Emotion Markup Language (EmotionML) 1.0

Abstract

As the web is becoming ubiquitous, interactive, and multimodal, technology needs to deal increasingly with human factors, including emotions. The present draft specification of Emotion Markup Language 1.0 aims to strike a balance between practical applicability and scientific well-foundedness. The language is conceived as a "plug-in" language suitable for use in three different areas: (1) manual annotation of data; (2) automatic recognition of emotion-related states from user behavior; and (3) generation of emotion-related system behavior.

Status of this document

This section describes the status of this document at the time of its publication. Other documents may supersede this document. A list of current W3C publications and the latest revision of this technical report can be found in the W3C technical reports index at http://www.w3.org/TR/.

This is the First Public Working Draft of "Emotion Markup Language (EmotionML) 1.0", published on 29 October 2009.

The present draft specification draws on previous work in the Emotion Markup Language Incubator Group (EmotionML XG, 2007-2008) that proposed elements of a generally usable markup language for emotions and related states, as well as the earlier Emotion Incubator Group (2006-2007) that had identified a comprehensive list of requirements arising from use cases of an Emotion Markup Language.

The present report reflects the starting point of formal specification. The group expects a process of condensing this document into a simpler, ready-to-use specification which removes unclear parts of the draft and cuts redundancy with related languages such as [EMMA].

This document was developed by the Multimodal Interaction Working Group. The Working Group expects to advance this Working Draft to Recommendation Status.

Please send comments about this document to www-multimodal@w3.org (with public archive).

Publication as a Working Draft does not imply endorsement by the W3C Membership. This is a draft document and may be updated, replaced or obsoleted by other documents at any time. It is inappropriate to cite this document as other than work in progress.

This document was produced by a group operating under the 5 February 2004 W3C Patent Policy. W3C maintains a public list of any patent disclosures made in connection with the deliverables of the group; that page also includes instructions for disclosing a patent. An individual who has actual knowledge of a patent which the individual believes contains Essential Claim(s) must disclose the information in accordance with section 6 of the W3C Patent Policy.

Conventions of this document

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].

Table of Contents

1 Introduction

Human emotions are increasingly understood to be a crucial aspect in human-machine interactive systems. Especially for non-expert end users, reactions to complex intelligent systems resemble social interactions, involving feelings such as frustration, impatience, or helplessness if things go wrong. Furthermore, technology is increasingly used to observe human-to-human interactions, such as customer frustration monitoring in call center applications. Dealing with these kinds of states in technological systems requires a suitable representation, which should make the concepts and descriptions developed in the affective sciences available for use in technological contexts.

This report specifies Emotion Markup Language (EmotionML) 1.0, a markup language designed to be usable in a broad variety of technological contexts while reflecting concepts from the affective sciences.

The report is work in progress. Issue notes are used to describe open questions as well as available choices.

1.1 Reasons for defining an Emotion Markup Language

As for any standard format, the first and main goal of an EmotionML is twofold: to allow a technological component to represent and process data, and to enable interoperability between different technological components processing the data.

Use cases for EmotionML can be grouped into three broad types:

1. Manual annotation of material involving emotionality, such as annotation of videos, of speech recordings, of faces, of texts, etc;
2. Automatic recognition of emotions from sensors, including physiological sensors, speech recordings, facial expressions, etc., as well as from multi-modal combinations of sensors;
3. Generation of emotion-related system responses, which may involve reasoning about the emotional implications of events, emotional prosody in synthetic speech, facial expressions and gestures of embodied agents or robots, the choice of music and colors of lighting in a room, etc.

Interactive systems are likely to involve both analysis and generation of emotion-related behavior; furthermore, systems are likely to benefit from data that was manually annotated, be it as training data or for rule-based modelling. Therefore, it is desirable to propose a single EmotionML that can be used in all three contexts.

Concrete examples of existing technology that could apply EmotionML include:

• Opinion mining / sentiment analysis in Web 2.0, to automatically track customer's attitude regarding a product across blogs;
• Affective monitoring, such as ambient assisted living applications for the elderly, fear detection for surveillance purposes, or using wearable sensors to test customer satisfaction;
• Character design and control for games and virtual worlds;
• Social robots, such as guide robots engaging with visitors;
• Expressive speech synthesis, generating synthetic speech with different emotions, such as happy or sad, friendly or apologetic;
• Emotion recognition (e.g., for spotting angry customers in speech dialog systems);
• Support for people with disabilities, such as educational programs for people with autism.

The Emotion Incubator Group has listed 39 individual use cases for an EmotionML.

A second reason for defining an EmotionML is the observation that ad hoc attempts to deal with emotions and related states often lead people to make the same mistakes that others have made before. The most typical mistake is to model emotions as a small number of intense states such as anger, fear, joy, and sadness; this choice is often made irrespective of the question whether these states are the most appropriate for the intended application. Crucially, the available alternatives that have been developed in the affective science literature are not sufficiently known, resulting in dead-end situations after the initial steps of work. Careful consideration of states to study and of representations for describing them can help avoid such situations.

EmotionML makes scientific concepts of emotions practically applicable. This can help potential users to identify the suitable representations for their respective applications.

1.2 The challenge of defining a generally usable Emotion Markup Language

Any attempt to standardize the description of emotions using a finite set of fixed descriptors is doomed to failure: even scientists cannot agree on the number of relevant emotions, or on the names that should be given to them. Even more basically, the list of emotion-related states that should be distinguished varies depending on the application domain and the aspect of emotions to be focused. Basically, the vocabulary needed depends on the context of use. On the other hand, the basic structure of concepts is less controversial: it is generally agreed that emotions involve triggers, appraisals, feelings, expressive behavior including physiological changes, and action tendencies; emotions in their entirety can be described in terms of categories or a small number of dimensions; emotions have an intensity, and so on. For details, see Scientific Descriptions of Emotions in the Final Report of the Emotion Incubator Group.

Given this lack of agreement on descriptors in the field, the only practical way of defining an EmotionML is the definition of possible structural elements, their valid child elements and attributes, but to allow users to "plug in" vocabularies that they consider appropriate for their work. A central repository of such vocabularies can serve as a recommended starting point; where that seems inappropriate, users can create their custom vocabularies.

An additional challenge lies in the aim to provide a generally usable markup, as the requirements arising from the three different use cases (annotation, recognition, and generation) are rather different. Whereas manual annotation tends to require all the fine-grained distinctions considered in the scientific literature, automatic recognition systems can usually distinguish only a very small number of different states.

For the reasons outlined here, it is clear that there is an inevitable tension between flexibility and interoperability, which need to be weighed in the formulation of an EmotionML. The guiding principle in the following specification has been to provide a choice only where it is needed; to propose reasonable default options for every choice; and, ultimately, to propose mapping mechanisms where that is possible and meaningful.

1.3 Glossary of terms

The terms related to emotions are not used consistently, neither in common use nor in the scientific literature. The following glossary describes the intended meaning of terms in this document.

Action tendency

Emotions have a strong influence on the motivational state of a subject. Emotion theory associates emotions to a small set of so-called action tendencies, e.g. avoidance (relates to fear), rejecting (disgust) etc. Action tendencies can be viewed as a link between the outcome of an appraisal process and actual actions.

Affect / Affective state

In the scientific literature, the term "affect" is often used as a general term covering a range of phenomena called "affective states", including emotions, moods, attitudes, etc. Proponents of the term consider it to be more generic than "emotion", in the sense that it covers both acute and long-term, specific and unspecific states. In this report, the term "affect" is avoided so that the scope of the intended markup language is more easily accessible to the non-expert; the term "affective state" is used interchangeably with "emotion-related state".

Appraisal

The term "appraisal" is used in the scientific literature to describe the evaluation process leading to an emotional response. Triggered by an "emotion-eliciting event", an individual carries out an automatic, subjective assessment of the event, in order to determine the relevance of the event to the individual. This assessment is carried out along a number of "appraisal dimensions" such as the novelty, pleasantness or goal conduciveness of the event.

Emotion

In this report, the term "emotion" is used in a very broad sense, covering both intense and weak states, short and long term, with and without event focus. This meaning is intended to reflect the understanding of the term "emotion" by the general public. In the scientific literature on emotion theories, the term "emotion" or "fullblown emotion" refers to intense states with a strong focus on current events, often in the context of the survival-benefiting function of behavioral responses such as "fight or flight". This reading of the term seems inappropriate for the vast majority of human-machine interaction contexts, in which more subtle states dominate; therefore, where this reading is intended, the term "fullblown emotion" is used in this report.

Emotion-related state

A cover term for the broad range of phenomena intended to be covered by this specification. In the scientific literature, several kinds of emotion-related or affective states are distinguished, see Emotions and related states in the final report of the Emotion Incubator Group.

Emotion dimensions

A small number of continuous scales describing the most basic properties of an emotion. Often three dimensions are used: valence (sometimes named pleasure), arousal (or activity/activation), and potency (sometimes called control, power or dominance). However, sometimes two, or more than three dimensions are used.

Fullblown emotion

Intense states with a strong focus on current events, often in the context of the survival-benefiting function of behavioral responses such as "fight or flight".

2 Elements of Emotion Markup

The following sections describe the syntax of the main elements of EmotionML. The specification is not yet fully complete. Feedback is highly appreciated.

2.1 Document structure

2.1.1 Document root: The <emotionml> element

<emotionml> is the root element of a standalone EmotionML document. It wraps a number of <emotion> elements into a single document. It may contain a single <metadata> element, providing document-level metadata.

The <emotionml> element MUST define the EmotionML namespace, and may define any other namespaces.

Example:

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
 ...
</emotionml>

or

<em:emotionml xmlns:em="http://www.w3.org/2009/10/emotionml">
 ...
</em:emotionml>

Note: One of the envisaged uses of EmotionML is to be used in the context of other markup languages. In such cases, there will be no <emotionml> root element, but <emotion> elements will be used directly in other markup -- see Examples of possible use with other markup languages.

2.1.2 A single emotion annotation: The <emotion> element

The <emotion> element represents an individual emotion annotation. No matter how simple or complex its substructure is, it represents a single statement about the emotional content of some annotated item. Where several statements about the emotion in a certain context are to be made, several <emotion> elements MUST be used. See Examples of emotion annotation for illustrations of this issue.

Whereas it is possible to use <emotion> elements in a standalone <emotionml> document, a typical use case is expected to be embedding an <emotion> into some other markup -- see Examples of possible use with other markup languages.

2.2 Representations of emotions and related states

2.2.1 The <category> element

<category> describes an emotion or a related state in terms of a single category name, given as the value of the name attribute. The name MUST belong to a clearly-identified set of category names, which MUST be defined according to Defining vocabularies for representing emotions.

The set of legal values of the name attribute is indicated in the set attribute of the <category> element. Different sets can be used, depending on the requirements of the use case. In particular, different types of emotion-related / affective states can be annotated by using appropriate value sets.

Examples:

In the following example, the emotion category "satisfaction" is being annotated; it must be contained in the set of values named "everydayEmotions".

<emotion>
    <category set="everydayEmotions" name="satisfaction"/>
</emotion>

The following is an annotation of an interpersonal stance "distant" which must belong to the set of values named "commonInterpersonalStances".

<emotion>
    <category set="commonInterpersonalStances" name="distant"/>
</emotion>

2.2.2 The <dimensions> element

A <dimensions> element describes an emotion or a related state in terms of a set of emotion dimensions. The names of the emotion dimensions MUST belong to a clearly-identified set of dimension names, which MUST be defined according to Defining vocabularies for representing emotions.

The set of values that can be used as tag names of child elements of the <dimensions> element is indicated in the set attribute of the <dimensions> element. Different sets can be used, depending on the requirements of the use case.

There are no constraints regarding the order of the dimension child elements within a <dimensions> element.

Any given dimension is either unipolar or bipolar; its value attribute MUST contain either discrete or continuous Scale values.

A dimension element MUST either contain a value attribute or a <trace> child element, corresponding to static and dynamic representations of Scale values, respectively.

If the dimension element has both a confidence attribute and a <trace> child, the <trace> child MUST NOT have a samples-confidence attribute. In other words, it is possible to either give a constant confidence on the dimension element or a confidence trace on the <trace> element, but not both.

Examples:

One of the most widespread sets of emotion dimensions used (sometimes by different names) is the combination of valence, arousal and potency. Assuming that arousal and potency are unipolar scales with typical values between 0 and 1, and valence is a bipolar scale with typical values between -1 and 1, the following example is a state of rather low arousal, very positive valence, and high potency -- in other words, a relaxed, positive state with a feeling of being in control of the situation:

<emotion>
    <dimensions set="valenceArousalPotency">
        <arousal value="0.3"/><!-- lower-than-average arousal -->
        <valence value="0.9"/><!-- very high positive valence -->
        <potency value="0.8"/><!-- relatively high potency    -->
    </dimensions>
</emotion>

In some use cases, custom sets of application-specific dimensions will be required. The following example uses a custom set of dimensions, defining a single, bipolar dimension "friendliness".

<emotion>
    <dimensions set="myFriendlinessDimension">
        <friendliness value="-0.7"/><!-- a pretty unfriendly person -->
    </dimensions>
</emotion>

Different use cases require continuous or discrete Scale values; the following example uses discrete values for a bipolar dimension "valence" and a unipolar dimension "arousal".

<emotion>
    <dimensions set="discreteValenceArousal">
        <arousal value="very high"/>
        <valence value="slightly negative"/>
    </dimensions>
</emotion>

2.2.3 The <appraisals> element

An <appraisals> element describes an emotion or a related state in terms of a set of appraisals. The names of the appraisals MUST belong to a clearly-identified set of appraisal names, which MUST be defined according to Defining vocabularies for representing emotions.

The set of values that can be used as tag names of child elements of the <appraisals> element is indicated in the set attribute of the <appraisals> element. Different sets can be used, depending on the requirements of the use case.

There are no constraints regarding the order of the appraisal child elements within a <appraisals> element.

Any given appraisal is either unipolar or bipolar; its value attribute MUST contain either discrete or continuous Scale values.

An appraisal element MUST either contain a value attribute or a <trace> child element, corresponding to static and dynamic representations of Scale values, respectively.

If the appraisal element has both a confidence attribute and a <trace> child, the <trace> child MUST NOT have a samples-confidence attribute. In other words, it is possible to either give a constant confidence on the appraisal element, or a confidence trace on the <trace> element, but not both.

Examples:

One of the most widespread sets of emotion appraisals used is the appraisals set proposed by K. Scherer, namely novelty, intrinsic pleasantness, goal/need significance, coping potential, and norm/self compatibility. Another very widespread set of emotion appraisals, used in particular in computational models of emotion, is the OCC set of appraisals (Ortony et al., 1988), which includes the consequences of events for oneself or for others, the actions of others and the perception of objects. Assuming some appraisal variables, say novelty is a unipolar scale with typical values between 0 and 1, and intrinsic pleasantness is a bipolar scale with typical values between -1 and 1, the following example is a state arising from the evaluation of an unpredicted and quite unpleasant event:

<emotion>
    <appraisals set="Scherer_appraisals_checks">
        <novelty value="0.8"/>
        <intrinsic-pleasantness value="-0.5"/>
    </appraisals>
</emotion>

In some use cases, custom sets of application-specific appraisals will be required. The following example uses a custom set of appraisals, defining single, bipolar appraisal "likelihood".

<emotion>
    <appraisals set="myLikelihoodAppraisal">
        <likelihood value="0.8"/><!-- a very predictable event -->
    </appraisals>
</emotion>

Different use cases require continuous or discrete Scale values; the following example uses discrete values for a bipolar appraisal "intrinsic-pleasantness" and a unipolar appraisal "novelty".

<emotion>
    <appraisals set="discreteSchererAppraisals">
        <novelty value="very high"/>
        <intrinsic-pleasantness value="slightly negative"/>
    </appraisals>
</emotion>

2.2.4 The <action-tendencies> element

An <action-tendencies> element describes an emotion or a related state in terms of a set of action-tendencies. The names of the action-tendencies MUST belong to a clearly-identified set of action-tendency names, which MUST be defined according to Defining vocabularies for representing emotions.

The set of values that can be used as tag names of child elements of the <action-tendencies> element is indicated in the set attribute of the <action-tendencies> element. Different sets can be used, depending on the requirements of the use case.

There are no constraints regarding the order of the action-tendency child elements within a <action-tendencies> element.

Any given action-tendency is either unipolar or bipolar; its value attribute MUST contain either discrete or continuous Scale values.

A action-tendency element MUST either contain a value attribute or a <trace> child element, corresponding to static and dynamic representations of Scale values, respectively.

If the action-tendency element has both a confidence attribute and a <trace> child, the <trace> child MUST NOT have a samples-confidence attribute. In other words, it is possible to either give a constant confidence on the action-tendency element, or a confidence trace on the <trace> element, but not both.

Examples:

One well known use of action tendencies is by N. Frijda who generally uses the term "action readiness". This model uses a number of action tendencies that are low level, diffuse behaviors from which more concrete actions could be determined. An example of someone attempting to attract someone they like by being confident, strong and attentive might look like this using unipolar values:

<emotion>
    <action-tendencies set="frijdaActionReadiness">
    	<approach value="0.7"/><!-- get close -->
    	<avoid value="0.0"/>
    	<being-with value="0.8"/><!-- be happy -->
    	<attending value="0.7"/><!-- pay attention -->
    	<rejecting value="0.0"/>
    	<non-attending value="0.0"/>
    	<agonistic value="0.0"/>
    	<interrupting value="0.0"/>
    	<dominating value="0.7"/><!-- be assertive -->
    	<submitting value="0.0"/>
    </action-tendencies>
</emotion>

In some use cases, custom sets of application-specific action-tendencies will be required. The following example shows control values for a robot who works in a factory and uses a custom set of action-tendencies, defining example actions for a robot using bipolar and unipolar values.

<emotion>
    <action-tendencies set="myRobotActionTendencies">
        <charge-battery value="0.9"/><!-- need to charge battery soon, be-with charger -->
        <pickup-boxes value="-0.2"/><!-- feeling tired, avoid work -->
    </action-tendencies>
</emotion>

Different use cases require continuous or discrete Scale values; the following example shows control values for a robot who works in a factory and uses discrete values for a bipolar action-tendency "pickup-boxes" and a unipolar action-tendency "seek-shelter".

<emotion>
    <action-tendencies set="myRobotActionTendencies">
    	<seek-shelter value="very high"/><!-- started to rain, approach shelter -->
    	<pickup-boxes value="slightly negative"/><!-- feeling tired, avoid work -->
    </action-tendencies>
</emotion>

2.2.5 The <intensity> element

<intensity> represents the intensity of an emotion. The <intensity> element MUST either contain a value attribute or a <trace> child element, corresponding to static and dynamic representations of scale values, respectively. <intensity> is a unipolar scale.

If the <intensity> element has both a confidence attribute and a <trace> child, the <trace> child MUST NOT have a samples-confidence attribute. In other words, it is possible to either give a constant confidence on the <intensity> element, or a confidence trace on the <trace> element, but not both.

A typical use of intensity is in combination with <category>. However, in some emotion models (e.g. Gebhard, 2005), the emotion's intensity can also be used in combination with a position in emotion dimension space, that is in combination with <dimensions>. Therefore, intensity is specified independently of <category>.

Example:

A weak surprise could accordingly be annotated as follows.

<emotion>
    <intensity value="0.2"/>
    <category set="everydayEmotions" name="surprise"/>
</emotion>

The fact that intensity is represented by an element makes it possible to add meta-information. For example, it is possible to express a high confidence that the intensity is low, but a low confidence regarding the emotion category, as shown as the last example in the description of confidence.

2.3 Meta-information

2.3.1 The confidence attribute

Confidence MAY be indicated separately for each of the Representations of emotions and related states. For example, the confidence that the <category> is assumed correctly is independent from the confidence that its <intensity> is correctly indicated.

Rooted in the tradition of statistics a confidence is usually given in an interval from 0 to 1, resembling a probability. This is an intuitive range opposing e.g. (logarithmic) score values. However, additionally a given yet limited number of discrete values may often be sufficient and more intuitive. Insofar, the confidence is a unipolar Scale value.

Legal values:

• a floating-point value from the interval [0;1];
• a fixed number of discrete values (see ISSUE note of the value attribute).

Examples:

In the following one simple example is provided for each element that MAY carry a confidence attribute.

The first example uses a verbal discrete scale value to indicate a very high confidence that surprise is the emotion to annotate.

<emotion>
    <category set="everydayEmotions" name="surprise" confidence="++"/> 
</emotion>

The next example illustrates using continuous scale values for confidence to indicate that the annotation of high arousal is probably correct, but the annotation of slightly positive valence may or may not be correct. Note that the choice of verbal vs. numeric scales between the emotion <dimension> and its confidence is totally independent, i.e. it is fully possible to use verbally specified emotion dimensions with numerically specified confidence (as in this example) or any other combination of verbal and numeric scales.

<emotion>
    <dimensions set="valenceArousal">
        <arousal value="++" confidence="0.9"/>
        <valence value="+" confidence="0.3"/>
    </dimensions>
</emotion>

Accordingly, an example of <appraisals> using verbal scales for both the appraisal dimensions themselves and for the confidence. Note that the confidence is always unipolar, but that some of the appraisal dimensions are bipolar.

<emotion>
    <appraisals set="Scherer_appraisals_checks">
        <novelty value="++" confidence="+"/>
        <intrinsic-pleasantness value="--" confidence="++"/>
    </appraisals>
</emotion>

The example for action tendencies demonstrates an alternative realisation: the example shows confidence as an attribute of the entire group of action tendencies; the confidence indicated (rather high) therefore applies to all action tendencies contained.

<emotion>
    <action-tendencies set="approachAvoidFlightFlight" confidence="0.8">
        <approach value="0.9"/>
        <avoid value="0.0"/>
        <flight-flight value="0.9"/>
    </action-tendencies>
</emotion>

Finally, an example for the case of <intensity>: A high confidence is named that the emotion has a low intensity.

<emotion>
    <intensity value="0.1" confidence="0.8"/>
</emotion>

Note that, as stated, obviously an emotional annotation can be a combination of some or all of the above, as in the following example: the intensity of the emotion is quite probably low, but if we have to guess, we would say the emotion is boredom.

<emotion>
    <intensity value="0.1" confidence="0.8"/>
    <category set="everydayEmotions" name="boredom" confidence="0.1"/>
</emotion>

2.3.2 The <modality> element

The <modality> element is used to annotate the modes in which the emotion is reflected. The mode attribute can contain values from a closed set of values, namely those specified by the set attribute. For example, a basic or default set could include values like face, voice, body and text. The mode and medium attributes can contain a list of space separated values, in order to indicate multimodal input or output.

Example:

In the following example the emotion is expressed through the voice, which is a modality included in the basicModalities set.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
    <emotion>
        <category set="everydayEmotions" name="satisfaction"/>
        <modality set="basicModalities" mode="voice"/>
    </emotion>
</emotionml>

In case of multimodal expression of an emotion, a list of space separated modalities can be indicated in the mode attribute, like in the following example in which the two values "face" and "voice" must be included in the basicModalities set.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
    <emotion>
        <category set="everydayEmotions" name="satisfaction"/>
        <modality set="basicModalities" mode="face voice"/>
    </emotion>
</emotionml>

See also the example at section 5.1.2 Automatic recognition of emotions

2.3.3 The <metadata> element

This element can contain arbitrary data (one option could be [RDF] data), either on a document global level or on a local "per annotation element" level.

Examples:

In the following example, the automatic classification for an annotation document was performed by a classifier based on Gaussian Mixture Models (GMM); the speakers of the annotated elements were of different German origins.

<emotionml>
    <metadata>
        <classifiers:classifier classifiers:name="GMM"/>
    </metadata>
    <emotion>
        <metadata>
            <origin:localization value="bavarian"/>
        </metadata>
        <category set="everydayEmotions" name="joy"/>
    </emotion>
    <emotion>
        <metadata>
            <origin:localization value="swabian"/>
        </metadata>
        <category set="everydayEmotions" name="sadness"/>
    </emotion>
</emotionml>

2.4 Links and time

2.4.1 The <link> element

A <link> element provides a link to media as a URI [RFC3986]. The semantics of links are described by the role attribute which MUST have one of four values:

• "expressedBy" indicates that the link points to observable behavior expressing the emotion. This is the default value if the role attribute is not explicitly stated;
• "experiencedBy" indicates that the link refers to the subject experiencing the emotion;
• "triggeredBy" indicates that the link identifies an emotion-eliciting event that caused the emotional reaction;
• "targetedAt" indicates that the link points to an object towards which an emotion-related action, or action tendency, is directed.

For resources representing a period of time, start and end time MAY be denoted by use of the optional attributes start and end that default to "0" and the time length of the media file, respectively.

There is no restriction regarding the number of <link> elements that MAY occur as children of <emotion>.

Example:

The following example illustrates the link to two different URIs having a different role with respect to the emotion: one link points to the emotion's expression, e.g. a video clip showing a user expressing the emotion; the other link points to the trigger that caused the emotion, e.g. another video clip that was seen by the person eliciting the expressed emotion. Note that no media sub-classing is used to differentiate between different media types as audio, video, text, etc. Several links may follow as children of one <emotion> tag, even having the same role: for example a video and physiological sensor data of the expressed emotion.

<emotion>
    <link uri="http:..." role="expressedBy"/>
    <link uri="http:..." role="triggeredBy"/>
</emotion>

2.4.2 Timestamps

2.4.2.1 Absolute time

date denotes the absolute timepoint at which an emotion or related state happened. This might be used for example with an "emotional diary" application. The attribute MAY be used with an <emotion> element, and MUST be a string in conformance to W3C datetime note based on the ISO-8601 standard.

Examples:

In the following example, the emotion category "joy" is annotated for the 23 November 2001, 14:36 hours UTC.

<emotion date="2001-11-23T14:36Z">
    <category set="everydayEmotions" name="joy"/>
</emotion>

2.4.2.2 Timing in media

start denotes the timepoint from which on an emotion or related state is displayed in a media file. It is optional and defaults to "0".

end denotes the timepoint at which an emotion or related state ends to be displayed in a media file. It is optional and defaults to the time length of the media file.

Both attributes MAY be used with a <link> element and MUST be a string in conformance to the SMIL clock value syntax.

Examples:

In the following example, the emotion category "joy" is displayed in a video file called "myVideo.avi" from the 3rd to the 9th second.

<emotion>
    <category set="everydayEmotions" name="joy"/>
    <link uri="myVideo.avi" start="3s" end="9s"/>
</emotion>

2.4.2.3 Timing reference

timeRefURI, timeRefAnchor and offsetToStart may be used to set the timing of an emotion or related state relative to the timing of another annotated element.

Examples:

In the following example, Fred is annotated as being sad on 23 November 2001 at 14:39 hours, three minutes later than the absolutely positioned reference element.

<emotion id="annasJoy" date="2001-11-23T14:36Z">
    <category set="everydayEmotions" name="joy"/>
</emotion>
<emotion id="fredsSadness" timeRefURI="#annasJoy"
    timeRefAnchor="end" offsetToStart="3min">
    <category set="everydayEmotions" name="sadness"/>
</emotion>

2.5 Scale values

Scale values are needed to represent content in dimension, appraisal and action-tendency elements, as well as in <intensity> and confidence.

Representations of scale values can vary along three axes:

• static vs. dynamic: a static, constant scale value is represented using the value attribute; for dynamic values, their evolution over time is expressed using the <trace> element.
• unipolar vs. bipolar: conceptually, a scale can represent concepts that vary from "nothing" to "a lot" (unipolar scales), or concepts that vary between two opposites, from "very negative" to "very positive" (bipolar scales).
• numeric vs. discrete: some use cases require scale values to be represented as continuous numeric values, whereas other use cases require a number of discrete values.

2.5.1 The value attribute

The value attribute represents a static scale value of the enclosing element.

Conceptually, each dimension, appraisal and action-tendency element is either unipolar or bipolar. The definition of a set of dimensions, appraisals or action tendencies MUST define, for each item in the set, whether it is unipolar or bipolar.

<intensity> is a unipolar scale.

Legal values:

• For unipolar scales, legal values are one of
  • a floating-point value from the interval [0;2], where usual values are in the range [0;1], and values in [1;2] can be used to represent exaggerated values;
  • a fixed number of discrete values (see ISSUE note below).
• For bipolar scales, legal values are one of
  • a floating-point value from the interval [-2;2], where usual values are in the range [-1;1], and values in [-2;-1] and [1;2] can be used to represent exaggerated values;
  • a fixed number of discrete values (see ISSUE note below).

Examples of the value attribute can be found in the context of the dimension, appraisal and action-tendency elements and of <intensity>.

2.5.2 The <trace> element

A <trace> element represents the time course of a numeric scale value. It cannot be used for discrete scale values.

The freq attribute indicates the sampling frequency at which the values listed in the samples attribute are given.

A <trace> MAY include a trace of the confidence alongside with the trace of the scale itself, in the samples-confidence attribute. If present, samples-confidence MUST use the same sampling frequency as the content scale, as given in the freq attribute. If the enclosing element contains a (static) confidence attribute, the <trace> MUST NOT have a samples-confidence attribute. In other words, it is possible to indicate either a static or a dynamic confidence for a given scale value, but not both.

NOTE: The <trace> representation requires a periodic sampling of values. In order to represent values that are sampled aperiodically, separate <emotion> annotations with appropriate timing information and individual value attributes may be used.

Examples:

The following example illustrates the use of a trace to represent an episode of fear during which intensity is rising, first gradually, then quickly to a very high value. Values are taken at a sampling frequency of 10 Hz, i.e. one value every 100 ms.

<emotion>
    <category set="everydayEmotions" name="fear"/>
    <intensity>
        <trace freq="10Hz" samples="0.1 0.1 0.15 0.2 0.2 0.25 0.25 0.25 0.3 0.3 0.35 0.5 0.7 0.8 0.85 0.85"/>
    </intensity>
</emotion>

The following example combines a trace of the appraisal "novelty" with a global confidence that the values represent the facts properly. There is a sudden peak of novelty; the annotator is reasonable certain that the annotation is correct:

<emotion>
    <appraisals set="someSetWithNovelty">
        <novelty confidence="0.75">
            <trace freq="10Hz" samples="0.1 0.1 0.1 0.1 0.1 0.7 0.8 0.8 0.8 0.8 0.4 0.2 0.1 0.1 0.1"/>
        </novelty>
    </appraisals>
</emotion>

In the following example, the confidence itself also changes over time. The observation is the same as before, but the confidence drops at the point where the novelty is rising, indicating some uncertainty where exactly the novelty appraisal is rising:

<emotion>
    <appraisals set="someSetWithNovelty">
        <novelty>
            <trace freq="10Hz" samples="0.1 0.1 0.1 0.1 0.1 0.7 0.8 0.8 0.8 0.8 0.4 0.2 0.1 0.1 0.1"
                    samples-confidence="0.7 0.7 0.7 0.4 0.3 0.3 0.3 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7"/>
        </novelty>
    </appraisals>
</emotion>

3 Defining vocabularies for representing emotions

EmotionML markup MUST refer to one or more vocabularies to be used for representing emotion-related states. Due to the lack of agreement in the community, the EmotionML specification does not preview a single default set which should apply if no set is indicated. Instead, the user MUST explicitly state the value set used.

3.1 Centrally defined default vocabularies

3.2 User-defined custom vocabularies

EmotionML markup makes no syntactic difference between referring to centrally-defined default vocabularies and referring to user-defined custom vocabularies. Therefore, one option to define a custom vocabulary is to create a definition XML file in the same way as it is done for the default vocabularies.

4 Conformance

4.1 EmotionML namespace

The EmotionML namespace is "http://www.w3.org/2009/10/emotionml". All EmotionML elements MUST use this namespace.

4.2 Use with other namespaces

The EmotionML namespace is intended to be used with other XML namespaces as per the Namespaces in XML Recommendation (1.0 [XML-NS10] or 1.1 [XML-NS11], depending on the version of XML being used).

4.3 Considerations regarding the validation of EmotionML documents

There is an intrinsic tension between the requirement of using plug-in vocabularies and the formal verification that a document is valid with respect to the specification. The issue has been pointed out repeatedly throughout this report, and is not yet solved. The following two subsections provide elements which may be part of a solution.

4.3.1 Use of QNAMES

A proposal under consideration is to use QNAMES to specify custom values for attributes. This solution allows to substitute the set attribute from many elements with a namespace declaration to be used as QNAME for the value of the attribute.

With this solution the attribute values are one or more white space separated QNames as defined in Section 4 of Namespaces in XML (1.0 [XML-NS10] or 1.1 [XML-NS11], depending on the version of XML being used).

When the attribute content is a QName, it is expanded into an expanded-name using the namespace declarations that are in scope for the relative element. Thus, each QName provides a reference to a specific item in the referred namespace.

In the example below, the QName "everydayEmotions:satisfaction" is the value of the name attribute and it will be expanded to the "satisfaction" item in the "http://www.example.com/everyday_emotion_catg_tags" namespace. The taxonomy for the everyday emotion categories has to be documented at the specified namespace URI.

<emotionml
  xmlns="http://www.w3.org/2009/10/emotionml" 
  xmlns:everydayEmotions="http://www.example.com/everyday_emotion_catg_tags">
    <emotion>
        <category name="everydayEmotions:satisfaction"/>
    </emotion>
</emotionml>

This solution allows for referencing different dictionaries depending on the namespace declarations. Moreover, the namespace qualification will make the new set of values unique. The drawbacks of this solution are the absence of a simple and clear way on how to validate the QNAME attribute values, and a more verbose syntax of the attribute contents.

4.3.2 Dynamic schema creation

A static schema document can only fully validate a language where the valid element names and attribute values are known at the time when the schema is written. For EmotionML, this is not possible because of the fundamental requirement to give users the option of using their own vocabularies.

The following is an idea for dynamically creating a schema document from a base schema and the vocabulary sets referenced in the document itself.

1. An EmotionML document refers to a centrally defined generic schema, one or more vocabularies which may be centrally defined or user-specific, and to a centrally defined XProc script;
2. the generic schema defines the legal structure of EmotionML documents using a schema language such as XML Schema, RelaxNG or Schematron, but using placeholders for concrete vocabulary items;
3. each vocabulary is defined using an XML format known to the XProc script;
4. the XProc script defines the workflow for validating an EmotionML document:
   1. from the EmotionML document, look up the generic schema and the custom vocabularies;
   2. through a suitable mechanism such as XSLT, merge the generic schema and the custom vocabularies into a (short-lived) custom schema;
   3. validate the EmotionML document using the custom schema in the usual way.

5 Examples (informative)

5.1 Examples of emotion annotation

5.1.1 Manual annotation of emotional material

Use case 1b-ii : Annotation of static images

An image gets annotated with several emotion categories at the same time, but different intensities.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
   <metadata>
      <media-type>image</media-type>
      <media-id>disgust</media-id>
      <media-set>JACFEE-database</media-set>
      <doc>Example adapted from (Hall & Matsumoto 2004) http://www.davidmatsumoto.info/Articles/2004_hall_and_matsumoto.pdf
      </doc>
   </metadata>
   <emotion>
       <category set="basicEmotions" name="Disgust"/>
       <intensity value="0.82"/>
   </emotion>
   <emotion>
       <category set="basicEmotions" name="Contempt"/>
       <intensity value="0.35"/>
   </emotion>
   <emotion>
       <category set="basicEmotions" name="Anger"/>
       <intensity value="0.12"/>
   </emotion>
   <emotion>
       <category set="basicEmotions" name="Surprise"/>
       <intensity value="0.53"/>
   </emotion>
</emotionml>

Use case 1c-i : Annotation of videos

Example 1: Annotation of a whole video: several emotions are annotated with different intensities.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
    <metadata>
        <media-type>video</media-type>
        <media-name>ed1_4</media-name>
        <media-set>humaine database</media-set>
        <coder-set>JM-AB-UH</coder-set>
    </metadata>
    <emotion>
        <category set="humaineDatabaseLabels" name="Amusement"/>
        <intensity value="0.52"/>
    </emotion>
    <emotion>
        <category set="humaineDatabaseLabels" name="Irritation"/>
        <intensity value="0.63"/>
    </emotion>
    <emotion>
        <category set="humaineDatabaseLabels" name="Relaxed"/>
        <intensity value="0.02"/>
    </emotion>
    <emotion>
        <category set="humaineDatabaseLabels" name="Frustration"/>
        <intensity value="0.87"/>
    </emotion>
    <emotion>
        <category set="humaineDatabaseLabels" name="Calm"/>
        <intensity value="0.21"/>
    </emotion>
    <emotion>
        <category set="humaineDatabaseLabels" name="Friendliness"/>
        <intensity value="0.28"/>
    </emotion>
</emotionml>

Example 2: Annotation of a video segment, where two emotions are annotated for the same timespan.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
    <metadata>
        <media-type>video</media-type>
        <media-name>ext-03</media-name>
        <media-set>EmoTV</media-set>
        <coder>4</coder>
    </metadata>
    <emotion>
        <category set="emoTV-labels" name="irritation"/>
        <intensity value="0.46"/>
        <link uri="ext03.avi" start="3.24s" end="15.4s">
    </emotion>
    <emotion>
        <category set="emoTV-labels" name="despair"/>
        <intensity value="0.48"/>
        <link uri="ext03.avi" start="3.24s" end="15.4s"/>
    </emotion>
</emotionml>

5.1.2 Automatic recognition of emotions

This example shows how automatically annotated data from three affective sensor devices might be stored or communicated.

It shows an excerpt of an episode experienced on 23 November 2001 from 14:36 onwards. Each device detects an emotion, but at slightly different times and for different durations.

The next entry of observed emotions occurs about 6 minutes later. Only the physiology sensor has detected a short glimpse of anger, for the visual and IR camera it was below their individual threshold so no entry from them.

For simplicity, all devices use categorical annotations and the same set of categories. Obviously it would be possible, and even likely, that different devices from different manufacturers provide their data annotated with different emotion sets.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
 ...
<emotion date="2001-11-23T14:36Z">
  <!--the first modality detects excitement.
      It is a camera observing the face. An URI to the database
      (a dedicated port at the server) is provided to access the
      video stream.-->
  <category set="everyday" name="excited"/>
  <modality medium="visual" mode="face"/>
  <link uri="http://www.example.com" start="26s" end="98s"/>
</emotion>

<emotion date="2001-11-23T14:36Z">
  <!--the second modality detects anger. It is an IR camera
      observing the face. An URI to the database (a dedicated port
      at the server) is provided to access the video stream.-->
  <category set="everyday" name="angry"/>
  <modality medium="infrared" mode="face"/>
  <link uri="http://www.example.com" start="23s" end="108s"/>
</emotion>

<emotion date="2001-11-23T14:36Z">
  <!--the third modality detects excitement again. It is a
      wearable device monitoring physiological changes in the
      body. An URI to the database (a dedicated port at the
      server) is provided to access the data stream.-->
  <category set="everyday" name="excited"/>
  <modality medium="physiological" mode="body"/>
  <link uri="http://www.example.com" start="19s" end="101s"/>
</emotion>

<emotion date="2001-11-23T14:42Z">
  <category set="everyday" name="angry"/>
  <modality medium="physiological" mode="body"/>
  <link uri="http://www.example.com" start="2s" end="6s"/>
</emotion>
 ...
</emotionml>

NOTE that handling of complex emotions is not explicitly specified. This example assumes that parallel occurrences of emotions will be determined on the time stamp.

NOTE that the used set of emotion descriptions needs to be specified for the document, see Defining vocabularies for representing emotions.

5.1.3 Generation of emotion-related system behavior

The following example describes various aspects of an emotionally competent robot.

<emotionml xmlns="http://www.w3.org/2009/10/emotionml">
    <metadata>
        <name>robbie the robot example</name>
    </metadata>

    <!-- Appraised value of incoming event -->
    <emotion>
        <modality mode="senses"/>
        <appraisals set="scherer_appraisals_checks">
            <novelty value="0.8" confidence="0.4"/>
            <intrinsic-pleasantness value="-0.5" confidence="0.8"/>
        </appraisals>
    </emotion>

    <!-- Robots current internal state configuration -->
    <emotion>
        <modality mode="internal"/>
        <dimensions set="arousal_valence_potency">
            <arousal value="0.3"/>
            <valence value="0.9"/>
            <potency value="0.8"/>
        </dimensions>
    </emotion>

    <!-- Robots output action tendencies -->
    <emotion>
        <modality mode="body"/>
        <action-tendencies set="myRobotActionTendencies">
            <charge-battery value="0.9"/>
            <seek-shelter value="0.7"/>
            <pickup-boxes value="-0.2"/>
        </action-tendencies>
    </emotion>

    <!-- Robots facial gestures -->
    <emotion>
        <modality mode="face"/>
        <category set="ekman_universal" name="joy"/>
        <link role="expressedBy" start="0" end="5s" uri="smile.xml"/>
    </emotion>
</emotionml>

5.2 Examples of possible use with other markup languages

One intended use of EmotionML is as a plug-in for existing markup languages. For compatibility with text-annotating markup languages such as SSML, EmotionML avoids the use of text nodes. All EmotionML information is encoded in element and attribute structures.

This section illustrates the concept using two existing W3C markup languages: EMMA and SSML.

5.2.1 Use with EMMA

EMMA is made for representing arbitrary analysis results; one of them could be the emotional state. The following example represents an analysis of a non-verbal vocalization; its emotion is described as most probably a low-intensity state, maybe boredom.

<emma:emma version="1.0" xmlns:emma="http://www.w3.org/2003/04/emma"
        xmlns="http://www.w3.org/2009/10/emotionml">
    <emma:interpretation start="12457990" end="12457995" mode="voice" verbal="false">

        <emotion>
            <intensity value="0.1" confidence="0.8"/>
            <category set="everydayEmotions" name="boredom" confidence="0.1"/>
        </emotion>

    </emma:interpretation>
</emma:emma>

5.2.2 Use with SSML

Two options for using EmotionML with SSML can be illustrated.

First, it is possible with the current draft version of SSML [SSML 1.1] to use arbitrary markup belonging to a different namespace anywhere in an SSML document; only SSML processors that support the markup would take it into account. Therefore, it is possible to insert EmotionML below, for example, an <s> element representing a sentence; the intended meaning is that the enclosing sentence should be spoken with the given emotion, in this case a moderately doubtful tone of voice:

<?xml version="1.0"?>
<speak version="1.1" xmlns="http://www.w3.org/2001/10/synthesis"
         xmlns:emo="http://www.w3.org/2009/10/emotionml"
         xml:lang="en-US">
    <s>
        <emo:emotion>
            <emo:category set="everydayEmotions" name="doubt"/>
            <emo:intensity value="0.4"/>
        </emo:emotion>

        Do you need help?
    </s>
</speak>

Second, a future version of SSML could explicitly preview the annotation of paralinguistic information, which could fill the gap between the extralinguistic, speaker-constant settings of the <voice> tag and the linguistic elements such as <s>, <emphasis>, <say-as> etc. The following example assumes that there is a <style> tag for paralinguistic information in a future version of SSML. The style could either embed an <emotion>, as follows:

<?xml version="1.0"?>
<speak version="x.y" xmlns="http://www.w3.org/2001/10/synthesis"
         xmlns:emo="http://www.w3.org/2009/10/emotionml"
         xml:lang="en-US">
    <s>
      <style>
        <emo:emotion>
            <emo:category set="everydayEmotion" name="doubt"/>
            <emo:intensity value="0.4"/>
        </emo:emotion>

        Do you need help?
      </style>
    </s>
</speak>

Alternatively, the <style> could refer to a previously defined <emotion>, for example:

<?xml version="1.0"?>
<speak version="x.y" xmlns="http://www.w3.org/2001/10/synthesis"
         xmlns:emo="http://www.w3.org/2009/10/emotionml"
         xml:lang="en-US">
    <emo:emotion id="somewhatDoubtful">
        <emo:category set="everydayEmotion" name="doubt"/>
        <emo:intensity value="0.4"/>
    </emo:emotion>

    <s>
        <style ref="#somewhatDoubtful">
            Do you need help?
        </style>
    </s>
</speak>

6 References

EMMA

EMMA: Extensible MultiModal Annotation markup language version 1.0, Michael Johnston, et al., Editors. World Wide Web Consortium, 11 December 2007.

Emotion Incubator Group

W3C Emotion Incubator Group, M. Schröder, E. Zovato, H. Pirker, C. Peter, F. Burkhardt, Editors. Final Report of the Emotion Incubator Group at the World Wide Web Consortium, 10 July 2007.

Emotion Markup Language Incubator Group

Elements of an EmotionML 1.0, M. Schröder, Editor. Final Report of the Emotion Markup Language Incubator Group at the World Wide Web Consortium, 20 November 2008.

EmotionML Requirements

Emotion Markup Language: Requirements with Priorities. F. Burkhardt and M. Schröder. W3C Incubator Group Report, 13 May 2008.

Gebhard, 2005

Gebhard, P. (2005). ALMA - A layered model of affect. In Proceedings of the Fourth International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS-05). Utrecht.

Ortony et al., 1988

Ortony, A., Clore, G. L., & Collins, A. (1988). The Cognitive Structure of Emotion. Cambridge, UK: Cambridge University Press.

RDF

RDF/XML Syntax Specification (Revised), Dave Beckett, Editor. World Wide Web Consortium, W3C Recommendation 10 February 2004.

RelaxNG

RELAX NG Specification, James Clark and Makoto Murata, Editors. OASIS, Committee Specification, 2001.

RFC2119

Key words for use in RFCs to Indicate Requirement Levels, S. Bradner, Editor. IETF RFC 2119, March 1997.

RFC3986

Uniform Resource Identifier (URI): Generic Syntax, T. Berners-Lee et al., Editors. IETF RFC 3986, January 2005.

Schematron

Information technology - Document Schema Definition Languages (DSDL) - Part 3: Rule-based validation - Schematron. ISO/IEC 19757-3:2006.

SMIL Clock Value Syntax

Synchronized Multimedia Integration Language (SMIL) Version 3.0 Time Clock Syntax, S. Mullender et al., Editors. W3C Proposed Recommendation, 6 October 2008.

SSML

Speech Synthesis Markup Language (SSML) Version 1.0, Daniel C. Burnett, et al., Editors. World Wide Web Consortium, W3C Recommendation, 7 September 2004.

SSML 1.1

Speech Synthesis Markup Language (SSML) Version 1.1, Daniel C. Burnett, et al., Editors. World Wide Web Consortium, W3C Candidate Recommendation, 7 November 2008.

XML-NS10

Namespaces in XML 1.0, Tim Bray et al., Editors. World Wide Web Consortium, W3C Recommendation, 16 August 2006.

XML-NS11

Namespaces in XML 1.1, Tim Bray et al., Editors. World Wide Web Consortium, W3C Recommendation, 2006.

XML Schema

XML Schema Part 1: Structures Second Edition, Henry S. Thompson et al., Editors. World Wide Web Consortium, W3C Recommendation, 2004.

XProc

XProc: An XML Pipeline Language, Norman Walsh et al., Editors. World Wide Web Consortium, Working Draft, 14 August 2008.

XSLT

XSL Transformations (XSLT) Version 1.0, James Clark, Editor. W3C Recommendation. 16 November 1999.

W3C Datetime Note

Date and Time Formats, M. Wolf and C. Wicksteed. W3C Note, 1997.

7 Acknowledgments

The authors wish to acknowledge the contributions by all members of the Emotion Markup Language Incubator Group and the Emotion Incubator Group, in particular the following persons (in alphabetic order):

• Andrew Breen, Nuance Communications
• Dylan Evans, University College Cork
• Nestor Garay Vitoria, University of the Basque Country
• Alain Giboin, INRIA Sophia Antipolis
• Bill Jarrold, SRI International
• Kostas Karpouzis, Image, Video and Multimedia Systems Lab (IVML-NTUA)
• Myriam Lamolle, University of Paris VIII
• Jean-Claude Martin, CNRS
• Hannes Pirker, Austrian Research Institute for Artificial Intelligence
• Björn Schuller, Technische Universität München
• Jianhua Tao, Chinese Academy of Sciences
• Ian Wilson, Emotion AI
• Idoia Zearreta, University of the Basque Country