ISO 24617-6:2016

SLOVENSKI STANDARD
01-oktober-2018
Upravljanje z jezikovnimi viri - Ogrodje za semantično označevanje (SemAF) - 6.
del: Načela semantičnega označevanja (načela SemAF)
Language resource management -- Semantic annotation framework -- Part 6: Principles
of semantic annotation (SemAF Principles)
Gestion des ressources linguistiques -- Cadre d'annotation sémantique -- Partie 6:
Principes d'annotation sémantique (SemAF Principes)
Ta slovenski standard je istoveten z: ISO 24617-6:2016
ICS:
01.020 Terminologija (načela in Terminology (principles and
koordinacija) coordination)
01.140.20 Informacijske vede Information sciences
35.240.30 Uporabniške rešitve IT v IT applications in information,
informatiki, dokumentiranju in documentation and
založništvu publishing
2003-01.Slovenski inštitut za standardizacijo. Razmnoževanje celote ali delov tega standarda ni dovoljeno.

INTERNATIONAL ISO
STANDARD 24617-6
First edition
2016-02-01
Language resource management —
Semantic annotation framework —
Part 6:
Principles of semantic annotation
(SemAF Principles)
Gestion des ressources linguistiques — Cadre d’annotation
sémantique —
Partie 6: Principes d’annotation sémantique (SemAF Principes)
Reference number
©
ISO 2016
© ISO 2016, Published in Switzerland
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized otherwise in any form
or by any means, electronic or mechanical, including photocopying, or posting on the internet or an intranet, without prior
written permission. Permission can be requested from either ISO at the address below or ISO’s member body in the country of
the requester.
ISO copyright office
Ch. de Blandonnet 8 • CP 401
CH-1214 Vernier, Geneva, Switzerland
Tel. +41 22 749 01 11
Fax +41 22 749 09 47
copyright@iso.org
www.iso.org
ii © ISO 2016 – All rights reserved

Contents Page
Foreword .iv
1 Scope . 1
2 Terms and definitions . 1
3 Purpose and motivation . 2
3.1 Purpose . 2
3.2 Motivation . 2
4 Overview . 3
5 Annotation principles and requirements . 4
5.1 Principles inherited from the Linguistic Annotation Framework . 4
5.2 Other general annotation principles . 5
5.3 Principles specific to semantic annotation . 5
6 The methodological basis of SemAF . 7
6.1 Steps in the design of an annotation scheme . 7
6.2 Metamodels . 8
6.3 Abstract syntax, concrete syntax and semantics .10
6.4 Steps forward and feedback in the design process .12
6.5 Optional elements in an annotation scheme .14
7 Overlaps between annotation schemes .15
7.1 Semantic and terminological consistency .15
7.2 Spatial and temporal relations as semantic roles .15
7.3 Events .17
7.4 Discourse relations in dialogue .18
8 Semantic phenomena that cut across annotation schemes .18
8.1 Ubiquitous semantic phenomena .18
8.2 Quantification .18
8.3 Quantities and measures .19
8.4 Negation, modality, factuality, and attribution .20
8.5 Modification and qualification .21
8.5.1 Modification and quantification .21
8.5.2 Qualification .22
8.5.3 Other issues .23
Annex A (informative) An approach to the annotation of quantification in natural language .24
Bibliography .28
Foreword
ISO (the International Organization for Standardization) is a worldwide federation of national standards
bodies (ISO member bodies). The work of preparing International Standards is normally carried out
through ISO technical committees. Each member body interested in a subject for which a technical
committee has been established has the right to be represented on that committee. International
organizations, governmental and non-governmental, in liaison with ISO, also take part in the work.
ISO collaborates closely with the International Electrotechnical Commission (IEC) on all matters of
electrotechnical standardization.
The procedures used to develop this document and those intended for its further maintenance are
described in the ISO/IEC Directives, Part 1. In particular the different approval criteria needed for the
different types of ISO documents should be noted. This document was drafted in accordance with the
editorial rules of the ISO/IEC Directives, Part 2 (see www.iso.org/directives).
Attention is drawn to the possibility that some of the elements of this document may be the subject of
patent rights. ISO shall not be held responsible for identifying any or all such patent rights. Details of
any patent rights identified during the development of the document will be in the Introduction and/or
on the ISO list of patent declarations received (see www.iso.org/patents).
Any trade name used in this document is information given for the convenience of users and does not
constitute an endorsement.
For an explanation on the meaning of ISO specific terms and expressions related to conformity
assessment, as well as information about ISO’s adherence to the WTO principles in the Technical
Barriers to Trade (TBT) see the following URL: Foreword - Supplementary information.
The committee responsible for this document is ISO/TC 37, Terminology and other language and content
resources, Subcommittee SC 4, Language resource management.
ISO 24617 consists of the following parts, under the general title Language resource management —
Semantic annotation framework (SemAF):
— Part 1: Time and events (SemAF-Time, ISOTimeML)
— Part 2: Dialogue acts (SemAF-Dacts)
— Part 4: Semantic roles (SemAF-SR)
— Part 5: Discourse structures (SemAF-DS) [Technical Specification]
— Part 6: Principles of semantic annotation (SemAF Principles)
— Part 7: Spatial information (ISOspace)
The following parts are in preparation:
— Part 8: Semantic relations in discourse (SemAF DR-core)
— Part 9: Reference (ISOref)
iv © ISO 2016 – All rights reserved

INTERNATIONAL STANDARD ISO 24617-6:2016(E)
Language resource management — Semantic annotation
framework —
Part 6:
Principles of semantic annotation (SemAF Principles)
1 Scope
This part of ISO 24617 specifies the approach to semantic annotation characterizing the ISO Semantic
annotation framework (SemAF). It outlines the SemAF strategy for developing separate annotation
schemes for certain classes of semantic phenomena, aiming in the long term to combine these into a
single, coherent scheme for semantic annotation with wide coverage. In particular, it sets out the
notions of both an abstract and a concrete syntax for semantic annotations, mirroring the distinction
between annotations and representations that is made in the ISO Linguistic Annotation Framework.
It describes the role of these notions in relation to the specification of a metamodel and a semantic
interpretation of annotations, with a view to defining a well-founded annotation scheme.
This part of ISO 24617 also provides guidelines for dealing with two issues regarding the annotation
schemes defined in SemAF-parts: a) conceptual and terminological inconsistencies that may arise due
to overlaps between annotation schemes and b) the treatment of semantic phenomena that cut across
SemAF-parts, such as negation, modality and quantification. Instances of both issues are identified, and
in some cases, direction is given as to how they may be tackled.
2 Terms and definitions
For the purposes of this document, the following terms and definitions apply.
NOTE In addition, the terms ‘event’ and ‘eventuality’ are used (as synonyms) as defined in ISO 24617-1 as
something that can be said to obtain or hold true, to happen or to occur.
2.1
primary data
electronic representation of text or communicative behaviour
EXAMPLE Digital representations of text, transcriptions of speech, gestures or multimodal dialogue.
Note 1 to entry: ISO 24612 defines primary data as the ‘electronic representation of language data’. This definition
is unsatisfactory for this part of ISO 24617 as semantic annotation may relate to non-verbal or multimodal data,
such as stretches of spoken dialogue with accompanying gestures and facial expressions, and even gestures
and/or facial expressions without any accompanying speech.
2.2
annotation
linguistic information added to primary data (2.1), independent of its representation
[SOURCE: ISO 24612:2012, 2.3]
2.3
semantic annotation
annotation (2.2) which contains information about the meaning of a segment or region of primary data
(2.1)
2.4
metamodel
schematic representation of the concepts that are used in the analysis and description of the phenomena
covered in annotations (2.2) and of the relationships between them
3 Purpose and motivation
3.1 Purpose
The purpose of this part of ISO 24617 is to provide support for the establishment of a consistent and
coherent set of international standards for semantic annotation within the Semantic Annotation
Framework (SemAF). It aims to do so in three ways.
First, by making explicit which basic principles underlie the approach that has been followed in
defining international standards in the SemAF parts that have been published so far (ISO 24617-1 and
ISO 24617-2, ISO 24617-4 and ISO 24617-7), and in parts that are close to publication (ISO 24617-6)
or in preparation (ISO 24617-8). This approach provides the Semantic Annotation Framework with
methodological coherence and helps to ensure mutual consistency between existing, developing, and
future SemAF parts.
Second, by identifying overlaps between SemAF parts and indicating how such overlaps may be dealt
with. Examples are the occurrence of temporal and spatial relations among semantic roles and of
discourse relations between dialogue acts.
Third, by identifying common issues that arise in various parts of SemAF (they are only partly covered
in these parts, if they are covered at all) and, where possible, by giving directions as to how these issues
may be tackled. Examples of such issues are polarity, modality, quantification, measures, qualification,
veridicity, attribution and non-literal language use.
3.2 Motivation
Semantic annotation enhances primary data with information about their meaning. The state of the art
in computational semantics makes it unlikely that a single existing formalism for annotating semantic
information would receive wide support from researchers and developers. Moreover, semantic
annotation tasks often have the limited aim of annotating certain specific semantic phenomena,
such as semantic roles, discourse relations or coreference relations, rather than annotating the full
meaning of stretches of primary data. A strategy was therefore adopted in ISO TC 37/SC 4 to devise the
SemAF standards in different parts, with separate annotation schemes for those classes of semantic
phenomenon for which the state of the art would justify the establishment of annotation standards;
these schemes could be extended and combined over time, growing into a wide-coverage framework for
semantic annotation.
This ‘crystal growth’ strategy has contributed significantly to the progress made in establishing
standardized annotation concepts and schemes supporting the development of interoperable resources,
but it also entails certain risks:
a) the annotation schemes defined in different SemAF parts are not necessarily mutually consistent,
especially in the case of overlaps in scope;
b) it may not be possible to combine the schemes, defined in different parts, into a coherent single
scheme with a wider coverage if they incorporate different views or employ different methodologies;
c) some semantic phenomena do not belong to the scope of any SemAF parts but cannot be disregarded
entirely in some parts, and this may result in these phenomena being unsatisfactorily treated.
The methodological principles and guidelines provided in this part of ISO 24617 are designed to
minimize these risks.
2 © ISO 2016 – All rights reserved

With regard to the issue of mutual consistency between SemAF parts, it may be noted that ISO 24617-1
for annotating time and events and ISO 24617-2 for annotating dialogue acts are concerned with
sufficiently distinct kinds of semantic information to allow their definitions to be established
independently. Other SemAF parts, such as those concerned with semantic roles, with relations in
discourse and with spatial information show a certain amount of overlap in the information that they
aim to capture, and the question therefore arises: can we ensure that the annotation schemes, defined
in these parts, are mutually consistent?
Mutual consistency of SemAF parts relates to the possible integration of annotation schemes defined
in different parts. For example, it would be desirable to use the ISO 24617-1 scheme (“ISO-TimeML”) for
annotating time and events in combination with the ISO 24617-4 scheme for semantic roles, thereby
annotating coherently not only the events identified in the data with their temporal properties, but
also the way in which these events are related to their participants. Integrating these annotations with
those of spatial information, using the ISO 24617-7 scheme for spatial information, would be another
plausible and desirable step, given that time and space are intertwined with concepts relating to motion
and velocity. More generally, the integration of SemAF parts would greatly enhance the significance
of the individual parts; in the end, SemAF’s ‘crystal growth’ strategy of SemAF is only really useful
if the annotation schemes defined in the various parts can grow into a single scheme with a wide
coverage of semantic phenomena. Only then can it effectively support such applications as text-based
question answering or extracting semantic information from text, and form the basis for automatically
recognizing semantic phenomena by means of machine-learning techniques. Clearly, this is only possible
if the annotation schemes are mutually consistent (e.g. they use the same classification of event types),
and are coherent whether, for example, temporal and spatial objects are viewed as event participants
or as the circumstances of an event.
With regard to the risk of unsatisfactory partial treatments of phenomena that are not among the
core issues of any (current) SemAF part, it should be noted that some of these phenomena cut across
several of these parts and are important for semantics-driven applications. Negation, or more generally
negative polarity, and quantification are two cases in point. Given that the aim in ISO-TimeML, for
instance, is to support the annotation of events, of their relation to time, and of the temporal relations
among temporal objects, it is desirable to be able to deal with sentences like the following:
(1) John teaches every Monday.
(2) Mary called twice this morning.
(3) John rang home twice a day.
Sentence (1) is about a set of “teach” events, each of which is related to a different element of the set of
temporal objects that are called “Monday”, so this is a case of quantification involving two sets, a set of
events and sets of days. Similarly, sentence (2) about a set of two “call” events, both related to the same
period of time. Sentence (3) is about a set of events and their frequency of occurrence.
In order to deal with such phenomena, ISO-TimeML has certain provisions for annotating quantification,
[13]
but they are not really adequate and do not generalize to cases of quantification where no events
are involved.
4 Overview
The ISO efforts aiming to develop standards for semantic annotation rest on certain basic principles,
some of which have been laid out by Reference [14] as requirements for semantic annotation, and
have been developed further in Reference [5]; others have been formulated as general principles for
linguistic annotation and are part of the ISO Linguistic Annotation Framework (LAF; see Reference [18]
and ISO 24623-1). The two sets of principles and requirements are considered in Clause 5.
The three kinds of risk associated with the SemAF ‘crystal growth’ strategy that have been identified
above correspond to the following issues of consistency and completeness that arise in the design of
semantic annotation schemes within the SemAF framework.
Consistency among annotation schemes:
— methodological consistency: the same basic approach is followed with respect to the distinction
between abstract and concrete syntax and their interrelation, and with respect to their semantics;
— conceptual consistency: different schemes are based on compatible underlying views and ontological
assumptions regarding their basic concepts, as reflected in metamodels (e.g. verbs are viewed as
denoting states or events, rather than relations);
— terminological consistency: terms that occur in different annotation schemes have the same meaning
in every scheme and the same term is used across annotation schemes to indicate the same concept.
Completeness of a set of annotation schemes: the combination of multiple annotation schemes leads to
a scheme that
— covers a wide range of semantic phenomena,
— does not have significant gaps when covering the semantic phenomena that it aims to cover, and
— deals in a satisfactory way with semantic phenomena that cut across the combined schemes but
which do not belong to the core phenomena that any of the combined schemes are designed to cover.
Clause 5 describes the methodological framework for defining annotation schemes in SemAF parts,
thereby ensuring methodological consistency. Clause 6 discusses conceptual and terminological
consistency issues that arise due to overlaps between SemAF parts, while Clause 7 identifies issues of
completeness regarding the annotation of semantic phenomena that cut across existing SemAF parts.
5 Annotation principles and requirements
5.1 Principles inherited from the Linguistic Annotation Framework
The annotation of semantic information when using SemAF inherits the principles for linguistic
annotation as formulated in LAF. These principles are often of a very general nature; they include the
principle that relevant segments of primary data are referred to in a uniform and TEI-compliant way,
and the principle that different layers of annotation over the primary data can co-exist by using stand-
off annotation and a uniform way of cross-referencing between layers.
The latter principle, which concerns the distinction of layers of annotation enabled by a stand-off
representation format, is of particular relevance for SemAF because it allows different annotation
layers to be used for different types of semantic information; for example, one layer could be used for
the annotation of events, time and space, and another one could be used to annotate semantic roles.
In principle, this allows for the use not only of layers that are not mutually consistent, but also of
alternative annotations that employ different annotation schemes for the same phenomena. However,
the SemAF ‘crystal growth’ strategy is designed to ensure that the annotation schemes for the various
types of semantic information can grow into a coherent annotation scheme for a wide range of semantic
phenomena, and it is therefore highly undesirable to have inconsistencies between annotation layers
concerned with different SemAF parts.
[18]
Also of particular relevance for SemAF is the distinction between ‘annotations’ and ‘representations’.
An annotation is any item of linguistic information that is added to primary data, independently of any
particular representation format. A representation is a format into which an annotation is rendered,
for example as an XML expression. ISO standards are assumed to be defined at the level of annotations,
rather than representations. The fundamental distinction between annotations and representations has
prompted the development of a methodology for developing semantic annotation schemes that draws
a distinction between the ‘abstract syntax’ of annotations and the ‘concrete syntax’ of representations.
This methodology is described in Clause 6.
4 © ISO 2016 – All rights reserved

5.2 Other general annotation principles
In addition to the principles that SemAF inherits from LAF, other general principles for designing
annotation schemes (in particular as part of an ISO standard) are worth mentioning; most of these
emerged during the development of the ISO 24617-2 standard for dialogue act annotation.
a) Theoretical validity: Annotation standards should consolidate existing knowledge and
accordingly should be firmly rooted in theoretical studies of the annotated phenomena. Any concept
that may occur in annotations according to the standard should therefore be well established in the
scientific literature.
b) Empirical validity: Annotation standards are designed to be useful for annotating corpora of
recorded empirical data; the annotation scheme defined in a standard should not therefore include
theoretical constructs that are not found in such corpora, but only concepts that correspond to
phenomena that are observed in empirical data.
c) Learnability: For an annotation scheme to be useful in the construction of annotated language
resources, it should be possible both for human annotators and for automatic annotation systems
to effectively learn how to apply the scheme with acceptable precision.
d) Generalizability: ISO standards should not be restricted in their applicability to particular
languages, subject domains or applications.
e) Extensibility: While ISO standard annotation schemes are designed to be language-independent,
domain-independent and application-independent, some applications and some languages may
require specific concepts that are not relevant in other applications or languages. Annotation
schemes should therefore be open, that is to say, they should allow extension with language-
specific, domain-specific and application-specific concepts.
f) Completeness: An annotation standard is designed to provide a good coverage of the phenomena of
which it is designed to enable the annotation; the set of concepts defined in an annotation standard
should, in that sense, be complete.
g) Variable granularity: One way to achieve good coverage is to include annotation concepts of a
high level of generality and which cover many specific instances. Since an annotation scheme
which uses only very general concepts would not be optimally useful, this leads to the principle
that annotation schemes should include concepts with different levels of granularity. This is also
beneficial for its interoperability, as it provides more possibilities for conversion between existing
annotation schemes and the standard scheme.
h) Compatibility: In order to enable mappings between alternative annotation schemes and thereby
contribute to the interoperability of annotated resources, concepts that are commonly found in
existing annotation schemes should preferably be included in an annotation standard.
5.3 Principles specific to semantic annotation
The idea behind annotating a text, which dates from long before the digital era, is to add information to a
primary text in order to support its understanding. The semantic annotation of digital source texts has
a similar purpose, namely to support the understanding of the text by humans, as well as by machines.
An annotation that does not add any information would therefore seem to make little sense, but the
[39]
following example of the annotation of a temporal expression using TimeML seems to do just that:
NOTE 1 For simplicity, the annotations of the events that are mentioned in the previous sentence is
suppressed here.
(4)

The CEO announced that he would resign as of

the first of December 2008


In this annotation, the subexpression
adds to the noun phrase “the first of December 2008” the information that this phrase describes: the
date 2008-12-01.This does not add any information; rather, it paraphrases the noun phrase in TimeML.
This could be useful if the expression in the annotation language had a well-specified semantics that
could be used directly by computer programs for applications like information extraction and question
answering. Unfortunately, TimeML does not have a semantics.
NOTE 2 It would be very simple to provide a semantics for the XML fragment shown here but it would be very
difficult to do so for the whole of TimeML. See also 7.3.
A case where the annotation of a date as in the above example does add something is (5). From the
utterance “Mr Brewster called a staff meeting today”, it is impossible to know the date on which the event
that is mentioned took place; in this case, the annotation, which is identical to (4), would be informative.
NOTE 3 Note that the examples of TimeML annotations shown here are “old-fashioned” in the sense that the
TIMEX3 element is wrapped around the annotated string. Modern annotation methods (e.g. in ISO-TimeML) use
stand-off representations.
(5) Mr Brewster called a staff meeting today.

Mr Brewster called a staff meeting

today


The examples in (4) and (5) illustrate two different functions that semantic annotations may have:
interpreting a natural language expression by recoding it in a formal annotation language with a well-
defined semantics, and adding context information, in order to allow the interpretation of context-
dependent expressions.
A third function that semantic annotations may have is to make explicit how certain parts of an
utterance are semantically related; for example, this is the function of semantic role labelling and of
indicating semantic relations between sentences in a discourse when no such relation is mentioned.
Note that the first function presupposes annotations to have a well-defined semantics; the other two
functions do not presuppose this, but since semantic annotations in digital corpora are typically
designed to support interpretation and inference, a desideratum for all functions that a semantic
annotation may have is that it has a well-defined semantics. Semantic annotations, like other linguistic
annotations, may also serve the purpose of supporting linguistic research, such as identifying syntactic
and semantic patterns in sentences and texts.
These considerations lead to the following two principles for semantic annotation:
— Semantic additivity: semantic annotations add semantic information to source data, or re-express
certain source data in a formal representation.
— Semantic adequacy: semantic annotations should have a well-defined semantics, making the
annotations machine-interpretable.
6 © ISO 2016 – All rights reserved

6 The methodological basis of SemAF
6.1 Steps in the design of an annotation scheme
An annotation scheme determines which information may be added to primary data and how that
information is expressed. The design of an annotation scheme from scratch should begin with a
conceptual analysis of the information that the annotations should capture. This analysis identifies
the concepts that form the building blocks of annotations and specifies how these blocks may be
used to build annotation structures. This may be broken down into two steps, the first establishing a
conceptual view of the phenomena to be annotated, the second an articulation of this view in the form
of a formal specification of categories of entities and relations, and of how annotation structures can be
built up from elements in these categories. The first of these steps corresponds to what is known in ISO
projects as the establishment of a ‘metamodel’, that is to say, the expression of a conceptual view of the
phenomena to be annotated in the form of a (UML) diagram. The formal specification produced in the
second step constitutes the abstract syntax of an annotation language.
While these two steps make explicit what information can be captured in the annotations, they do not
concern the use of representation formats, such as XML strings, logical formulas, graphs or feature
structures; the abstract syntax defines the specification of information in terms of set-theoretic
structures, called annotation structures. An annotation structure is a collection of two kinds of structure,
entity structures and link structures. An entity structure contains semantic information about a segment
of primary data and is formally a pair consisting of a markable, which refers to a segment of
primary data, and of certain semantic information. A link structure contains information about the way
two segments of primary data are semantically related; for example, in semantic role annotation a link
structure is a triple where e is an entity structure that contains information about an event,
1 2 i 1
e is an entity structure that contains information about a participant in the event, and R is a relation
2 i
denoting a semantic role.
The third step in the definition of an annotation scheme is the specification of the meaning of the structures
defined by the abstract syntax, that is to say, the specification of semantics for annotation structures.
The fourth and final step is the definition of a format for representing annotation structures, such as a
serialization in XML. The semantics of an expression in this format is that of the annotation structure
that it represents. In sum, this approach to designing annotation schemes consists of four steps, with as
many feedback loops as may be needed.
The method consisting of these four steps is called ‘CASCADES’: Conceptual analysis, Abstract syntax,
Semantics, and Concrete syntax for Annotation language DESign. Figure 1 visualizes the CASCADES
method, of which the central concept, that of an abstract syntax for annotations with the specification
of a semantics for annotation structures (rather than for their representations in a particular format),
was introduced in Reference [5].
The CASCADES method is useful for enabling a systematic design process, in which due attention is
given to the conceptual and semantic choices on which more superficial decisions, such as the choice
of particular XML attributes and values, should be based. Apart from supporting the design of an
annotation scheme starting from scratch, the method also provides support for improving an existing
annotation scheme. This support consists not only of the distinction of four well-defined design steps,
but also of procedures and guidelines for taking these steps. These procedures are outlined in 6.3.
But before that, the use of metamodels, resulting from the conceptual analysis phase of designing an
annotation scheme, is discussed in 6.2.
Figure 1 — Steps and feedback loops in the CASCADES method
6.2 Metamodels
A metamodel of an annotation standard is a schematic representation of the relations between
the concepts that are important in the analysis and description of the phenomena covered in the
annotations. Over the years, two slightly different notions of a metamodel have been used in ISO
projects, namely the following:
— A: as a representation of the relations between the most important concepts that are mentioned in
the document in which the standard is defined;
— B: as a representation of the relations between the concepts denoted by terms that occur in
annotations, according to the standard.
Type-A metamodels may help nontechnical readers to have a better understanding of annotation
schemes; those of type-B are a visual representation of the abstract syntax of the annotations according
to the scheme and may help readers see at a glance what information the annotations may contain.
(Note that a type-A metamodel may have a type-B metamodel as a proper part.) For example, Figure 2
shows the metamodel of ISO 24617-4 for semantic role labelling. This metamodel reflects the view that
semantic roles are relations between eventualities and their participants; different roles correspond
to different ways in which a participant is involved in an eventuality. For example, in “Chris wrote a
poem”, a poem participates in an event in the Result role (i.e. created by the event); in “Chris revised the
poem”, the poem is involved in the Patient role (i.e. affected by the event); and in “Chris read a poem”, a
poem participates in the Theme role (i.e. not affected by the event). The annotation of semantic roles
therefore involves the use both of eventualities, most often corresponding to verbs, and of participants,
typically expressed by noun phrases. The metamodel associates eventualities and participants with
markables, whose source is a stretch of text in the primary data. The participants in an eventuality are
typically individuals, but may also be properties, numbers, quantities, propositions, sets of individuals
or embedded eventualities, as the examples in (6) illustrate.
(6)  a. He painted his house blue.
b. The birth of the twins increased the number of children to four.
c. The weight of this suitcase exceeds 20 kilos.
8 © ISO 2016 – All rights reserved

d. John believes he will get the money together.
e. Marie, Jill and Chris got together.
f. Eric and Nicole went to the concert.
Figure 2 — Metamodel for semantic role annotation (ISO 24617-4)
The term ‘entity’ is used in ISO 24617-4 to denote anything except an eventuality; the fact that a
participant in an eventuality can be an ‘entity’ of any sort as well as an (embedded) eventuality, as in (6f),
is reflected in the metamodel by the “entityFunctionsAs” relation between entities and participants,
and the “eventualityMayFunctionAs” relation between eventualities and participants. Many approaches
to semantic role labelling (e.g. PropBank, FrameNet and VerbNet) make use of “eventuality frames” that
specify the set of semantic roles expected in a type of eventuality, and take the eventuality frame as a
whole into consideration when determining the choice of individual semantic roles. For that reason,
eventuality frames have also been included in the metamodel.
Is this a type-A metamodel or a type-B metamodel? The following is an example of an annotation
representation according to ISO 24617-4 (where m1 is a markable that refers to “The soprano”; m2 is a
markable that refers to “sang”, and m3 is a markable that refers to “an aria”):
(7)  a. The soprano sang an aria.
b.





In this annotation, eventualities are directly related to entities by means of an element like event=”#e1” participant=”#x1” semRole=”agent”/>. There is no ‘participant’ object,
related to an entity via an ‘entityFunctionsAs relation’, as there is in the metamodel. It follows that
this metamodel is not just a visual representation of the (abstract) syntax of annotation structures, but
rather a type-A model, since it has elements that do not appear in annotations. This is corroborated by
the fact that eventuality frames do not occur as such in annotations; eventuality frames may be used in
the process of semantic role labelling, but are not part of the resulting annotations.
6.3 Abstract syntax, concrete syntax and semantics
The abstract syntax of an annotation scheme, as mentioned above, specifies the information in
annotations in terms of set-theoretical structures such as pairs and triples, like the triple
1 2 i
which relates the two arguments e and e through the relation R . More generally, these structures
1 2 i
are n-tuples of elements that are either basic concepts taken from a store of basic concepts called the
‘conceptual inventory’ of the abstract syntax specification, or n-tuples of such structures.
A concrete syntax specifies a representation format for annotation structures, such as the XML format
illustrated in (7), where a triple like is represented by a sequence of three XML elements,
1 2 i
of which the element
represents the relation and the other two elements represent two entity structures.
A representation format for annotation structures should ideally give an exact expression of the
information contained in annotation structures. A concrete syntax, defining a representation format
for a given abstract syntax, is said to be ideal if it has the following properties:
— completeness: every annotation structure defined by the abstract syntax can be represented by an
expression defined by the concrete syntax;
— unambiguity: every representation defined by the concrete syntax is the rendering of exactly one
annotation structure defined by the abstract syntax.
The representation format defined by an ideal concrete syntax is called an ideal representation format.
Due to its completeness, an ideal concrete syntax CS defines a function F from annotation structures
i i
−1
to CS -representations, and due to its ‘unambiguity’, there is also an inverse function F from CS -
i i i
representations to annotation structures. If I is the interpretation function defining the semantics
a
of the abstract syntax, then the meaning of a representation r in the ideal format CS is defined by
k
−1
I (F (r)). It follows that, for any two ideal representation formats CS and CS , there is a meaning-
a k i j
preserving conversion C defined by:
ij
−1
(8)  C (r) = F [F (r)]
ij j i
This mapping is meaning-preserving due to the fact that the meaning of a representation is the meaning
of the annotation structure that it encodes. Since this holds for any ideal concrete syntax, it follows that
any two ideal representation formats are semantically equivalent.
Figure 3 visualizes the relations between abstract syntax, semantics, and alternative ideal concrete-
syntax specifications. It is immediately clear that a given representation r defined by concrete syntax
CS can be converted into a semantically equivalent representation r’ in the representation format CS
i j
−1
by first applying the function F in order to determine the annotation structure which it encodes, and
i
applying to that
...

INTERNATIONAL ISO
STANDARD 24617-6
First edition
2016-02-01
Language resource management —
Semantic annotation framework —
Part 6:
Principles of semantic annotation
(SemAF Principles)
Gestion des ressources linguistiques — Cadre d’annotation
sémantique —
Partie 6: Principes d’annotation sémantique (SemAF Principes)
Reference number
©
ISO 2016
© ISO 2016, Published in Switzerland
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized otherwise in any form
or by any means, electronic or mechanical, including photocopying, or posting on the internet or an intranet, without prior
written permission. Permission can be requested from either ISO at the address below or ISO’s member body in the country of
the requester.
ISO copyright office
Ch. de Blandonnet 8 • CP 401
CH-1214 Vernier, Geneva, Switzerland
Tel. +41 22 749 01 11
Fax +41 22 749 09 47
copyright@iso.org
www.iso.org
ii © ISO 2016 – All rights reserved

Contents Page
Foreword .iv
1 Scope . 1
2 Terms and definitions . 1
3 Purpose and motivation . 2
3.1 Purpose . 2
3.2 Motivation . 2
4 Overview . 3
5 Annotation principles and requirements . 4
5.1 Principles inherited from the Linguistic Annotation Framework . 4
5.2 Other general annotation principles . 5
5.3 Principles specific to semantic annotation . 5
6 The methodological basis of SemAF . 7
6.1 Steps in the design of an annotation scheme . 7
6.2 Metamodels . 8
6.3 Abstract syntax, concrete syntax and semantics .10
6.4 Steps forward and feedback in the design process .12
6.5 Optional elements in an annotation scheme .14
7 Overlaps between annotation schemes .15
7.1 Semantic and terminological consistency .15
7.2 Spatial and temporal relations as semantic roles .15
7.3 Events .17
7.4 Discourse relations in dialogue .18
8 Semantic phenomena that cut across annotation schemes .18
8.1 Ubiquitous semantic phenomena .18
8.2 Quantification .18
8.3 Quantities and measures .19
8.4 Negation, modality, factuality, and attribution .20
8.5 Modification and qualification .21
8.5.1 Modification and quantification .21
8.5.2 Qualification .22
8.5.3 Other issues .23
Annex A (informative) An approach to the annotation of quantification in natural language .24
Bibliography .28
Foreword
ISO (the International Organization for Standardization) is a worldwide federation of national standards
bodies (ISO member bodies). The work of preparing International Standards is normally carried out
through ISO technical committees. Each member body interested in a subject for which a technical
committee has been established has the right to be represented on that committee. International
organizations, governmental and non-governmental, in liaison with ISO, also take part in the work.
ISO collaborates closely with the International Electrotechnical Commission (IEC) on all matters of
electrotechnical standardization.
The procedures used to develop this document and those intended for its further maintenance are
described in the ISO/IEC Directives, Part 1. In particular the different approval criteria needed for the
different types of ISO documents should be noted. This document was drafted in accordance with the
editorial rules of the ISO/IEC Directives, Part 2 (see www.iso.org/directives).
Attention is drawn to the possibility that some of the elements of this document may be the subject of
patent rights. ISO shall not be held responsible for identifying any or all such patent rights. Details of
any patent rights identified during the development of the document will be in the Introduction and/or
on the ISO list of patent declarations received (see www.iso.org/patents).
Any trade name used in this document is information given for the convenience of users and does not
constitute an endorsement.
For an explanation on the meaning of ISO specific terms and expressions related to conformity
assessment, as well as information about ISO’s adherence to the WTO principles in the Technical
Barriers to Trade (TBT) see the following URL: Foreword - Supplementary information.
The committee responsible for this document is ISO/TC 37, Terminology and other language and content
resources, Subcommittee SC 4, Language resource management.
ISO 24617 consists of the following parts, under the general title Language resource management —
Semantic annotation framework (SemAF):
— Part 1: Time and events (SemAF-Time, ISOTimeML)
— Part 2: Dialogue acts (SemAF-Dacts)
— Part 4: Semantic roles (SemAF-SR)
— Part 5: Discourse structures (SemAF-DS) [Technical Specification]
— Part 6: Principles of semantic annotation (SemAF Principles)
— Part 7: Spatial information (ISOspace)
The following parts are in preparation:
— Part 8: Semantic relations in discourse (SemAF DR-core)
— Part 9: Reference (ISOref)
iv © ISO 2016 – All rights reserved

INTERNATIONAL STANDARD ISO 24617-6:2016(E)
Language resource management — Semantic annotation
framework —
Part 6:
Principles of semantic annotation (SemAF Principles)
1 Scope
This part of ISO 24617 specifies the approach to semantic annotation characterizing the ISO Semantic
annotation framework (SemAF). It outlines the SemAF strategy for developing separate annotation
schemes for certain classes of semantic phenomena, aiming in the long term to combine these into a
single, coherent scheme for semantic annotation with wide coverage. In particular, it sets out the
notions of both an abstract and a concrete syntax for semantic annotations, mirroring the distinction
between annotations and representations that is made in the ISO Linguistic Annotation Framework.
It describes the role of these notions in relation to the specification of a metamodel and a semantic
interpretation of annotations, with a view to defining a well-founded annotation scheme.
This part of ISO 24617 also provides guidelines for dealing with two issues regarding the annotation
schemes defined in SemAF-parts: a) conceptual and terminological inconsistencies that may arise due
to overlaps between annotation schemes and b) the treatment of semantic phenomena that cut across
SemAF-parts, such as negation, modality and quantification. Instances of both issues are identified, and
in some cases, direction is given as to how they may be tackled.
2 Terms and definitions
For the purposes of this document, the following terms and definitions apply.
NOTE In addition, the terms ‘event’ and ‘eventuality’ are used (as synonyms) as defined in ISO 24617-1 as
something that can be said to obtain or hold true, to happen or to occur.
2.1
primary data
electronic representation of text or communicative behaviour
EXAMPLE Digital representations of text, transcriptions of speech, gestures or multimodal dialogue.
Note 1 to entry: ISO 24612 defines primary data as the ‘electronic representation of language data’. This definition
is unsatisfactory for this part of ISO 24617 as semantic annotation may relate to non-verbal or multimodal data,
such as stretches of spoken dialogue with accompanying gestures and facial expressions, and even gestures
and/or facial expressions without any accompanying speech.
2.2
annotation
linguistic information added to primary data (2.1), independent of its representation
[SOURCE: ISO 24612:2012, 2.3]
2.3
semantic annotation
annotation (2.2) which contains information about the meaning of a segment or region of primary data
(2.1)
2.4
metamodel
schematic representation of the concepts that are used in the analysis and description of the phenomena
covered in annotations (2.2) and of the relationships between them
3 Purpose and motivation
3.1 Purpose
The purpose of this part of ISO 24617 is to provide support for the establishment of a consistent and
coherent set of international standards for semantic annotation within the Semantic Annotation
Framework (SemAF). It aims to do so in three ways.
First, by making explicit which basic principles underlie the approach that has been followed in
defining international standards in the SemAF parts that have been published so far (ISO 24617-1 and
ISO 24617-2, ISO 24617-4 and ISO 24617-7), and in parts that are close to publication (ISO 24617-6)
or in preparation (ISO 24617-8). This approach provides the Semantic Annotation Framework with
methodological coherence and helps to ensure mutual consistency between existing, developing, and
future SemAF parts.
Second, by identifying overlaps between SemAF parts and indicating how such overlaps may be dealt
with. Examples are the occurrence of temporal and spatial relations among semantic roles and of
discourse relations between dialogue acts.
Third, by identifying common issues that arise in various parts of SemAF (they are only partly covered
in these parts, if they are covered at all) and, where possible, by giving directions as to how these issues
may be tackled. Examples of such issues are polarity, modality, quantification, measures, qualification,
veridicity, attribution and non-literal language use.
3.2 Motivation
Semantic annotation enhances primary data with information about their meaning. The state of the art
in computational semantics makes it unlikely that a single existing formalism for annotating semantic
information would receive wide support from researchers and developers. Moreover, semantic
annotation tasks often have the limited aim of annotating certain specific semantic phenomena,
such as semantic roles, discourse relations or coreference relations, rather than annotating the full
meaning of stretches of primary data. A strategy was therefore adopted in ISO TC 37/SC 4 to devise the
SemAF standards in different parts, with separate annotation schemes for those classes of semantic
phenomenon for which the state of the art would justify the establishment of annotation standards;
these schemes could be extended and combined over time, growing into a wide-coverage framework for
semantic annotation.
This ‘crystal growth’ strategy has contributed significantly to the progress made in establishing
standardized annotation concepts and schemes supporting the development of interoperable resources,
but it also entails certain risks:
a) the annotation schemes defined in different SemAF parts are not necessarily mutually consistent,
especially in the case of overlaps in scope;
b) it may not be possible to combine the schemes, defined in different parts, into a coherent single
scheme with a wider coverage if they incorporate different views or employ different methodologies;
c) some semantic phenomena do not belong to the scope of any SemAF parts but cannot be disregarded
entirely in some parts, and this may result in these phenomena being unsatisfactorily treated.
The methodological principles and guidelines provided in this part of ISO 24617 are designed to
minimize these risks.
2 © ISO 2016 – All rights reserved

With regard to the issue of mutual consistency between SemAF parts, it may be noted that ISO 24617-1
for annotating time and events and ISO 24617-2 for annotating dialogue acts are concerned with
sufficiently distinct kinds of semantic information to allow their definitions to be established
independently. Other SemAF parts, such as those concerned with semantic roles, with relations in
discourse and with spatial information show a certain amount of overlap in the information that they
aim to capture, and the question therefore arises: can we ensure that the annotation schemes, defined
in these parts, are mutually consistent?
Mutual consistency of SemAF parts relates to the possible integration of annotation schemes defined
in different parts. For example, it would be desirable to use the ISO 24617-1 scheme (“ISO-TimeML”) for
annotating time and events in combination with the ISO 24617-4 scheme for semantic roles, thereby
annotating coherently not only the events identified in the data with their temporal properties, but
also the way in which these events are related to their participants. Integrating these annotations with
those of spatial information, using the ISO 24617-7 scheme for spatial information, would be another
plausible and desirable step, given that time and space are intertwined with concepts relating to motion
and velocity. More generally, the integration of SemAF parts would greatly enhance the significance
of the individual parts; in the end, SemAF’s ‘crystal growth’ strategy of SemAF is only really useful
if the annotation schemes defined in the various parts can grow into a single scheme with a wide
coverage of semantic phenomena. Only then can it effectively support such applications as text-based
question answering or extracting semantic information from text, and form the basis for automatically
recognizing semantic phenomena by means of machine-learning techniques. Clearly, this is only possible
if the annotation schemes are mutually consistent (e.g. they use the same classification of event types),
and are coherent whether, for example, temporal and spatial objects are viewed as event participants
or as the circumstances of an event.
With regard to the risk of unsatisfactory partial treatments of phenomena that are not among the
core issues of any (current) SemAF part, it should be noted that some of these phenomena cut across
several of these parts and are important for semantics-driven applications. Negation, or more generally
negative polarity, and quantification are two cases in point. Given that the aim in ISO-TimeML, for
instance, is to support the annotation of events, of their relation to time, and of the temporal relations
among temporal objects, it is desirable to be able to deal with sentences like the following:
(1) John teaches every Monday.
(2) Mary called twice this morning.
(3) John rang home twice a day.
Sentence (1) is about a set of “teach” events, each of which is related to a different element of the set of
temporal objects that are called “Monday”, so this is a case of quantification involving two sets, a set of
events and sets of days. Similarly, sentence (2) about a set of two “call” events, both related to the same
period of time. Sentence (3) is about a set of events and their frequency of occurrence.
In order to deal with such phenomena, ISO-TimeML has certain provisions for annotating quantification,
[13]
but they are not really adequate and do not generalize to cases of quantification where no events
are involved.
4 Overview
The ISO efforts aiming to develop standards for semantic annotation rest on certain basic principles,
some of which have been laid out by Reference [14] as requirements for semantic annotation, and
have been developed further in Reference [5]; others have been formulated as general principles for
linguistic annotation and are part of the ISO Linguistic Annotation Framework (LAF; see Reference [18]
and ISO 24623-1). The two sets of principles and requirements are considered in Clause 5.
The three kinds of risk associated with the SemAF ‘crystal growth’ strategy that have been identified
above correspond to the following issues of consistency and completeness that arise in the design of
semantic annotation schemes within the SemAF framework.
Consistency among annotation schemes:
— methodological consistency: the same basic approach is followed with respect to the distinction
between abstract and concrete syntax and their interrelation, and with respect to their semantics;
— conceptual consistency: different schemes are based on compatible underlying views and ontological
assumptions regarding their basic concepts, as reflected in metamodels (e.g. verbs are viewed as
denoting states or events, rather than relations);
— terminological consistency: terms that occur in different annotation schemes have the same meaning
in every scheme and the same term is used across annotation schemes to indicate the same concept.
Completeness of a set of annotation schemes: the combination of multiple annotation schemes leads to
a scheme that
— covers a wide range of semantic phenomena,
— does not have significant gaps when covering the semantic phenomena that it aims to cover, and
— deals in a satisfactory way with semantic phenomena that cut across the combined schemes but
which do not belong to the core phenomena that any of the combined schemes are designed to cover.
Clause 5 describes the methodological framework for defining annotation schemes in SemAF parts,
thereby ensuring methodological consistency. Clause 6 discusses conceptual and terminological
consistency issues that arise due to overlaps between SemAF parts, while Clause 7 identifies issues of
completeness regarding the annotation of semantic phenomena that cut across existing SemAF parts.
5 Annotation principles and requirements
5.1 Principles inherited from the Linguistic Annotation Framework
The annotation of semantic information when using SemAF inherits the principles for linguistic
annotation as formulated in LAF. These principles are often of a very general nature; they include the
principle that relevant segments of primary data are referred to in a uniform and TEI-compliant way,
and the principle that different layers of annotation over the primary data can co-exist by using stand-
off annotation and a uniform way of cross-referencing between layers.
The latter principle, which concerns the distinction of layers of annotation enabled by a stand-off
representation format, is of particular relevance for SemAF because it allows different annotation
layers to be used for different types of semantic information; for example, one layer could be used for
the annotation of events, time and space, and another one could be used to annotate semantic roles.
In principle, this allows for the use not only of layers that are not mutually consistent, but also of
alternative annotations that employ different annotation schemes for the same phenomena. However,
the SemAF ‘crystal growth’ strategy is designed to ensure that the annotation schemes for the various
types of semantic information can grow into a coherent annotation scheme for a wide range of semantic
phenomena, and it is therefore highly undesirable to have inconsistencies between annotation layers
concerned with different SemAF parts.
[18]
Also of particular relevance for SemAF is the distinction between ‘annotations’ and ‘representations’.
An annotation is any item of linguistic information that is added to primary data, independently of any
particular representation format. A representation is a format into which an annotation is rendered,
for example as an XML expression. ISO standards are assumed to be defined at the level of annotations,
rather than representations. The fundamental distinction between annotations and representations has
prompted the development of a methodology for developing semantic annotation schemes that draws
a distinction between the ‘abstract syntax’ of annotations and the ‘concrete syntax’ of representations.
This methodology is described in Clause 6.
4 © ISO 2016 – All rights reserved

5.2 Other general annotation principles
In addition to the principles that SemAF inherits from LAF, other general principles for designing
annotation schemes (in particular as part of an ISO standard) are worth mentioning; most of these
emerged during the development of the ISO 24617-2 standard for dialogue act annotation.
a) Theoretical validity: Annotation standards should consolidate existing knowledge and
accordingly should be firmly rooted in theoretical studies of the annotated phenomena. Any concept
that may occur in annotations according to the standard should therefore be well established in the
scientific literature.
b) Empirical validity: Annotation standards are designed to be useful for annotating corpora of
recorded empirical data; the annotation scheme defined in a standard should not therefore include
theoretical constructs that are not found in such corpora, but only concepts that correspond to
phenomena that are observed in empirical data.
c) Learnability: For an annotation scheme to be useful in the construction of annotated language
resources, it should be possible both for human annotators and for automatic annotation systems
to effectively learn how to apply the scheme with acceptable precision.
d) Generalizability: ISO standards should not be restricted in their applicability to particular
languages, subject domains or applications.
e) Extensibility: While ISO standard annotation schemes are designed to be language-independent,
domain-independent and application-independent, some applications and some languages may
require specific concepts that are not relevant in other applications or languages. Annotation
schemes should therefore be open, that is to say, they should allow extension with language-
specific, domain-specific and application-specific concepts.
f) Completeness: An annotation standard is designed to provide a good coverage of the phenomena of
which it is designed to enable the annotation; the set of concepts defined in an annotation standard
should, in that sense, be complete.
g) Variable granularity: One way to achieve good coverage is to include annotation concepts of a
high level of generality and which cover many specific instances. Since an annotation scheme
which uses only very general concepts would not be optimally useful, this leads to the principle
that annotation schemes should include concepts with different levels of granularity. This is also
beneficial for its interoperability, as it provides more possibilities for conversion between existing
annotation schemes and the standard scheme.
h) Compatibility: In order to enable mappings between alternative annotation schemes and thereby
contribute to the interoperability of annotated resources, concepts that are commonly found in
existing annotation schemes should preferably be included in an annotation standard.
5.3 Principles specific to semantic annotation
The idea behind annotating a text, which dates from long before the digital era, is to add information to a
primary text in order to support its understanding. The semantic annotation of digital source texts has
a similar purpose, namely to support the understanding of the text by humans, as well as by machines.
An annotation that does not add any information would therefore seem to make little sense, but the
[39]
following example of the annotation of a temporal expression using TimeML seems to do just that:
NOTE 1 For simplicity, the annotations of the events that are mentioned in the previous sentence is
suppressed here.
(4)

The CEO announced that he would resign as of

the first of December 2008


In this annotation, the subexpression
adds to the noun phrase “the first of December 2008” the information that this phrase describes: the
date 2008-12-01.This does not add any information; rather, it paraphrases the noun phrase in TimeML.
This could be useful if the expression in the annotation language had a well-specified semantics that
could be used directly by computer programs for applications like information extraction and question
answering. Unfortunately, TimeML does not have a semantics.
NOTE 2 It would be very simple to provide a semantics for the XML fragment shown here but it would be very
difficult to do so for the whole of TimeML. See also 7.3.
A case where the annotation of a date as in the above example does add something is (5). From the
utterance “Mr Brewster called a staff meeting today”, it is impossible to know the date on which the event
that is mentioned took place; in this case, the annotation, which is identical to (4), would be informative.
NOTE 3 Note that the examples of TimeML annotations shown here are “old-fashioned” in the sense that the
TIMEX3 element is wrapped around the annotated string. Modern annotation methods (e.g. in ISO-TimeML) use
stand-off representations.
(5) Mr Brewster called a staff meeting today.

Mr Brewster called a staff meeting

today


The examples in (4) and (5) illustrate two different functions that semantic annotations may have:
interpreting a natural language expression by recoding it in a formal annotation language with a well-
defined semantics, and adding context information, in order to allow the interpretation of context-
dependent expressions.
A third function that semantic annotations may have is to make explicit how certain parts of an
utterance are semantically related; for example, this is the function of semantic role labelling and of
indicating semantic relations between sentences in a discourse when no such relation is mentioned.
Note that the first function presupposes annotations to have a well-defined semantics; the other two
functions do not presuppose this, but since semantic annotations in digital corpora are typically
designed to support interpretation and inference, a desideratum for all functions that a semantic
annotation may have is that it has a well-defined semantics. Semantic annotations, like other linguistic
annotations, may also serve the purpose of supporting linguistic research, such as identifying syntactic
and semantic patterns in sentences and texts.
These considerations lead to the following two principles for semantic annotation:
— Semantic additivity: semantic annotations add semantic information to source data, or re-express
certain source data in a formal representation.
— Semantic adequacy: semantic annotations should have a well-defined semantics, making the
annotations machine-interpretable.
6 © ISO 2016 – All rights reserved

6 The methodological basis of SemAF
6.1 Steps in the design of an annotation scheme
An annotation scheme determines which information may be added to primary data and how that
information is expressed. The design of an annotation scheme from scratch should begin with a
conceptual analysis of the information that the annotations should capture. This analysis identifies
the concepts that form the building blocks of annotations and specifies how these blocks may be
used to build annotation structures. This may be broken down into two steps, the first establishing a
conceptual view of the phenomena to be annotated, the second an articulation of this view in the form
of a formal specification of categories of entities and relations, and of how annotation structures can be
built up from elements in these categories. The first of these steps corresponds to what is known in ISO
projects as the establishment of a ‘metamodel’, that is to say, the expression of a conceptual view of the
phenomena to be annotated in the form of a (UML) diagram. The formal specification produced in the
second step constitutes the abstract syntax of an annotation language.
While these two steps make explicit what information can be captured in the annotations, they do not
concern the use of representation formats, such as XML strings, logical formulas, graphs or feature
structures; the abstract syntax defines the specification of information in terms of set-theoretic
structures, called annotation structures. An annotation structure is a collection of two kinds of structure,
entity structures and link structures. An entity structure contains semantic information about a segment
of primary data and is formally a pair consisting of a markable, which refers to a segment of
primary data, and of certain semantic information. A link structure contains information about the way
two segments of primary data are semantically related; for example, in semantic role annotation a link
structure is a triple where e is an entity structure that contains information about an event,
1 2 i 1
e is an entity structure that contains information about a participant in the event, and R is a relation
2 i
denoting a semantic role.
The third step in the definition of an annotation scheme is the specification of the meaning of the structures
defined by the abstract syntax, that is to say, the specification of semantics for annotation structures.
The fourth and final step is the definition of a format for representing annotation structures, such as a
serialization in XML. The semantics of an expression in this format is that of the annotation structure
that it represents. In sum, this approach to designing annotation schemes consists of four steps, with as
many feedback loops as may be needed.
The method consisting of these four steps is called ‘CASCADES’: Conceptual analysis, Abstract syntax,
Semantics, and Concrete syntax for Annotation language DESign. Figure 1 visualizes the CASCADES
method, of which the central concept, that of an abstract syntax for annotations with the specification
of a semantics for annotation structures (rather than for their representations in a particular format),
was introduced in Reference [5].
The CASCADES method is useful for enabling a systematic design process, in which due attention is
given to the conceptual and semantic choices on which more superficial decisions, such as the choice
of particular XML attributes and values, should be based. Apart from supporting the design of an
annotation scheme starting from scratch, the method also provides support for improving an existing
annotation scheme. This support consists not only of the distinction of four well-defined design steps,
but also of procedures and guidelines for taking these steps. These procedures are outlined in 6.3.
But before that, the use of metamodels, resulting from the conceptual analysis phase of designing an
annotation scheme, is discussed in 6.2.
Figure 1 — Steps and feedback loops in the CASCADES method
6.2 Metamodels
A metamodel of an annotation standard is a schematic representation of the relations between
the concepts that are important in the analysis and description of the phenomena covered in the
annotations. Over the years, two slightly different notions of a metamodel have been used in ISO
projects, namely the following:
— A: as a representation of the relations between the most important concepts that are mentioned in
the document in which the standard is defined;
— B: as a representation of the relations between the concepts denoted by terms that occur in
annotations, according to the standard.
Type-A metamodels may help nontechnical readers to have a better understanding of annotation
schemes; those of type-B are a visual representation of the abstract syntax of the annotations according
to the scheme and may help readers see at a glance what information the annotations may contain.
(Note that a type-A metamodel may have a type-B metamodel as a proper part.) For example, Figure 2
shows the metamodel of ISO 24617-4 for semantic role labelling. This metamodel reflects the view that
semantic roles are relations between eventualities and their participants; different roles correspond
to different ways in which a participant is involved in an eventuality. For example, in “Chris wrote a
poem”, a poem participates in an event in the Result role (i.e. created by the event); in “Chris revised the
poem”, the poem is involved in the Patient role (i.e. affected by the event); and in “Chris read a poem”, a
poem participates in the Theme role (i.e. not affected by the event). The annotation of semantic roles
therefore involves the use both of eventualities, most often corresponding to verbs, and of participants,
typically expressed by noun phrases. The metamodel associates eventualities and participants with
markables, whose source is a stretch of text in the primary data. The participants in an eventuality are
typically individuals, but may also be properties, numbers, quantities, propositions, sets of individuals
or embedded eventualities, as the examples in (6) illustrate.
(6)  a. He painted his house blue.
b. The birth of the twins increased the number of children to four.
c. The weight of this suitcase exceeds 20 kilos.
8 © ISO 2016 – All rights reserved

d. John believes he will get the money together.
e. Marie, Jill and Chris got together.
f. Eric and Nicole went to the concert.
Figure 2 — Metamodel for semantic role annotation (ISO 24617-4)
The term ‘entity’ is used in ISO 24617-4 to denote anything except an eventuality; the fact that a
participant in an eventuality can be an ‘entity’ of any sort as well as an (embedded) eventuality, as in (6f),
is reflected in the metamodel by the “entityFunctionsAs” relation between entities and participants,
and the “eventualityMayFunctionAs” relation between eventualities and participants. Many approaches
to semantic role labelling (e.g. PropBank, FrameNet and VerbNet) make use of “eventuality frames” that
specify the set of semantic roles expected in a type of eventuality, and take the eventuality frame as a
whole into consideration when determining the choice of individual semantic roles. For that reason,
eventuality frames have also been included in the metamodel.
Is this a type-A metamodel or a type-B metamodel? The following is an example of an annotation
representation according to ISO 24617-4 (where m1 is a markable that refers to “The soprano”; m2 is a
markable that refers to “sang”, and m3 is a markable that refers to “an aria”):
(7)  a. The soprano sang an aria.
b.





In this annotation, eventualities are directly related to entities by means of an element like event=”#e1” participant=”#x1” semRole=”agent”/>. There is no ‘participant’ object,
related to an entity via an ‘entityFunctionsAs relation’, as there is in the metamodel. It follows that
this metamodel is not just a visual representation of the (abstract) syntax of annotation structures, but
rather a type-A model, since it has elements that do not appear in annotations. This is corroborated by
the fact that eventuality frames do not occur as such in annotations; eventuality frames may be used in
the process of semantic role labelling, but are not part of the resulting annotations.
6.3 Abstract syntax, concrete syntax and semantics
The abstract syntax of an annotation scheme, as mentioned above, specifies the information in
annotations in terms of set-theoretical structures such as pairs and triples, like the triple
1 2 i
which relates the two arguments e and e through the relation R . More generally, these structures
1 2 i
are n-tuples of elements that are either basic concepts taken from a store of basic concepts called the
‘conceptual inventory’ of the abstract syntax specification, or n-tuples of such structures.
A concrete syntax specifies a representation format for annotation structures, such as the XML format
illustrated in (7), where a triple like is represented by a sequence of three XML elements,
1 2 i
of which the element
represents the relation and the other two elements represent two entity structures.
A representation format for annotation structures should ideally give an exact expression of the
information contained in annotation structures. A concrete syntax, defining a representation format
for a given abstract syntax, is said to be ideal if it has the following properties:
— completeness: every annotation structure defined by the abstract syntax can be represented by an
expression defined by the concrete syntax;
— unambiguity: every representation defined by the concrete syntax is the rendering of exactly one
annotation structure defined by the abstract syntax.
The representation format defined by an ideal concrete syntax is called an ideal representation format.
Due to its completeness, an ideal concrete syntax CS defines a function F from annotation structures
i i
−1
to CS -representations, and due to its ‘unambiguity’, there is also an inverse function F from CS -
i i i
representations to annotation structures. If I is the interpretation function defining the semantics
a
of the abstract syntax, then the meaning of a representation r in the ideal format CS is defined by
k
−1
I (F (r)). It follows that, for any two ideal representation formats CS and CS , there is a meaning-
a k i j
preserving conversion C defined by:
ij
−1
(8)  C (r) = F [F (r)]
ij j i
This mapping is meaning-preserving due to the fact that the meaning of a representation is the meaning
of the annotation structure that it encodes. Since this holds for any ideal concrete syntax, it follows that
any two ideal representation formats are semantically equivalent.
Figure 3 visualizes the relations between abstract syntax, semantics, and alternative ideal concrete-
syntax specifications. It is immediately clear that a given representation r defined by concrete syntax
CS can be converted into a semantically equivalent representation r’ in the representation format CS
i j
−1
by first applying the function F in order to determine the annotation structure which it encodes, and
i
applying to that annotation structure the function F which encodes it in the format CS .
j j
NOTE 1 The conditions defining an ideal concrete syntax require every annotation structure to have at least
one representation (but they do not rule out the possibility of an annotation structure having more than one
representation; the encoding functions, like F , may assign multiple CS -representations to a given annotation
j i
structure). If these conditions are met, a conversion function like the one defined in (8) assigns a set of CS -
j
equivalents to a given CS -representation.
i
An ideal concrete syntax can be derived systematically from an abstract syntax. The procedure for
doing so depends on the kind of concrete representations that are aimed for. For example, a concrete
syntax defining XML representations can be constructed using the following procedure:
NOTE 2 The description given here is simplified, for illustrative purposes. For more details, see Reference [7].
a) For each element of the conceptual inventory specify an XML name;
b) For each type of entity structure , define an XML element with the following attributes and
values:
— the special attribute ‘xml:id’, whose value is an identifier of the element;
10 © ISO 2016 – All rights reserved

— the special attribute ‘target’,
...