed prior to use of the text since even a single critical error is likely to cause serious problems.

8.2 Scoring algorithm

MQM can generate target document quality scores according to the following formula:

where:

TQ = quality score

The overall rating of quality

TP = penalties for the target content

Sum of all weighted penalty points assigned to the target text

SP = penalties for the source content

Sum of all weighted penalty points assigned to the target text

All penalties are relative to the sample size (in words) and are calculated as follows (assuming default weights and severity levels):

where:

${\mathrm{Issues}}_{\mathrm{minor}}$	= Number of issues with a “minor” severity
${\mathrm{Issues}}_{\mathrm{major}}$	= Number of issues with a “major” severity
${\mathrm{Issues}}_{\mathrm{critical}}$	= Number of issues with a “critical” severity

A score can thus be generated through the following (pseudo-code) algorithm:

foreach targetIssue {
	targetIssueTotal = targetIssueTotal +
	(targetIssue * weight[sourceIssueType] * severityMultiplier);
}

foreach sourceIssue {
	sourceIssueTotal = sourceIssueTotal +
	(sourceIssue * weight[sourceIssueType] * severityMultiplier);
}

// Generate overall score
translationQualityScore = 100 - (targetIssueTotal / wordcount) + (sourceIssueTotal / wordcount);

In this algorithm, each issue type has a weight assigned by the metric that is retrieved and used to determine the individual penalties. Penalties are cumulative. Note that if the source is examined, penalties against the source are effectively added to the overall score for the translation, reflecting the fact that they indicate problems in the source the translator had to deal with. If the source is not assessed, the source penalties are by definition 0 and do not count for or against the translation’s quality score.

(Scores can be generated for any dimension or branch in the MQM hierarchy by counting only those issues in that selection. Note that counting source issues is optional and that if a score for a source document is desired then the formula should ignore target issues and instead subtract the total of source issues divided by the wordcount from 100 to arrive at a source content score.)

This algorithm can serve as a model for other systems, such as metrics with two severity levels or those with four. However, using other models will impede comparability of scores generated by various metrics.

8.3. Default severity multipliers from versions earlier than 0.3.0 (deprecated)

The following severity multipliers were recommended as default multipliers prior to version 0.3.0. The former default severity weights were taken from the LISA QA Model and represent common industry practice. Discussion with experts in psychometrics, however, revealed that the range of values was too close to provide sufficient discrimination between relatively insignificant errors and those considered serious enough to reject a project. As these values were implemented in a number of tools they are documented here:

• minor: 1
• major: 5
• critical: 10

Scores using these multipliers can be easily updated to reflect the new values simply by changing the multipliers in the formula. Similarly, new scores can be compared with old scores by using this values in place of the new ones. However, as the old multipliers are deprecated, they SHOULD NOT be used as the default model for any new implementations.

9.1. Example of defining a metric

The following example will help clarify how the process works. The example is for a case in which a company that makes network diagnostic gear wishes to evaluate whether automatic (machine) translations into Japanese of user-generated forum content written in English is helping their Japanese users solve technical problems with their equipment.

1. Selection of assessment method.
   1. What: The company wishes to assess a translation product (the forum content) and also the MT system they are using to translate the content.
   2. Who: The company wants to use its customers to evaluate the translation since they are the only ones who can determine whether the content meets their needs.
   3. Where: The assessment must be done on the user-to-user forum with end users who are not experts in translation or language and who cannot be trained in advance.
   4. When: The assessment will take place after texts have already been published on the website. The texts will be raw MT output with no post-editing or other correction.
   5. Why: The assessment will be used to determine if the MT system’s results help users meet their needs or whether more manual processes (e.g., MT + post-editing) are required.
   
   Based on these answers, the company decides to use a holistic assessment method with a low number of dimensions (no more than three).
2. Creating the specifications. The company fills out a worksheet to define the values for the parameters in their specification (described in section 8.2. Definition of MQM parameters below) and creates a full set of translation specifications.
3. Selection of Dimensions. Based on their translation specifications they determine that the following dimensions are relevant to this task: Accuracy, Fluency, and Verity. Because of the nature of the assessment method and the assessors, however, the company decides to limit assessment to three issues: Terminology, Fluency, and Verity. Although Accuracy is highly important, they cannot expect their users to understand English well enough to assess the accuracy of translated texts.
4. Building the metric. Based on the selection of a holistic metric and three issues, the company selects three issue types and implements a metric with three questions on their website at the end of each translated forum entry:
   1. Did this answer enable you to solve your problem? (Yes/No) (Addresses Verity)
   2. Was this answer grammatically correct? (Yes/No) (Addresses Grammar)
   3. Did this answer use the correct words to describe your product and the solution? (Yes/No) (Addresses Terminology)
   
   In addition, because the company realizes that their customers cannot assess some core aspects and help them evaluate their MT system, they decide to create a second, analytic metric for human assessors to check a subset of the output.

Although simple, this example, shows how it is possible to build customized metrics to meet specific requirements using MQM.

9.2. Definition of MQM parameters

MQM makes use of a selection of 11 of the 21 parameters defined in ASTM F2575, with the addition of one additional parameter, Output modality, which is subsumed under Text type in ASTM F2575 but which is broken out in MQM because of its special impact on some translationed. The parameters are defined as follows:

9.3. Analytic metrics

Analytic metrics are created by making a selection of relevant issues from the listing of MQM issue types. The following procedure may be used to create a metric:

1. Complete a full set of project specifications, including the 12 MQM parameters. Ensure that all stakeholders are in agreement about the values of the parameters. (Note that the value of some parameters, such as the target language, may change from project to project, so implementers should consider the range of likely values. For example, if a project will be translated into 15 languages, the impact each language might have should be considered.)
2. For the value of each parameter, consider what features of the text would be needed to verify that the text meets specifications and note these issue types down. Note that “doesn’t matter” is an acceptable value for many parameters and if this value is chosen, the parameter may be skipped. (E.g., if Style is judged to be insignificant, then this parameter will be skipped in assessment.
3. After deciding what features need to be checked, determine which issue types can be used to assess that feature and note these types.
4. From the list of issue types, prioritize them based on the importance of each parameter and then make a selection of issue types based on this list and the priorities. (Note that it may be impractical to do fine-grained analysis of every potential issue type identified. Feedback from LSPs suggests that six to seven issue types is sufficient for most assessment tasks, although some use up to twenty.
5. If a score is to be assigned, assign weights to the issues. Assigning weights is a tricky process and should be done by assessing existing translations deemed to be acceptable, borderline acceptable, and unacceptable to see what impact each issue type has on that judgment. Note that some existing metrics, such as SAE J2450, have predefined weights that should be honored. The default issue weight in MQM is 1.0 and any positive decimal value may be used.
6. If the resulting metric is to be implemented in an MQM-compliant tool chain, it should be declared as described in Section 7.1. MQM metrics description.

When considering which issues to check, creators of metrics should consider the following practical guidelines:

1. Are there any requirements for compatibility with legacy systems or standard/semi-standard specifications? If so, choose issue types that correspond to those used by those systems/specifications. In most cases it is possible to emulate legacy metrics in MQM with little or no modification, although some might require the use of custom extensions.
2. Select the least granular issue types that allow assessment of whether the text meets specifications. For example, in many cases use of the category would be sufficient because it is not particularly relevant to know what subcategory is used. On the other hand, when trying to diagnose problems generated by an MT system, finer-grained types might be necessary.
3. When possible, choose issues from the MQM Core. Using these issues helps ensure compatibility. However, the Core does not cover all cases, including common ones such as checking formatting, because it is focused on text translations.
4. Consider not just requirements for one set of specifications/parameters, but also for other likely sets. For example, if two types of translations are frequently assessed, it may make sense to develop one list of issues with different sets of weights and to use the single (master) set of issues. This practice is recommended to prevent the need to train evaluators on multiple metrics.

9.4. Holistic metrics

Holistic assessment methods are more flexible in some respects than error-count metrics. They are designed to provide an assessment of the translated text as a whole rather than a detailed accounting of all errors. As analytic assessment can be time consuming and is not needed in all cases (e.g., when the question is whether a text should be accepted or not), holistic methods may be more appropriate in some cases. Most of the MQM issue types can be easily used as either analytic types or holistic types that apply to the text as a whole. For example, the MQM issue type can be used by asking assessors using a holistic tool whether the text is punctuated correctly. In this context some issues will be more useful than others. For example, the issue type is unlikely to be useful in most holistic assessments since it generally makes sense only with regard to very specific sections of a text. By contrast, categories like can more readily be applied to entire texts.

Note that there is no single method for building holistic scores. In a holistic approach specific issues are addressed through qualitative questions that may be assessed via ranking or on a binary- or scalar-value system. For example, a holistic assessment might address the issue via questions like the following:

• The translated text is spelled correctly:
  [ ] Yes
  [ ] No
• The translated text is spelled correctly:
  [ ] Strongly disagree
  [ ] Disagree
  [ ] Neither disagree nor agree
  [ ] Agree
  [ ] Strongly agree
• Does the translated text meet expectations with regard to correct spelling?
  [ ] The text does not meet expectations
  [ ] The text meets expectations
  [ ] The text exceeds expectations

Because the scoring for holistic systems is highly dependent on the type of assessment scale used, no specific scoring system is provided here. Users of MQM who wish to implement it in a holistic environment should tie holistic questions to specific MQM issue types and develop appropriate scoring systems. This version of MQM does not define a system for describing holistic scoring systems, although future versions may do so. However, by using the MQM issue types and associating specific holistic questions with them, implementers can make their metrics more transparent and tie them to project parameters in the same way that can be done with error-count metrics.

The following guidelines may assist in designing appropriate holistic assessments and selecting issue types:

• Use the highest-level issues that suffice to provide the needed assessment. For example, “Is the text translated accurately?” (assessing ) is more likely to be answered accurately than “Does the text show any problems related to ‘false friends’?” (assessing ), especially as false friends are less likely to occur than generalized mistranslation. However, for some assessment purposes, fine-grained holistics may be appropriate.
• Write holistic statements or questions that clearly address the desired issue type. For example, “Does the text comply with specified guidelines for style?” is more likely to obtain an accurate assessment for than “Does the text read well?”, which is vague in its intent.
• Design holistic questions that allow assessors to give credit, not just penalties. Holistic assessment tools often allow assessors to indicate that a text outperforms expectations in order to give credit for jobs well done. Using scalars that give translators credit is in line with the MQM principle of fairness and is roughly analogous to the MQM practice of assigning credits to translators for errors detected in the source.

11.1. SAE J2450

The mapping from SAE J2450 is somewhat complex in that the distinction between severity levels is, in part, based on the whether the issue changes the meaning between target and source, meaning that—at least in principle—a minor error in J2450 would correspond to the Fluency branch in MQM and a major error would correspond to the Accuracy branch. Nevertheless, for most purposes, the following mapping should suffice.