ISO 24610-1:2006

Language resource management - Feature structures - Part 1: Feature structure representation

ISO 24610-1:2006 provides a format for the representation, storage and exchange of feature structures in natural language applications concerned with the annotation, production or analysis of linguistic data. It also defines a computer format for the description of constraints that bear on a set of features, feature values, feature specifications and operations on feature structures, thus offering a means of checking the conformance of each feature structure with regards to a reference specification.

Gestion des ressources linguistiques — Structures de traits — Partie 1: Représentation de structures de traits

Upravljanje z jezikovnimi viri - Strukture lastnosti - 1. del: Predstavitev struktur lastnosti

Strukture lastnosti so ključni sestavni del številnih jezikovnih formalizmov in osnovni mehanizmi za predstavitev informacij, ki jih predelajo ali oblikujejo aplikacije za jezikovno inženirstvo, ali ki so oblikovane za te aplikacije. Ta del standarda ISO 24610 zagotavlja format za predstavitev, shranjevanje in izmenjavo struktur lastnosti v aplikacijah za naravni jezik, ki označujejo, ustvarjajo ali analizirajo jezikovne podatke. Prav tako določa računalniški format za opis omejitev, ki veljajo za sklop lastnosti, vrednosti lastnosti, specifikacij lastnosti in operacij na strukturah lastnosti, in tako nudi način za preverjanje skladnosti vsake strukture lastnosti glede na referenčno specifikacijo.

General Information

• Status: Published
• Publication Date: 09-Apr-2006
• ICS: 01.140.20 - Information sciences
• Technical Committee: ISO/TC 37/SC 4 - Language resource management
• Drafting Committee: ISO/TC 37/SC 4/WG 1 - Basic descriptors and mechanisms for language resources
• Current Stage: 9093 - International Standard confirmed
• Start Date: 26-Jan-2022
• Completion Date: 13-Dec-2025
• Ref Project: SIST ISO 24610-1:2013 - Language resource management -- Feature structures -- Part 1: Feature structure representation

Overview

ISO 24610-1:2006 / ISO 24610-1:2006 specifies a standardized format for representing, storing and exchanging feature structures used in language resource management and natural language processing (NLP). Part 1 defines the concepts, informal semantics and an XML-based vocabulary for untyped feature structures, feature values and feature libraries. It also provides a computer-readable way to describe constraints on features and values so implementations can check conformance against a reference specification.

Keywords: feature structures, language resource management, XML representation, ISO 24610-1, linguistic annotation, natural language processing.

Key Topics and Technical Requirements

• General characteristics and notations: formalizes graph, matrix and XML-based notations for feature structures.
• Structure sharing (re-entrancy): representation of shared substructures to model linguistic phenomena efficiently.
• Collections as feature values: support for lists, sets and multisets as complex feature values.
• Typed feature structures (overview): discussion of types, type notation and type-related constraints (with full typed validation addressed in Part 2).
• Relations and operations: formal treatment of subsumption, compatibility, unification (including shared structures) and operations on feature values such as concatenation, union, alternation and negation.
• XML Representation: a standard XML vocabulary and organization for elementary and complex feature structures, feature/value libraries, default values, and linking analysis to text.
• Formal and implementation materials (Annexes): formal definitions, RELAX NG module, illustrative examples, type inheritance guidance and denotational semantics to aid implementation and validation.
• Normative references: integrates existing standards such as ISO 8879 (SGML) and ISO 19757‑2 (RELAX NG) for schema and validation.

Practical Applications and Users

• NLP engineers and computational linguists: interchange of parsed or annotated linguistic structures across tools and pipelines.
• Corpus and annotation projects: storing multi-layer annotations (morphosyntax, semantics, phonology) with shared substructures and complex values.
• Tool developers and library authors: implementing parsers, unifiers or feature-based grammars consistent with a standard XML interchange format.
• Language resource managers and standards teams: ensuring long-term interoperability of language resources and validation of feature-structure conformance.
• Linguists and lexicographers: encoding feature-based descriptions for research and publication.

Related Standards

• ISO 8879 - SGML (reference for markup principles)
• ISO 19757-2 - RELAX NG (used for schema/validation)
• TEI (Text Encoding Initiative) - collaborative origin of feature-structure chapters

SIST ISO 24610-1 is a practical, interoperable foundation for representing feature structures in linguistic data exchange, validation and tool integration.

Standards Content (Sample)

SLOVENSKI STANDARD
01-julij-2013
Upravljanje z jezikovnimi viri - Strukture lastnosti - 1. del: Predstavitev struktur
lastnosti
Language resource management -- Feature structures -- Part 1: Feature structure
representation
Gestion des ressources linguistiques -- Structures de traits -- Partie 1: Représentation de
structures de traits
Ta slovenski standard je istoveten z: ISO 24610-1:2006
ICS:
01.140.20 Informacijske vede Information sciences
2003-01.Slovenski inštitut za standardizacijo. Razmnoževanje celote ali delov tega standarda ni dovoljeno.

INTERNATIONAL ISO
STANDARD 24610-1
FIrst edition
2006-04-15
Language resource management —
Feature structures —
Part 1:
Feature structure representation
Gestion des ressources linguistiques — Structures de traits —
Partie 1: Représentation de structures de traits

Reference number
©
ISO 2006
PDF disclaimer
This PDF file may contain embedded typefaces. In accordance with Adobe's licensing policy, this file may be printed or viewed but
shall not be edited unless the typefaces which are embedded are licensed to and installed on the computer performing the editing. In
downloading this file, parties accept therein the responsibility of not infringing Adobe's licensing policy. The ISO Central Secretariat
accepts no liability in this area.
Adobe is a trademark of Adobe Systems Incorporated.
Details of the software products used to create this PDF file can be found in the General Info relative to the file; the PDF-creation
parameters were optimized for printing. Every care has been taken to ensure that the file is suitable for use by ISO member bodies. In
the unlikely event that a problem relating to it is found, please inform the Central Secretariat at the address given below.

©  ISO 2006
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized in any form or by any means,
electronic or mechanical, including photocopying and microfilm, without permission in writing from either ISO at the address below or
ISO's member body in the country of the requester.
ISO copyright office
Case postale 56 • CH-1211 Geneva 20
Tel. + 41 22 749 01 11
Fax + 41 22 749 09 47
E-mail [e-mail]
Web www.iso.org
Published in Switzerland
ii © ISO 2006 – All rights reserved

Contents Page
Foreword. v
Introduction . vi
1 Scope. 1
2 Normative references. 1
3 Terms and definitions. 1
4 General characteristics of feature structure. 4
4.1 Overview. 4
4.2 Use of feature structures . 4
4.3 Basic concepts. 5
4.4 Notations . 5
4.4.1 Overview. 5
4.4.2 Graph notation . 6
4.4.3 Matrix notation . 7
4.4.4 XML-based notation. 8
4.5 Structure sharing. 10
4.6 Collections as complex feature values. 12
4.6.1 Overview. 12
4.6.2 Lists as feature values . 12
4.6.3 Sets as feature values . 14
4.6.4 Multisets as feature values . 15
4.7 Typed feature structure. 16
4.7.1 Overview. 16
4.7.2 Types. 16
4.7.3 Notations . 16
4.8 Subsumption: relation on feature structures . 18
4.8.1 Overview. 18
4.8.2 Definition . 18
4.8.3 Condition A on path values . 19
4.8.4 Condition B on structure sharing . 19
4.8.5 Condition C on type ordering . 20
4.9 Operations on feature structures and feature values. 21
4.9.1 Overview. 21
4.9.2 Compatibility . 21
4.9.3 Unification . 22
4.9.4 Unification of shared structures . 22
4.10 Operations on feature values and types . 23
4.10.1 Concatenation and union operations . 23
4.10.2 Alternation. 24
4.10.3 Negation. 25
4.11 Informal semantics of feature structures. 27
5 XML Representation of feature structures. 29
5.1 Overview. 29
5.2 Organization . 29
5.3 Elementary feature structures and the binary feature value. 30
5.4 Other atomic feature values . 32
5.5 Feature and feature-value libraries . 35
5.6 Feature structures as complex feature values .37
5.7 Re-entrant feature structures . 40
5.8 Collections as complex feature values. 41
5.9 Feature value expressions. 44
5.9.1 Overview. 44
5.9.2 Alternation. 44
5.9.3 Negation. 47
5.9.4 Collection of values . 48
5.10 Default values. 48
5.11 Linking text and analysis . 50
Annex A (informative) Formal definitions and implementation of the XML representation of
feature structures. 54
A.1 Overview. 54
A.2 RELAX NG specification for the module . 54
Annex B (informative) Examples for illustration . 60
Annex C (informative) Type inheritance hierarchies. 62
C.1 Overview. 62
C.2 Definition. 62
C.3 Multiple inheritance . 64
C.4 Type constraints . 64
Annex D (informative) Denotational semantics of feature structure. 66
D.1 Feature structure signatures . 66
D.2 Feature structure algebra. 66
D.3 FS domains. 67
D.4 Feature structure interpretations . 68
D.5 Satisfiability . 68
D.6 Subsumption . 68
D.7 Unification. 69
Annex E (informative) Use of feature structures in applications. 70
E.1 Overview. 70
E.2 Phonological representation. 70
E.3 Grammar formalisms or theories . 70
E.4 Computational implementations . 71
Bibliography . 75

iv © ISO 2006 – All rights reserved

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.
International Standards are drafted in accordance with the rules given in the ISO/IEC Directives, Part 2.
The main task of technical committees is to prepare International Standards. Draft International Standards
adopted by the technical committees are circulated to the member bodies for voting. Publication as an
International Standard requires approval by at least 75 % of the member bodies casting a vote.
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.
ISO 24610-1 was prepared by Technical Committee ISO/TC 37, Terminology and other language and content
resources, Subcommittee SC 4, Language resource management.
ISO 24610 consists of the following parts, under the general title Language resource management — Feature
structures:
⎯ Part 1: Feature structure representation
The following part is under preparation:
⎯ Part 2: Feature system declaration

Introduction
This part of ISO 24610 results from the agreement between the Text Encoding Initiative Consortium (TEI) and
the ISO TC 37/SC 4 that a joint activity should take place to revise the two existing chapters on feature
structures and feature system declaration in The TEI Guidelines called P4.
It is foreseen that ISO 24610 will have the following two parts.
⎯ Part 1, Feature structure representation, describes feature structures and their representation. It provides
an informal but explicit overview of their basic characteristics and formal semantics. In addition, part 1
defines a standard XML (eXtended Markup Language) vocabulary for the representation of untyped
feature structures, feature values, and feature libraries. It thus provides a reference format for the
exchange of feature structure representations between different application systems.
⎯ Part 2, Feature system declaration, discusses ways of validating typed feature structures which are
conformant to part 1, and of enforcing application-specific constraints. It proposes an XML vocabulary for
the representation of such constraints with reference to a set of features and the range of values
appropriate for them, and thus facilitates representation and validation of a type hierarchy as well as other
well-formedness conditions for particular applications, in particular those related to the goal of language
resource management.
vi © ISO 2006 – All rights reserved

INTERNATIONAL STANDARD ISO 24610-1:2006(E)

Language resource management — Feature structures —
Part 1:
Feature structure representation
1 Scope
Feature structures are an essential part of many linguistic formalisms as well as an underlying mechanism for
representing the information consumed or produced by and for language engineering applications. This part of
ISO 24610 provides a format for the representation, storage and exchange of feature structures in natural
language applications concerned with the annotation, production or analysis of linguistic data. It also defines a
computer format for the description of constraints that bear on a set of features, feature values, feature
specifications and operations on feature structures, thus offering a means of checking the conformance of
each feature structure with regards to a reference specification.
2 Normative references
The following referenced documents are indispensable for the application of this document. For dated
references, only the edition cited applies. For undated references, the latest edition of the referenced
document (including any amendments) applies.
ISO 8879, Information processing — Text and office systems — Standard Generalized Markup Language
(SGML), as extended by TC 2 (ISO/IEC JTC 1/SC 34 N029: 1998-12-06).
ISO 19757-2, Information technology — Document Schema Definition Language (DSDL) — Part 2: Regular-
grammar-based validation — RELAX NG
NOTE The first reference permits the use of XML and the second, RELAX NG,provides a specification for XML
modules. RELAX NG is a schema language for XML, standing for REgular LAnguage for XML for Next Generation, and
simplifies and extends the features of DTDs, Document Type Definitions.
3 Terms and definitions
For the purposes of this document, the terms and definitions given in ISO 8879 and ISO 19757-2 and the
following apply. This list is provided to clarify the terminology relating to feature structures used throughout
this part of ISO 24610. Terminology derived from XLM and other formal languages is not defined here.
3.1
alternation
operation on feature values (3.23) that returns one and only one of the values supplied as its argument
NOTE Given a feature specification F : a|b, where a|b denotes the alternation of a and b, F has either the value a or
the value b, but not both.
3.2
atomic value
value (3.23) without internal structure, i.e. value other than feature structure (3.10) and collection (3.4)
3.3
boxed label
label in box used in a matrix notation to denote a value shared by several features (3.8)
NOTE The label may be any alphanumeric symbol.
3.4
collection
list, set, or multiset of values (3.23)
NOTE A list is an ordered collection of entities some of which may be identical. A set is an unordered collection of
unique entities. A multiset is an unordered collection of entities that may or may not be unique; it is sometimes referred to
as a bag.
3.5
complex value
value (3.23) represented either as a feature structure (3.10) or as collection (3.4)
3.6
concatenation
operation of combining two lists of values (3.23) into a single list
3.7
empty feature structure
feature structure (3.10) containing no feature specifications (3.9)
3.8
feature
property of an entity
NOTE The combination of feature and feature-value constitutes a feature specification (3.9). For example, number
is a feature, singular is a value, and a pair is a feature specification.
3.9
feature specification
assignment of a value (3.23) to a feature (3.8)
NOTE Formally, it is treated as a pair of a feature and its value.
3.10
feature structure
set of feature specifications (3.9)
NOTE The minimum feature structure is the empty feature structure (3.7).
3.11
graph notation
notation of feature structure (3.10) in a single rooted graph
3.12
incompatibility
relation between two feature structures (3.10) which have conflicting types (3.19) or at least one common
feature (3.8) with incompatible values (3.23)
NOTE Two feature structures that are incompatible cannot be unified. The empty feature structure (3.7) is
compatible with any other feature structure.
2 © ISO 2006 – All rights reserved

3.13
matrix notation
attribute-value matrix
AVM
notation that uses square brackets to represent feature structures (3.10)
NOTE In a matrix notation, each row represents a feature specification (3.9), with the feature name and the feature
value separated by a colon (:), space ( ) or the equals sign (=).
3.14
merge
generic operation that includes union (3.22) of sets or multisets and concatenation (3.6) of lists
3.15
negation
(unary) operation on a value (3.23) denoting any other value incompatible with it
NOTE In this part of ISO 24610, negation applies to values only and is not understood as a truth function as in
ordinary bivalent logics.
3.16
path
sequence of labeled arcs connecting nodes in a graph
3.17
structure sharing
re-entrancy
relation between two or more features (3.8) within a feature structure (3.10) that share a value (3.23)
3.18
subsumption
relationship between two feature structures (3.10) in which one is more specific than the other
NOTE A feature structure A is said to subsume a feature structure B if A is at least as informative as B. Subsumption
is a reflexive, antisymmetric, and transitive relation between two feature structures.
3.19
type
name of a class of entities
NOTE Feature structures (3.10) may be characterized by grouping them into certain classes. Types are used to
name such classes.
3.20
typed feature structure
feature structure (3.10) labelled by a type (3.19)
NOTE In the graph notation (3.11), each node is labelled with a type. In the matrix notation (3.13), a type is
ordinarily placed at the upper left corner of the inside of the pair of square brackets that represents a typed feature
structure. In XML notation, the type is supplied as the value (3.23) of a type attribute on the element.
3.21
unification
operation that combines two compatible feature structures (3.10) into the least informative feature structure
that contains the information from the two
3.22
union
operation that combines two sets, or multisets, into one
NOTE The corresponding operation for lists is concatenation (3.6).
3.23
value
information about an entity
NOTE There are two kinds of feature values: atomic value (3.2) and complex value (3.5).
4 General characteristics of feature structure
4.1 Overview
A feature structure is a general purpose data structure that identifies and groups together individual features
by assigning a particular value to each. Because of the generality of feature structures, they can be used to
represent many different kinds of information. Interrelations among various pieces of information and their
instantiation in markup provide a meta-language for representing analysis and interpretation of linguistic
content. Moreover, this instantiation allows a specification of a set of features with values of specific types and
1)
restrictions, by means of feature system declarations, or other XML mechanisms discussed in ISO 24610-2 .
4.2 Use of feature structures
Feature structures provide partial information about an object by specifying values for some or all of its
features. For example, if a female employee named Sandy Jones who is 30 years old is of the present
concern, then that person’s sex, name and age can be specified in a succinct manner by assigning a value to
each of these three features. These pieces of information can be put into a simple set notation, as in:
(1) Employee
{, , }
Feature structures are generally used as a vehicle for linguistic descriptions. For example, the phoneme /p/ in
English can be analysed in terms of its distinctive features: consonantal, anterior, voiceless, non-continuant or
stop sound segment, etc. Each of these features may be combined with one or other of the binary values
plus(+) and minus(-) to provide a feature specification. In a phonemic analysis such as the following, the value
of a feature specifies the presence or absence of that feature:
(2) Sound segment /p/
{, , , }
In such an analysis, the sound segment /p/ is distinguished from other phonemes in terms of the presence or
absence of specific features. For example, /p/ differs from the phoneme /b/ in VOICING, and from /k/ in
articulatory position: one is articulated at the anterior, (the lip or alveolar area of the mouth), and the other at
the nonanterior, namely the back of the oral cavity.
This feature analysis can be extended to other kinds of description. Consider a verb like “love”. Its features
include both syntactic and semantic properties: as a transitive verb, it takes an object as well as a subject as
its arguments, expressing the semantic relation of loving between two persons or animate beings. The exact
representation of these feature specifications requires a detailed elaboration of what feature structures are.
For now, these grammatical features can be roughly represented in a set format like the following:
(3) Grammatical features of the verb “love”
{, , },
where POS stands for part of speech.

1) Under preparation.
4 © ISO 2006 – All rights reserved

Since its first extensive use in generative phonology in mid-60s, the feature structure formalism has become
an essential tool not only for phonology, but more generally in support of syntax and semantics as well as
lexicon building, especially in computational work. Feature structures are used to describe and model
linguistic entities and phenomena by specifying their properties. In the next clauses, some of the formal
properties of feature structures are outlined together with means of representing them in a systematic manner.
4.3 Basic concepts
Feature structures may be viewed in a variety of ways. The most common and perhaps the most intuitive
views are the following:
⎯ a set of feature specifications that consists of pairs of features and their values;
⎯ labelled directed graphs with a single root where each arc is labelled with the name of a feature and
directed to its value.
In set-theoretic terms, a feature structure FS can be defined as a partial function from a set Feat of features to
a set FeatVal of values, where FeatVal consists of a set AtomVal of atomic values and a set FS of feature
structures.
(4) A feature structure as a set or partial function
FS ⊆ {⎥F∈Feat,v∈FeatVal}
i i i i
or
FS : Feat ⎯→FeatVal
where FeatVal = AtomVal∪ FS and where FeatVal stands for all possible values.
Values may be regarded as either atomic or complex. Atomic values are entities without internal structure,
while complex values may be feature structures or collections of values (either complex or atomic).
NOTE These two definitions are not absolutely equivalent, for in a set there may be more than one value, v, for a
feature, F.
For example, the part of speech (POS) feature can take the name of an atomic morpho-syntactic category
such as verb as its value. Conversely, the agreement (AGR) feature in English takes a complex value in the
form of a feature structure with features PERSON and NUMBER. The word “loves”, for instance, has a POS
feature with the value “verb”, while the value of its AGR feature consists of a feature structure comprising two
feature specifications: PERSON with the value “3rd”, and NUMBER with the value “singular”.
4.4 Notations
4.4.1 Overview
As a list of feature-value pairs, the overall form of a feature structure is simple. However, the internal structure
of a feature structure may be complex when a feature structure contains either
a) a feature whose value is itself a feature structure, or
b) a multi-valued feature whose value is a list, set, or multiset.
Lists, sets and multisets may be made up of atomic values, or they may include complex values that are
themselves feature structures. That is, feature structures allow limitless recursive embedding. It is therefore
necessary to represent them in an understandable and mathematically precise notation.
There are two commonly used notations for the representation of feature structures: 1) graphs; and 2)
matrices. This part of ISO 24610 proposes a third notation based on the XML standard. The following
subclauses discuss how the same feature structure may be represented using each of these equivalent
notations. Graphs are suitable for mathematical discourses, matrices for linguistic descriptions, and XML
notations for computational implementation.
4.4.2 Graph notation
Feature structures are often represented as labelled directed graphs with a single root.
NOTE This graph can be either
a) acyclic, in which case it is referred to as a directed acyclic graph (DAG), or
b) cyclic for handling cases like the Liar’s paradox.
Feature structures are usually represented as DAGs, but it has been suggested that cyclical feature structures
should be introduced to model some linguistic phenomena.
Each graph starts with a single particular node called the root. From this root, any number of arcs may branch
out to other nodes and then some of them may terminate or extend to other nodes. The extension of directed
arcs shall, however, stop at some terminal nodes. On such a graph, representing a feature structure, each arc
is labelled with a feature name and its directed node is labelled with its value.
Here is a very simple example for a directed graph representing a feature structure.
(5) Feature structure in graph notation

In this graph, the two features, FEATURE1 and FEATURE2, are atomic-valued, taking value1 and value2 on
the terminal nodes respectively as their value. The feature FEATURE3 is, however, complex-valued, since it
takes as its value the feature structure which is represented by the two arcs FEATURE31 and FEATURE32
with their respective values, value31 and value32.
A graph may just consist of the root node, that is, without any branching arcs. Such a graph represents the
empty feature structure.
There may be one or more arcs branching out from the root, each of them bearing a single feature name.
Some of these labelled arcs originating from the root may again stretch out to another node and then from this
node to another, forming an indefinitely long sequence of feature names. Such a sequence of feature names,
labelling the arcs from the unique root node to each of the terminal nodes on a graph, is called a path. For
example, there are four paths in (5): FEATURE1, FEATURE2, FEATURE3.FEATURE31 and
FEATURE3.FEATURE32.
6 © ISO 2006 – All rights reserved

Here is a linguistically more relevant example.
(6) Linguistic example in graph notation

This graph consists of three paths: orthography (ORTH), SYNTAX.POS and SYNTAX.VALENCE. The path
consisting of a single feature name ORTH is directed to the terminal node “love”, which is an atomic value.
The path SYNTAX.POS terminates with the atomic value “verb” and the path SYNTAX.VALENCE with the
atomic value “transitive”. The nonterminal feature SYNTAX takes a complex value, namely a feature structure
consisting of the two feature specifications, and .
4.4.3 Matrix notation
Despite their mathematical elegance, graphs are difficult to typeset or read, especially when complex. For this
reason, feature structures are more often depicted using a matrix notation called attribute-value matrix, or
simply AVM.
NOTE 1 The term “attribute” is an alternative term for the concept which is called “feature”.
(7) Matrix notation
Each row in the square bracket with a FEATURE name followed by its value name represents a feature
specification in a feature structure.
NOTE 2 A colon, an equals sign or a space separates a feature from its value on each row of an AVM.
Feature values can be either atomic or complex. Each row with an atomic value terminates at that value. But if
the value is complex, then that row leads to another feature structure, as in the case of FEATURE3 above.
The notion of path is also important in several applications of the attribute-value matrix (AVM) notation, as
discussed further below. A path in an AVM is a sequence of feature names, as is the case with feature
structure graphs. An AVM with no rows and thus no occurrences of features, is represented as [ ]; this
represents the empty feature structure. If an AVM has at least one row consisting of a feature name and its
value, then there is a path of length 1 corresponding to each occurrence of a feature name in each row. Given
a path of length i, if the last member of that path takes a nonempty feature structure as value, then that path
forms a new path of length i +1 by taking each one of the features in that feature structure as its member.
For illustration, consider the following AVM which represents the same feature structure as is represented by
the graph notation (6).
(8) Example of an AVM notation

This AVM has three paths: ORTH, SYNTAX.POS and SYNTAX.VALENCE.
4.4.4 XML-based notation
Like any other formalism, XML has its own terminology. An XML document consists of typed elements,
occurrences of which are marked by start- and end-tags which enclose the content of the element. Every XML
element may be regarded as a linearization of a tree with a single root node. Each node contains either
terminal text or other element nodes. Elements have a specific type. Element occurrences can also carry
named attributes (or, more exactly, attribute-value pairs). For example:
(9) love
This is the XML way of representing an occurrence of the element type word, the content of which
is the string “love”. The word element is defined as having an attribute called “class”, whose value
in this case is the string “noun”.
The term “attribute” is thus used in a way which potentially clashes with its traditional usage in discussions of
feature structures. In AVM, feature structures are represented in a square bracket form, with each row
representing an attribute-value pair. In this usage, an attribute does not correspond necessarily to an XML
attribute. To avoid this potential source of confusion, the reader should be aware that in the remainder of this
clause, the term “attribute” will be used only in its technical XML sense.
To illustrate further the distinction, it should now be considered how a feature structure in AVM may be
represented in XML. Consider the following AVM that represents a non-typed feature structure:
(10) Representation of a feature structure in AVM

This feature structure consists of two feature specifications: one combines the feature ORTH with the string
value “love” and the other combines the feature POS with the string value “noun”.
NOTE This representation says nothing about the range of possible values for the two features: in particular, it does
not indicate that the range of possible values for the ORTH feature is not constrained, whereas (at least in principle) the
range of possible values for the POS feature is likely to be constrained to one of a small set of valid codes. In a constraint-
based grammar formalism, possible values for the feature ORTH need to be strings consisting of a finite number of
characters drawn from a well-defined character set, say the Roman Alphabet plus some special characters, for each
particular language.
In the XML representation proposed here, a feature structure is represented by means of an XML element fs,
and a feature-value specification by means of an XML element f:
(11)
.
.

8 © ISO 2006 – All rights reserved

This is a more generic approach than another, equally plausible, representation, in which each feature
specification might be represented by a specifically-named element:
(12)
.
.

Using this more generic approach means that systems can be developed which are independent of the
particular feature set, at the expense of slightly complicating the representation in any particular case. By
representing the specification of a feature as an f element, rather than regarding each specific feature as a
different element type, the overall processing model is simplified considerably.
A feature specification, as has been seen, has two components: a name, and a value. Again, there are
several equally valid ways of representing this combination in XML. Any of the following three, for example,
could be chosen:
(13)
a)  ;
b) 
noun
;
c) 
pos
noun
.
The fact that all three of these are possible arises from a redundancy introduced in the design of the XML
language largely for historical reasons. The present recommendation is to use a formulation like b) above,
though c) may be preferable on the grounds of its greater simplicity.
Finally, the representation of the value part of a feature specification is discussed. A name is simply a name,
but a value in the system discussed here may be of many different types: it might for example be an arbitrary
string, one of a predefined set of codes, a Boolean value, or a reference to another (nested) feature structure.
In the interests of greater expressivity, the present system proposes to distinguish amongst these kinds of
value in the XML representation itself.
Once more, there are a number of more or less equivalent ways of doing this. For example:
(14)
a) 

love

noun
;
b) 
love

.
The number of possible different value types is comparatively small, and the advantages of handling values
generically rather than specifically seem less persuasive. The approach currently taken is therefore to define
different element types for the different kinds of value (for example, string to enclose a string, symbol to
denote a symbol, binary to denote a binary value, fs to denote a nested feature structure).
Representing feature structures in XML notation is thus possible and rather straightforward. For example, the
following is the XML representation for the feature structure given in (6) and (8) above.
(15) Feature structure in XML notation


love







Despite the apparent difference between the representations (6), (8) and (15), the information contained in
each is identical. For example, the arc labels in graph notation like ORTH and SYNTAX are now represented
by tags with appropriate names. Furthermore, the XML representation can be processed by general
purpose software tools and thus can be converted automatically to a variety of other formats.
NOTE A more detailed discussion including ways to simplify the representation in (15) is provided in Clause 5.
4.5 Structure sharing
A feature structure representation may need to represent shared components. For example, in representing
the structure of a noun phrase (“la pomme") which includes a determiner (article) “la”, it may be desired to
show that the feature NUMBER is shared between the determiner and the noun. This sharing can be directly
represented in the graphic notation as follows:
(16) Merging paths in graph notation

where POS again stands for part of speech, ORTH for orthography and AGR for agreement feature.
10 © ISO 2006 – All rights reserved

Here, the two SPECIFIER.AGR and HEAD.AGR paths merge on the node NUMBER, indicating that they
share one and the same feature structure as their value.
In AVM notation, structure sharing is represented by providing a boxed integer label at the point where
structures are shared, as in the following example:
(17) Structure sharing in AVM notation

In a similar way, in the XML notation the element var is used to provide a name for the sharing point. It
supplies a label for the point which can then be referenced elsewhere in the structure, as in the following
example:
(18) Structure sharing in XML notation

SLOVENSKI STANDARD
01-julij-2013
Upravljanje z jezikovnimi viri - Strukture lastnosti - 1. del: Predstavitev struktur
lastnosti
Language resource management -- Feature structures -- Part 1: Feature structure
representation
Gestion des ressources linguistiques -- Structures de traits -- Partie 1: Représentation de
structures de traits
Ta slovenski standard je istoveten z: ISO 24610-1:2006
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 24610-1
FIrst edition
2006-04-15
Language resource management —
Feature structures —
Part 1:
Feature structure representation
Gestion des ressources linguistiques — Structures de traits —
Partie 1: Représentation de structures de traits

Reference number
©
ISO 2006
PDF disclaimer
This PDF file may contain embedded typefaces. In accordance with Adobe's licensing policy, this file may be printed or viewed but
shall not be edited unless the typefaces which are embedded are licensed to and installed on the computer performing the editing. In
downloading this file, parties accept therein the responsibility of not infringing Adobe's licensing policy. The ISO Central Secretariat
accepts no liability in this area.
Adobe is a trademark of Adobe Systems Incorporated.
Details of the software products used to create this PDF file can be found in the General Info relative to the file; the PDF-creation
parameters were optimized for printing. Every care has been taken to ensure that the file is suitable for use by ISO member bodies. In
the unlikely event that a problem relating to it is found, please inform the Central Secretariat at the address given below.

©  ISO 2006
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized in any form or by any means,
electronic or mechanical, including photocopying and microfilm, without permission in writing from either ISO at the address below or
ISO's member body in the country of the requester.
ISO copyright office
Case postale 56 • CH-1211 Geneva 20
Tel. + 41 22 749 01 11
Fax + 41 22 749 09 47
E-mail [e-mail]
Web www.iso.org
Published in Switzerland
ii © ISO 2006 – All rights reserved

Contents Page
Foreword. v
Introduction . vi
1 Scope. 1
2 Normative references. 1
3 Terms and definitions. 1
4 General characteristics of feature structure. 4
4.1 Overview. 4
4.2 Use of feature structures . 4
4.3 Basic concepts. 5
4.4 Notations . 5
4.4.1 Overview. 5
4.4.2 Graph notation . 6
4.4.3 Matrix notation . 7
4.4.4 XML-based notation. 8
4.5 Structure sharing. 10
4.6 Collections as complex feature values. 12
4.6.1 Overview. 12
4.6.2 Lists as feature values . 12
4.6.3 Sets as feature values . 14
4.6.4 Multisets as feature values . 15
4.7 Typed feature structure. 16
4.7.1 Overview. 16
4.7.2 Types. 16
4.7.3 Notations . 16
4.8 Subsumption: relation on feature structures . 18
4.8.1 Overview. 18
4.8.2 Definition . 18
4.8.3 Condition A on path values . 19
4.8.4 Condition B on structure sharing . 19
4.8.5 Condition C on type ordering . 20
4.9 Operations on feature structures and feature values. 21
4.9.1 Overview. 21
4.9.2 Compatibility . 21
4.9.3 Unification . 22
4.9.4 Unification of shared structures . 22
4.10 Operations on feature values and types . 23
4.10.1 Concatenation and union operations . 23
4.10.2 Alternation. 24
4.10.3 Negation. 25
4.11 Informal semantics of feature structures. 27
5 XML Representation of feature structures. 29
5.1 Overview. 29
5.2 Organization . 29
5.3 Elementary feature structures and the binary feature value. 30
5.4 Other atomic feature values . 32
5.5 Feature and feature-value libraries . 35
5.6 Feature structures as complex feature values .37
5.7 Re-entrant feature structures . 40
5.8 Collections as complex feature values. 41
5.9 Feature value expressions. 44
5.9.1 Overview. 44
5.9.2 Alternation. 44
5.9.3 Negation. 47
5.9.4 Collection of values . 48
5.10 Default values. 48
5.11 Linking text and analysis . 50
Annex A (informative) Formal definitions and implementation of the XML representation of
feature structures. 54
A.1 Overview. 54
A.2 RELAX NG specification for the module . 54
Annex B (informative) Examples for illustration . 60
Annex C (informative) Type inheritance hierarchies. 62
C.1 Overview. 62
C.2 Definition. 62
C.3 Multiple inheritance . 64
C.4 Type constraints . 64
Annex D (informative) Denotational semantics of feature structure. 66
D.1 Feature structure signatures . 66
D.2 Feature structure algebra. 66
D.3 FS domains. 67
D.4 Feature structure interpretations . 68
D.5 Satisfiability . 68
D.6 Subsumption . 68
D.7 Unification. 69
Annex E (informative) Use of feature structures in applications. 70
E.1 Overview. 70
E.2 Phonological representation. 70
E.3 Grammar formalisms or theories . 70
E.4 Computational implementations . 71
Bibliography . 75

iv © ISO 2006 – All rights reserved

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.
International Standards are drafted in accordance with the rules given in the ISO/IEC Directives, Part 2.
The main task of technical committees is to prepare International Standards. Draft International Standards
adopted by the technical committees are circulated to the member bodies for voting. Publication as an
International Standard requires approval by at least 75 % of the member bodies casting a vote.
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.
ISO 24610-1 was prepared by Technical Committee ISO/TC 37, Terminology and other language and content
resources, Subcommittee SC 4, Language resource management.
ISO 24610 consists of the following parts, under the general title Language resource management — Feature
structures:
⎯ Part 1: Feature structure representation
The following part is under preparation:
⎯ Part 2: Feature system declaration

Introduction
This part of ISO 24610 results from the agreement between the Text Encoding Initiative Consortium (TEI) and
the ISO TC 37/SC 4 that a joint activity should take place to revise the two existing chapters on feature
structures and feature system declaration in The TEI Guidelines called P4.
It is foreseen that ISO 24610 will have the following two parts.
⎯ Part 1, Feature structure representation, describes feature structures and their representation. It provides
an informal but explicit overview of their basic characteristics and formal semantics. In addition, part 1
defines a standard XML (eXtended Markup Language) vocabulary for the representation of untyped
feature structures, feature values, and feature libraries. It thus provides a reference format for the
exchange of feature structure representations between different application systems.
⎯ Part 2, Feature system declaration, discusses ways of validating typed feature structures which are
conformant to part 1, and of enforcing application-specific constraints. It proposes an XML vocabulary for
the representation of such constraints with reference to a set of features and the range of values
appropriate for them, and thus facilitates representation and validation of a type hierarchy as well as other
well-formedness conditions for particular applications, in particular those related to the goal of language
resource management.
vi © ISO 2006 – All rights reserved

INTERNATIONAL STANDARD ISO 24610-1:2006(E)

Language resource management — Feature structures —
Part 1:
Feature structure representation
1 Scope
Feature structures are an essential part of many linguistic formalisms as well as an underlying mechanism for
representing the information consumed or produced by and for language engineering applications. This part of
ISO 24610 provides a format for the representation, storage and exchange of feature structures in natural
language applications concerned with the annotation, production or analysis of linguistic data. It also defines a
computer format for the description of constraints that bear on a set of features, feature values, feature
specifications and operations on feature structures, thus offering a means of checking the conformance of
each feature structure with regards to a reference specification.
2 Normative references
The following referenced documents are indispensable for the application of this document. For dated
references, only the edition cited applies. For undated references, the latest edition of the referenced
document (including any amendments) applies.
ISO 8879, Information processing — Text and office systems — Standard Generalized Markup Language
(SGML), as extended by TC 2 (ISO/IEC JTC 1/SC 34 N029: 1998-12-06).
ISO 19757-2, Information technology — Document Schema Definition Language (DSDL) — Part 2: Regular-
grammar-based validation — RELAX NG
NOTE The first reference permits the use of XML and the second, RELAX NG,provides a specification for XML
modules. RELAX NG is a schema language for XML, standing for REgular LAnguage for XML for Next Generation, and
simplifies and extends the features of DTDs, Document Type Definitions.
3 Terms and definitions
For the purposes of this document, the terms and definitions given in ISO 8879 and ISO 19757-2 and the
following apply. This list is provided to clarify the terminology relating to feature structures used throughout
this part of ISO 24610. Terminology derived from XLM and other formal languages is not defined here.
3.1
alternation
operation on feature values (3.23) that returns one and only one of the values supplied as its argument
NOTE Given a feature specification F : a|b, where a|b denotes the alternation of a and b, F has either the value a or
the value b, but not both.
3.2
atomic value
value (3.23) without internal structure, i.e. value other than feature structure (3.10) and collection (3.4)
3.3
boxed label
label in box used in a matrix notation to denote a value shared by several features (3.8)
NOTE The label may be any alphanumeric symbol.
3.4
collection
list, set, or multiset of values (3.23)
NOTE A list is an ordered collection of entities some of which may be identical. A set is an unordered collection of
unique entities. A multiset is an unordered collection of entities that may or may not be unique; it is sometimes referred to
as a bag.
3.5
complex value
value (3.23) represented either as a feature structure (3.10) or as collection (3.4)
3.6
concatenation
operation of combining two lists of values (3.23) into a single list
3.7
empty feature structure
feature structure (3.10) containing no feature specifications (3.9)
3.8
feature
property of an entity
NOTE The combination of feature and feature-value constitutes a feature specification (3.9). For example, number
is a feature, singular is a value, and a pair is a feature specification.
3.9
feature specification
assignment of a value (3.23) to a feature (3.8)
NOTE Formally, it is treated as a pair of a feature and its value.
3.10
feature structure
set of feature specifications (3.9)
NOTE The minimum feature structure is the empty feature structure (3.7).
3.11
graph notation
notation of feature structure (3.10) in a single rooted graph
3.12
incompatibility
relation between two feature structures (3.10) which have conflicting types (3.19) or at least one common
feature (3.8) with incompatible values (3.23)
NOTE Two feature structures that are incompatible cannot be unified. The empty feature structure (3.7) is
compatible with any other feature structure.
2 © ISO 2006 – All rights reserved

3.13
matrix notation
attribute-value matrix
AVM
notation that uses square brackets to represent feature structures (3.10)
NOTE In a matrix notation, each row represents a feature specification (3.9), with the feature name and the feature
value separated by a colon (:), space ( ) or the equals sign (=).
3.14
merge
generic operation that includes union (3.22) of sets or multisets and concatenation (3.6) of lists
3.15
negation
(unary) operation on a value (3.23) denoting any other value incompatible with it
NOTE In this part of ISO 24610, negation applies to values only and is not understood as a truth function as in
ordinary bivalent logics.
3.16
path
sequence of labeled arcs connecting nodes in a graph
3.17
structure sharing
re-entrancy
relation between two or more features (3.8) within a feature structure (3.10) that share a value (3.23)
3.18
subsumption
relationship between two feature structures (3.10) in which one is more specific than the other
NOTE A feature structure A is said to subsume a feature structure B if A is at least as informative as B. Subsumption
is a reflexive, antisymmetric, and transitive relation between two feature structures.
3.19
type
name of a class of entities
NOTE Feature structures (3.10) may be characterized by grouping them into certain classes. Types are used to
name such classes.
3.20
typed feature structure
feature structure (3.10) labelled by a type (3.19)
NOTE In the graph notation (3.11), each node is labelled with a type. In the matrix notation (3.13), a type is
ordinarily placed at the upper left corner of the inside of the pair of square brackets that represents a typed feature
structure. In XML notation, the type is supplied as the value (3.23) of a type attribute on the element.
3.21
unification
operation that combines two compatible feature structures (3.10) into the least informative feature structure
that contains the information from the two
3.22
union
operation that combines two sets, or multisets, into one
NOTE The corresponding operation for lists is concatenation (3.6).
3.23
value
information about an entity
NOTE There are two kinds of feature values: atomic value (3.2) and complex value (3.5).
4 General characteristics of feature structure
4.1 Overview
A feature structure is a general purpose data structure that identifies and groups together individual features
by assigning a particular value to each. Because of the generality of feature structures, they can be used to
represent many different kinds of information. Interrelations among various pieces of information and their
instantiation in markup provide a meta-language for representing analysis and interpretation of linguistic
content. Moreover, this instantiation allows a specification of a set of features with values of specific types and
1)
restrictions, by means of feature system declarations, or other XML mechanisms discussed in ISO 24610-2 .
4.2 Use of feature structures
Feature structures provide partial information about an object by specifying values for some or all of its
features. For example, if a female employee named Sandy Jones who is 30 years old is of the present
concern, then that person’s sex, name and age can be specified in a succinct manner by assigning a value to
each of these three features. These pieces of information can be put into a simple set notation, as in:
(1) Employee
{, , }
Feature structures are generally used as a vehicle for linguistic descriptions. For example, the phoneme /p/ in
English can be analysed in terms of its distinctive features: consonantal, anterior, voiceless, non-continuant or
stop sound segment, etc. Each of these features may be combined with one or other of the binary values
plus(+) and minus(-) to provide a feature specification. In a phonemic analysis such as the following, the value
of a feature specifies the presence or absence of that feature:
(2) Sound segment /p/
{, , , }
In such an analysis, the sound segment /p/ is distinguished from other phonemes in terms of the presence or
absence of specific features. For example, /p/ differs from the phoneme /b/ in VOICING, and from /k/ in
articulatory position: one is articulated at the anterior, (the lip or alveolar area of the mouth), and the other at
the nonanterior, namely the back of the oral cavity.
This feature analysis can be extended to other kinds of description. Consider a verb like “love”. Its features
include both syntactic and semantic properties: as a transitive verb, it takes an object as well as a subject as
its arguments, expressing the semantic relation of loving between two persons or animate beings. The exact
representation of these feature specifications requires a detailed elaboration of what feature structures are.
For now, these grammatical features can be roughly represented in a set format like the following:
(3) Grammatical features of the verb “love”
{, , },
where POS stands for part of speech.

1) Under preparation.
4 © ISO 2006 – All rights reserved

Since its first extensive use in generative phonology in mid-60s, the feature structure formalism has become
an essential tool not only for phonology, but more generally in support of syntax and semantics as well as
lexicon building, especially in computational work. Feature structures are used to describe and model
linguistic entities and phenomena by specifying their properties. In the next clauses, some of the formal
properties of feature structures are outlined together with means of representing them in a systematic manner.
4.3 Basic concepts
Feature structures may be viewed in a variety of ways. The most common and perhaps the most intuitive
views are the following:
⎯ a set of feature specifications that consists of pairs of features and their values;
⎯ labelled directed graphs with a single root where each arc is labelled with the name of a feature and
directed to its value.
In set-theoretic terms, a feature structure FS can be defined as a partial function from a set Feat of features to
a set FeatVal of values, where FeatVal consists of a set AtomVal of atomic values and a set FS of feature
structures.
(4) A feature structure as a set or partial function
FS ⊆ {⎥F∈Feat,v∈FeatVal}
i i i i
or
FS : Feat ⎯→FeatVal
where FeatVal = AtomVal∪ FS and where FeatVal stands for all possible values.
Values may be regarded as either atomic or complex. Atomic values are entities without internal structure,
while complex values may be feature structures or collections of values (either complex or atomic).
NOTE These two definitions are not absolutely equivalent, for in a set there may be more than one value, v, for a
feature, F.
For example, the part of speech (POS) feature can take the name of an atomic morpho-syntactic category
such as verb as its value. Conversely, the agreement (AGR) feature in English takes a complex value in the
form of a feature structure with features PERSON and NUMBER. The word “loves”, for instance, has a POS
feature with the value “verb”, while the value of its AGR feature consists of a feature structure comprising two
feature specifications: PERSON with the value “3rd”, and NUMBER with the value “singular”.
4.4 Notations
4.4.1 Overview
As a list of feature-value pairs, the overall form of a feature structure is simple. However, the internal structure
of a feature structure may be complex when a feature structure contains either
a) a feature whose value is itself a feature structure, or
b) a multi-valued feature whose value is a list, set, or multiset.
Lists, sets and multisets may be made up of atomic values, or they may include complex values that are
themselves feature structures. That is, feature structures allow limitless recursive embedding. It is therefore
necessary to represent them in an understandable and mathematically precise notation.
There are two commonly used notations for the representation of feature structures: 1) graphs; and 2)
matrices. This part of ISO 24610 proposes a third notation based on the XML standard. The following
subclauses discuss how the same feature structure may be represented using each of these equivalent
notations. Graphs are suitable for mathematical discourses, matrices for linguistic descriptions, and XML
notations for computational implementation.
4.4.2 Graph notation
Feature structures are often represented as labelled directed graphs with a single root.
NOTE This graph can be either
a) acyclic, in which case it is referred to as a directed acyclic graph (DAG), or
b) cyclic for handling cases like the Liar’s paradox.
Feature structures are usually represented as DAGs, but it has been suggested that cyclical feature structures
should be introduced to model some linguistic phenomena.
Each graph starts with a single particular node called the root. From this root, any number of arcs may branch
out to other nodes and then some of them may terminate or extend to other nodes. The extension of directed
arcs shall, however, stop at some terminal nodes. On such a graph, representing a feature structure, each arc
is labelled with a feature name and its directed node is labelled with its value.
Here is a very simple example for a directed graph representing a feature structure.
(5) Feature structure in graph notation

In this graph, the two features, FEATURE1 and FEATURE2, are atomic-valued, taking value1 and value2 on
the terminal nodes respectively as their value. The feature FEATURE3 is, however, complex-valued, since it
takes as its value the feature structure which is represented by the two arcs FEATURE31 and FEATURE32
with their respective values, value31 and value32.
A graph may just consist of the root node, that is, without any branching arcs. Such a graph represents the
empty feature structure.
There may be one or more arcs branching out from the root, each of them bearing a single feature name.
Some of these labelled arcs originating from the root may again stretch out to another node and then from this
node to another, forming an indefinitely long sequence of feature names. Such a sequence of feature names,
labelling the arcs from the unique root node to each of the terminal nodes on a graph, is called a path. For
example, there are four paths in (5): FEATURE1, FEATURE2, FEATURE3.FEATURE31 and
FEATURE3.FEATURE32.
6 © ISO 2006 – All rights reserved

Here is a linguistically more relevant example.
(6) Linguistic example in graph notation

This graph consists of three paths: orthography (ORTH), SYNTAX.POS and SYNTAX.VALENCE. The path
consisting of a single feature name ORTH is directed to the terminal node “love”, which is an atomic value.
The path SYNTAX.POS terminates with the atomic value “verb” and the path SYNTAX.VALENCE with the
atomic value “transitive”. The nonterminal feature SYNTAX takes a complex value, namely a feature structure
consisting of the two feature specifications, and .
4.4.3 Matrix notation
Despite their mathematical elegance, graphs are difficult to typeset or read, especially when complex. For this
reason, feature structures are more often depicted using a matrix notation called attribute-value matrix, or
simply AVM.
NOTE 1 The term “attribute” is an alternative term for the concept which is called “feature”.
(7) Matrix notation
Each row in the square bracket with a FEATURE name followed by its value name represents a feature
specification in a feature structure.
NOTE 2 A colon, an equals sign or a space separates a feature from its value on each row of an AVM.
Feature values can be either atomic or complex. Each row with an atomic value terminates at that value. But if
the value is complex, then that row leads to another feature structure, as in the case of FEATURE3 above.
The notion of path is also important in several applications of the attribute-value matrix (AVM) notation, as
discussed further below. A path in an AVM is a sequence of feature names, as is the case with feature
structure graphs. An AVM with no rows and thus no occurrences of features, is represented as [ ]; this
represents the empty feature structure. If an AVM has at least one row consisting of a feature name and its
value, then there is a path of length 1 corresponding to each occurrence of a feature name in each row. Given
a path of length i, if the last member of that path takes a nonempty feature structure as value, then that path
forms a new path of length i +1 by taking each one of the features in that feature structure as its member.
For illustration, consider the following AVM which represents the same feature structure as is represented by
the graph notation (6).
(8) Example of an AVM notation

This AVM has three paths: ORTH, SYNTAX.POS and SYNTAX.VALENCE.
4.4.4 XML-based notation
Like any other formalism, XML has its own terminology. An XML document consists of typed elements,
occurrences of which are marked by start- and end-tags which enclose the content of the element. Every XML
element may be regarded as a linearization of a tree with a single root node. Each node contains either
terminal text or other element nodes. Elements have a specific type. Element occurrences can also carry
named attributes (or, more exactly, attribute-value pairs). For example:
(9) love
This is the XML way of representing an occurrence of the element type word, the content of which
is the string “love”. The word element is defined as having an attribute called “class”, whose value
in this case is the string “noun”.
The term “attribute” is thus used in a way which potentially clashes with its traditional usage in discussions of
feature structures. In AVM, feature structures are represented in a square bracket form, with each row
representing an attribute-value pair. In this usage, an attribute does not correspond necessarily to an XML
attribute. To avoid this potential source of confusion, the reader should be aware that in the remainder of this
clause, the term “attribute” will be used only in its technical XML sense.
To illustrate further the distinction, it should now be considered how a feature structure in AVM may be
represented in XML. Consider the following AVM that represents a non-typed feature structure:
(10) Representation of a feature structure in AVM

This feature structure consists of two feature specifications: one combines the feature ORTH with the string
value “love” and the other combines the feature POS with the string value “noun”.
NOTE This representation says nothing about the range of possible values for the two features: in particular, it does
not indicate that the range of possible values for the ORTH feature is not constrained, whereas (at least in principle) the
range of possible values for the POS feature is likely to be constrained to one of a small set of valid codes. In a constraint-
based grammar formalism, possible values for the feature ORTH need to be strings consisting of a finite number of
characters drawn from a well-defined character set, say the Roman Alphabet plus some special characters, for each
particular language.
In the XML representation proposed here, a feature structure is represented by means of an XML element fs,
and a feature-value specification by means of an XML element f:
(11)
.
.

8 © ISO 2006 – All rights reserved

This is a more generic approach than another, equally plausible, representation, in which each feature
specification might be represented by a specifically-named element:
(12)
.
.

Using this more generic approach means that systems can be developed which are independent of the
particular feature set, at the expense of slightly complicating the representation in any particular case. By
representing the specification of a feature as an f element, rather than regarding each specific feature as a
different element type, the overall processing model is simplified considerably.
A feature specification, as has been seen, has two components: a name, and a value. Again, there are
several equally valid ways of representing this combination in XML. Any of the following three, for example,
could be chosen:
(13)
a)  ;
b) 
noun
;
c) 
pos
noun
.
The fact that all three of these are possible arises from a redundancy introduced in the design of the XML
language largely for historical reasons. The present recommendation is to use a formulation like b) above,
though c) may be preferable on the grounds of its greater simplicity.
Finally, the representation of the value part of a feature specification is discussed. A name is simply a name,
but a value in the system discussed here may be of many different types: it might for example be an arbitrary
string, one of a predefined set of codes, a Boolean value, or a reference to another (nested) feature structure.
In the interests of greater expressivity, the present system proposes to distinguish amongst these kinds of
value in the XML representation itself.
Once more, there are a number of more or less equivalent ways of doing this. For example:
(14)
a) 

love

noun
;
b) 
love

.
The number of possible different value types is comparatively small, and the advantages of handling values
generically rather than specifically seem less persuasive. The approach currently taken is therefore to define
different element types for the different kinds of value (for example, string to enclose a string, symbol to
denote a symbol, binary to denote a binary value, fs to denote a nested feature structure).
Representing feature structures in XML notation is thus possible and rather straightforward. For example, the
following is the XML representation for the feature structure given in (6) and (8) above.
(15) Feature structure in XML notation


love







Despite the apparent difference between the representations (6), (8) and (15), the information contained in
each is identical. For example, the arc labels in graph notation like ORTH and SYNTAX are now represented
by tags with appropriate names. Furthermore, the XML representation can be processed by general
purpose software tools and thus can be converted automatically to a variety of other formats.
NOTE A more detailed discussion including ways to simplify the representation in (15) is provided in Clause 5.
4.5 Structure sharing
A feature structure representation may need to represent shared components. For example, in representing
the structure of a noun phrase (“la pomme") which includes a determiner (article) “la”, it may be desired to
show that the feature NUMBER is shared between the determiner and the noun. This sharing can be directly
represented in the graphic notation as follows:
(16) Merging paths in graph notation

where POS again stands for part of speech, ORTH for orthography and AGR for agreement feature.
10 © ISO 2006 – All rights reserved

Here, the two SPECIFIER.AGR and HEAD.AGR paths merge on the node NUMBER, indicating that they
share one and the same feature structure as their value.
In AVM notation, structure sharing is represented by providing a boxed integer label at the point where
structures are shared, as in the following example:
(17) Structure sharing in AVM notation

In a similar way, in the XML notation the element var is used to provide a name for the sharing point. It
supplies a label for the point which can then be referenced elsewhere in the structure, as in the following
example:
(18) Structure sharing in XML notation








...

INTERNATIONAL ISO
STANDARD 24610-1
FIrst edition
2006-04-15
Language resource management —
Feature structures —
Part 1:
Feature structure representation
Gestion des ressources linguistiques — Structures de traits —
Partie 1: Représentation de structures de traits

Reference number
©
ISO 2006
PDF disclaimer
This PDF file may contain embedded typefaces. In accordance with Adobe's licensing policy, this file may be printed or viewed but
shall not be edited unless the typefaces which are embedded are licensed to and installed on the computer performing the editing. In
downloading this file, parties accept therein the responsibility of not infringing Adobe's licensing policy. The ISO Central Secretariat
accepts no liability in this area.
Adobe is a trademark of Adobe Systems Incorporated.
Details of the software products used to create this PDF file can be found in the General Info relative to the file; the PDF-creation
parameters were optimized for printing. Every care has been taken to ensure that the file is suitable for use by ISO member bodies. In
the unlikely event that a problem relating to it is found, please inform the Central Secretariat at the address given below.

©  ISO 2006
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized in any form or by any means,
electronic or mechanical, including photocopying and microfilm, without permission in writing from either ISO at the address below or
ISO's member body in the country of the requester.
ISO copyright office
Case postale 56 • CH-1211 Geneva 20
Tel. + 41 22 749 01 11
Fax + 41 22 749 09 47
E-mail [e-mail]
Web www.iso.org
Published in Switzerland
ii © ISO 2006 – All rights reserved

Contents Page
Foreword. v
Introduction . vi
1 Scope. 1
2 Normative references. 1
3 Terms and definitions. 1
4 General characteristics of feature structure. 4
4.1 Overview. 4
4.2 Use of feature structures . 4
4.3 Basic concepts. 5
4.4 Notations . 5
4.4.1 Overview. 5
4.4.2 Graph notation . 6
4.4.3 Matrix notation . 7
4.4.4 XML-based notation. 8
4.5 Structure sharing. 10
4.6 Collections as complex feature values. 12
4.6.1 Overview. 12
4.6.2 Lists as feature values . 12
4.6.3 Sets as feature values . 14
4.6.4 Multisets as feature values . 15
4.7 Typed feature structure. 16
4.7.1 Overview. 16
4.7.2 Types. 16
4.7.3 Notations . 16
4.8 Subsumption: relation on feature structures . 18
4.8.1 Overview. 18
4.8.2 Definition . 18
4.8.3 Condition A on path values . 19
4.8.4 Condition B on structure sharing . 19
4.8.5 Condition C on type ordering . 20
4.9 Operations on feature structures and feature values. 21
4.9.1 Overview. 21
4.9.2 Compatibility . 21
4.9.3 Unification . 22
4.9.4 Unification of shared structures . 22
4.10 Operations on feature values and types . 23
4.10.1 Concatenation and union operations . 23
4.10.2 Alternation. 24
4.10.3 Negation. 25
4.11 Informal semantics of feature structures. 27
5 XML Representation of feature structures. 29
5.1 Overview. 29
5.2 Organization . 29
5.3 Elementary feature structures and the binary feature value. 30
5.4 Other atomic feature values . 32
5.5 Feature and feature-value libraries . 35
5.6 Feature structures as complex feature values .37
5.7 Re-entrant feature structures . 40
5.8 Collections as complex feature values. 41
5.9 Feature value expressions. 44
5.9.1 Overview. 44
5.9.2 Alternation. 44
5.9.3 Negation. 47
5.9.4 Collection of values . 48
5.10 Default values. 48
5.11 Linking text and analysis . 50
Annex A (informative) Formal definitions and implementation of the XML representation of
feature structures. 54
A.1 Overview. 54
A.2 RELAX NG specification for the module . 54
Annex B (informative) Examples for illustration . 60
Annex C (informative) Type inheritance hierarchies. 62
C.1 Overview. 62
C.2 Definition. 62
C.3 Multiple inheritance . 64
C.4 Type constraints . 64
Annex D (informative) Denotational semantics of feature structure. 66
D.1 Feature structure signatures . 66
D.2 Feature structure algebra. 66
D.3 FS domains. 67
D.4 Feature structure interpretations . 68
D.5 Satisfiability . 68
D.6 Subsumption . 68
D.7 Unification. 69
Annex E (informative) Use of feature structures in applications. 70
E.1 Overview. 70
E.2 Phonological representation. 70
E.3 Grammar formalisms or theories . 70
E.4 Computational implementations . 71
Bibliography . 75

iv © ISO 2006 – All rights reserved

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.
International Standards are drafted in accordance with the rules given in the ISO/IEC Directives, Part 2.
The main task of technical committees is to prepare International Standards. Draft International Standards
adopted by the technical committees are circulated to the member bodies for voting. Publication as an
International Standard requires approval by at least 75 % of the member bodies casting a vote.
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.
ISO 24610-1 was prepared by Technical Committee ISO/TC 37, Terminology and other language and content
resources, Subcommittee SC 4, Language resource management.
ISO 24610 consists of the following parts, under the general title Language resource management — Feature
structures:
⎯ Part 1: Feature structure representation
The following part is under preparation:
⎯ Part 2: Feature system declaration

Introduction
This part of ISO 24610 results from the agreement between the Text Encoding Initiative Consortium (TEI) and
the ISO TC 37/SC 4 that a joint activity should take place to revise the two existing chapters on feature
structures and feature system declaration in The TEI Guidelines called P4.
It is foreseen that ISO 24610 will have the following two parts.
⎯ Part 1, Feature structure representation, describes feature structures and their representation. It provides
an informal but explicit overview of their basic characteristics and formal semantics. In addition, part 1
defines a standard XML (eXtended Markup Language) vocabulary for the representation of untyped
feature structures, feature values, and feature libraries. It thus provides a reference format for the
exchange of feature structure representations between different application systems.
⎯ Part 2, Feature system declaration, discusses ways of validating typed feature structures which are
conformant to part 1, and of enforcing application-specific constraints. It proposes an XML vocabulary for
the representation of such constraints with reference to a set of features and the range of values
appropriate for them, and thus facilitates representation and validation of a type hierarchy as well as other
well-formedness conditions for particular applications, in particular those related to the goal of language
resource management.
vi © ISO 2006 – All rights reserved

INTERNATIONAL STANDARD ISO 24610-1:2006(E)

Language resource management — Feature structures —
Part 1:
Feature structure representation
1 Scope
Feature structures are an essential part of many linguistic formalisms as well as an underlying mechanism for
representing the information consumed or produced by and for language engineering applications. This part of
ISO 24610 provides a format for the representation, storage and exchange of feature structures in natural
language applications concerned with the annotation, production or analysis of linguistic data. It also defines a
computer format for the description of constraints that bear on a set of features, feature values, feature
specifications and operations on feature structures, thus offering a means of checking the conformance of
each feature structure with regards to a reference specification.
2 Normative references
The following referenced documents are indispensable for the application of this document. For dated
references, only the edition cited applies. For undated references, the latest edition of the referenced
document (including any amendments) applies.
ISO 8879, Information processing — Text and office systems — Standard Generalized Markup Language
(SGML), as extended by TC 2 (ISO/IEC JTC 1/SC 34 N029: 1998-12-06).
ISO 19757-2, Information technology — Document Schema Definition Language (DSDL) — Part 2: Regular-
grammar-based validation — RELAX NG
NOTE The first reference permits the use of XML and the second, RELAX NG,provides a specification for XML
modules. RELAX NG is a schema language for XML, standing for REgular LAnguage for XML for Next Generation, and
simplifies and extends the features of DTDs, Document Type Definitions.
3 Terms and definitions
For the purposes of this document, the terms and definitions given in ISO 8879 and ISO 19757-2 and the
following apply. This list is provided to clarify the terminology relating to feature structures used throughout
this part of ISO 24610. Terminology derived from XLM and other formal languages is not defined here.
3.1
alternation
operation on feature values (3.23) that returns one and only one of the values supplied as its argument
NOTE Given a feature specification F : a|b, where a|b denotes the alternation of a and b, F has either the value a or
the value b, but not both.
3.2
atomic value
value (3.23) without internal structure, i.e. value other than feature structure (3.10) and collection (3.4)
3.3
boxed label
label in box used in a matrix notation to denote a value shared by several features (3.8)
NOTE The label may be any alphanumeric symbol.
3.4
collection
list, set, or multiset of values (3.23)
NOTE A list is an ordered collection of entities some of which may be identical. A set is an unordered collection of
unique entities. A multiset is an unordered collection of entities that may or may not be unique; it is sometimes referred to
as a bag.
3.5
complex value
value (3.23) represented either as a feature structure (3.10) or as collection (3.4)
3.6
concatenation
operation of combining two lists of values (3.23) into a single list
3.7
empty feature structure
feature structure (3.10) containing no feature specifications (3.9)
3.8
feature
property of an entity
NOTE The combination of feature and feature-value constitutes a feature specification (3.9). For example, number
is a feature, singular is a value, and a pair is a feature specification.
3.9
feature specification
assignment of a value (3.23) to a feature (3.8)
NOTE Formally, it is treated as a pair of a feature and its value.
3.10
feature structure
set of feature specifications (3.9)
NOTE The minimum feature structure is the empty feature structure (3.7).
3.11
graph notation
notation of feature structure (3.10) in a single rooted graph
3.12
incompatibility
relation between two feature structures (3.10) which have conflicting types (3.19) or at least one common
feature (3.8) with incompatible values (3.23)
NOTE Two feature structures that are incompatible cannot be unified. The empty feature structure (3.7) is
compatible with any other feature structure.
2 © ISO 2006 – All rights reserved

3.13
matrix notation
attribute-value matrix
AVM
notation that uses square brackets to represent feature structures (3.10)
NOTE In a matrix notation, each row represents a feature specification (3.9), with the feature name and the feature
value separated by a colon (:), space ( ) or the equals sign (=).
3.14
merge
generic operation that includes union (3.22) of sets or multisets and concatenation (3.6) of lists
3.15
negation
(unary) operation on a value (3.23) denoting any other value incompatible with it
NOTE In this part of ISO 24610, negation applies to values only and is not understood as a truth function as in
ordinary bivalent logics.
3.16
path
sequence of labeled arcs connecting nodes in a graph
3.17
structure sharing
re-entrancy
relation between two or more features (3.8) within a feature structure (3.10) that share a value (3.23)
3.18
subsumption
relationship between two feature structures (3.10) in which one is more specific than the other
NOTE A feature structure A is said to subsume a feature structure B if A is at least as informative as B. Subsumption
is a reflexive, antisymmetric, and transitive relation between two feature structures.
3.19
type
name of a class of entities
NOTE Feature structures (3.10) may be characterized by grouping them into certain classes. Types are used to
name such classes.
3.20
typed feature structure
feature structure (3.10) labelled by a type (3.19)
NOTE In the graph notation (3.11), each node is labelled with a type. In the matrix notation (3.13), a type is
ordinarily placed at the upper left corner of the inside of the pair of square brackets that represents a typed feature
structure. In XML notation, the type is supplied as the value (3.23) of a type attribute on the element.
3.21
unification
operation that combines two compatible feature structures (3.10) into the least informative feature structure
that contains the information from the two
3.22
union
operation that combines two sets, or multisets, into one
NOTE The corresponding operation for lists is concatenation (3.6).
3.23
value
information about an entity
NOTE There are two kinds of feature values: atomic value (3.2) and complex value (3.5).
4 General characteristics of feature structure
4.1 Overview
A feature structure is a general purpose data structure that identifies and groups together individual features
by assigning a particular value to each. Because of the generality of feature structures, they can be used to
represent many different kinds of information. Interrelations among various pieces of information and their
instantiation in markup provide a meta-language for representing analysis and interpretation of linguistic
content. Moreover, this instantiation allows a specification of a set of features with values of specific types and
1)
restrictions, by means of feature system declarations, or other XML mechanisms discussed in ISO 24610-2 .
4.2 Use of feature structures
Feature structures provide partial information about an object by specifying values for some or all of its
features. For example, if a female employee named Sandy Jones who is 30 years old is of the present
concern, then that person’s sex, name and age can be specified in a succinct manner by assigning a value to
each of these three features. These pieces of information can be put into a simple set notation, as in:
(1) Employee
{, , }
Feature structures are generally used as a vehicle for linguistic descriptions. For example, the phoneme /p/ in
English can be analysed in terms of its distinctive features: consonantal, anterior, voiceless, non-continuant or
stop sound segment, etc. Each of these features may be combined with one or other of the binary values
plus(+) and minus(-) to provide a feature specification. In a phonemic analysis such as the following, the value
of a feature specifies the presence or absence of that feature:
(2) Sound segment /p/
{, , , }
In such an analysis, the sound segment /p/ is distinguished from other phonemes in terms of the presence or
absence of specific features. For example, /p/ differs from the phoneme /b/ in VOICING, and from /k/ in
articulatory position: one is articulated at the anterior, (the lip or alveolar area of the mouth), and the other at
the nonanterior, namely the back of the oral cavity.
This feature analysis can be extended to other kinds of description. Consider a verb like “love”. Its features
include both syntactic and semantic properties: as a transitive verb, it takes an object as well as a subject as
its arguments, expressing the semantic relation of loving between two persons or animate beings. The exact
representation of these feature specifications requires a detailed elaboration of what feature structures are.
For now, these grammatical features can be roughly represented in a set format like the following:
(3) Grammatical features of the verb “love”
{, , },
where POS stands for part of speech.

1) Under preparation.
4 © ISO 2006 – All rights reserved

Since its first extensive use in generative phonology in mid-60s, the feature structure formalism has become
an essential tool not only for phonology, but more generally in support of syntax and semantics as well as
lexicon building, especially in computational work. Feature structures are used to describe and model
linguistic entities and phenomena by specifying their properties. In the next clauses, some of the formal
properties of feature structures are outlined together with means of representing them in a systematic manner.
4.3 Basic concepts
Feature structures may be viewed in a variety of ways. The most common and perhaps the most intuitive
views are the following:
⎯ a set of feature specifications that consists of pairs of features and their values;
⎯ labelled directed graphs with a single root where each arc is labelled with the name of a feature and
directed to its value.
In set-theoretic terms, a feature structure FS can be defined as a partial function from a set Feat of features to
a set FeatVal of values, where FeatVal consists of a set AtomVal of atomic values and a set FS of feature
structures.
(4) A feature structure as a set or partial function
FS ⊆ {⎥F∈Feat,v∈FeatVal}
i i i i
or
FS : Feat ⎯→FeatVal
where FeatVal = AtomVal∪ FS and where FeatVal stands for all possible values.
Values may be regarded as either atomic or complex. Atomic values are entities without internal structure,
while complex values may be feature structures or collections of values (either complex or atomic).
NOTE These two definitions are not absolutely equivalent, for in a set there may be more than one value, v, for a
feature, F.
For example, the part of speech (POS) feature can take the name of an atomic morpho-syntactic category
such as verb as its value. Conversely, the agreement (AGR) feature in English takes a complex value in the
form of a feature structure with features PERSON and NUMBER. The word “loves”, for instance, has a POS
feature with the value “verb”, while the value of its AGR feature consists of a feature structure comprising two
feature specifications: PERSON with the value “3rd”, and NUMBER with the value “singular”.
4.4 Notations
4.4.1 Overview
As a list of feature-value pairs, the overall form of a feature structure is simple. However, the internal structure
of a feature structure may be complex when a feature structure contains either
a) a feature whose value is itself a feature structure, or
b) a multi-valued feature whose value is a list, set, or multiset.
Lists, sets and multisets may be made up of atomic values, or they may include complex values that are
themselves feature structures. That is, feature structures allow limitless recursive embedding. It is therefore
necessary to represent them in an understandable and mathematically precise notation.
There are two commonly used notations for the representation of feature structures: 1) graphs; and 2)
matrices. This part of ISO 24610 proposes a third notation based on the XML standard. The following
subclauses discuss how the same feature structure may be represented using each of these equivalent
notations. Graphs are suitable for mathematical discourses, matrices for linguistic descriptions, and XML
notations for computational implementation.
4.4.2 Graph notation
Feature structures are often represented as labelled directed graphs with a single root.
NOTE This graph can be either
a) acyclic, in which case it is referred to as a directed acyclic graph (DAG), or
b) cyclic for handling cases like the Liar’s paradox.
Feature structures are usually represented as DAGs, but it has been suggested that cyclical feature structures
should be introduced to model some linguistic phenomena.
Each graph starts with a single particular node called the root. From this root, any number of arcs may branch
out to other nodes and then some of them may terminate or extend to other nodes. The extension of directed
arcs shall, however, stop at some terminal nodes. On such a graph, representing a feature structure, each arc
is labelled with a feature name and its directed node is labelled with its value.
Here is a very simple example for a directed graph representing a feature structure.
(5) Feature structure in graph notation

In this graph, the two features, FEATURE1 and FEATURE2, are atomic-valued, taking value1 and value2 on
the terminal nodes respectively as their value. The feature FEATURE3 is, however, complex-valued, since it
takes as its value the feature structure which is represented by the two arcs FEATURE31 and FEATURE32
with their respective values, value31 and value32.
A graph may just consist of the root node, that is, without any branching arcs. Such a graph represents the
empty feature structure.
There may be one or more arcs branching out from the root, each of them bearing a single feature name.
Some of these labelled arcs originating from the root may again stretch out to another node and then from this
node to another, forming an indefinitely long sequence of feature names. Such a sequence of feature names,
labelling the arcs from the unique root node to each of the terminal nodes on a graph, is called a path. For
example, there are four paths in (5): FEATURE1, FEATURE2, FEATURE3.FEATURE31 and
FEATURE3.FEATURE32.
6 © ISO 2006 – All rights reserved

Here is a linguistically more relevant example.
(6) Linguistic example in graph notation

This graph consists of three paths: orthography (ORTH), SYNTAX.POS and SYNTAX.VALENCE. The path
consisting of a single feature name ORTH is directed to the terminal node “love”, which is an atomic value.
The path SYNTAX.POS terminates with the atomic value “verb” and the path SYNTAX.VALENCE with the
atomic value “transitive”. The nonterminal feature SYNTAX takes a complex value, namely a feature structure
consisting of the two feature specifications, and .
4.4.3 Matrix notation
Despite their mathematical elegance, graphs are difficult to typeset or read, especially when complex. For this
reason, feature structures are more often depicted using a matrix notation called attribute-value matrix, or
simply AVM.
NOTE 1 The term “attribute” is an alternative term for the concept which is called “feature”.
(7) Matrix notation
Each row in the square bracket with a FEATURE name followed by its value name represents a feature
specification in a feature structure.
NOTE 2 A colon, an equals sign or a space separates a feature from its value on each row of an AVM.
Feature values can be either atomic or complex. Each row with an atomic value terminates at that value. But if
the value is complex, then that row leads to another feature structure, as in the case of FEATURE3 above.
The notion of path is also important in several applications of the attribute-value matrix (AVM) notation, as
discussed further below. A path in an AVM is a sequence of feature names, as is the case with feature
structure graphs. An AVM with no rows and thus no occurrences of features, is represented as [ ]; this
represents the empty feature structure. If an AVM has at least one row consisting of a feature name and its
value, then there is a path of length 1 corresponding to each occurrence of a feature name in each row. Given
a path of length i, if the last member of that path takes a nonempty feature structure as value, then that path
forms a new path of length i +1 by taking each one of the features in that feature structure as its member.
For illustration, consider the following AVM which represents the same feature structure as is represented by
the graph notation (6).
(8) Example of an AVM notation

This AVM has three paths: ORTH, SYNTAX.POS and SYNTAX.VALENCE.
4.4.4 XML-based notation
Like any other formalism, XML has its own terminology. An XML document consists of typed elements,
occurrences of which are marked by start- and end-tags which enclose the content of the element. Every XML
element may be regarded as a linearization of a tree with a single root node. Each node contains either
terminal text or other element nodes. Elements have a specific type. Element occurrences can also carry
named attributes (or, more exactly, attribute-value pairs). For example:
(9) love
This is the XML way of representing an occurrence of the element type word, the content of which
is the string “love”. The word element is defined as having an attribute called “class”, whose value
in this case is the string “noun”.
The term “attribute” is thus used in a way which potentially clashes with its traditional usage in discussions of
feature structures. In AVM, feature structures are represented in a square bracket form, with each row
representing an attribute-value pair. In this usage, an attribute does not correspond necessarily to an XML
attribute. To avoid this potential source of confusion, the reader should be aware that in the remainder of this
clause, the term “attribute” will be used only in its technical XML sense.
To illustrate further the distinction, it should now be considered how a feature structure in AVM may be
represented in XML. Consider the following AVM that represents a non-typed feature structure:
(10) Representation of a feature structure in AVM

This feature structure consists of two feature specifications: one combines the feature ORTH with the string
value “love” and the other combines the feature POS with the string value “noun”.
NOTE This representation says nothing about the range of possible values for the two features: in particular, it does
not indicate that the range of possible values for the ORTH feature is not constrained, whereas (at least in principle) the
range of possible values for the POS feature is likely to be constrained to one of a small set of valid codes. In a constraint-
based grammar formalism, possible values for the feature ORTH need to be strings consisting of a finite number of
characters drawn from a well-defined character set, say the Roman Alphabet plus some special characters, for each
particular language.
In the XML representation proposed here, a feature structure is represented by means of an XML element fs,
and a feature-value specification by means of an XML element f:
(11)
.
.

8 © ISO 2006 – All rights reserved

This is a more generic approach than another, equally plausible, representation, in which each feature
specification might be represented by a specifically-named element:
(12)
.
.

Using this more generic approach means that systems can be developed which are independent of the
particular feature set, at the expense of slightly complicating the representation in any particular case. By
representing the specification of a feature as an f element, rather than regarding each specific feature as a
different element type, the overall processing model is simplified considerably.
A feature specification, as has been seen, has two components: a name, and a value. Again, there are
several equally valid ways of representing this combination in XML. Any of the following three, for example,
could be chosen:
(13)
a)  ;
b) 
noun
;
c) 
pos
noun
.
The fact that all three of these are possible arises from a redundancy introduced in the design of the XML
language largely for historical reasons. The present recommendation is to use a formulation like b) above,
though c) may be preferable on the grounds of its greater simplicity.
Finally, the representation of the value part of a feature specification is discussed. A name is simply a name,
but a value in the system discussed here may be of many different types: it might for example be an arbitrary
string, one of a predefined set of codes, a Boolean value, or a reference to another (nested) feature structure.
In the interests of greater expressivity, the present system proposes to distinguish amongst these kinds of
value in the XML representation itself.
Once more, there are a number of more or less equivalent ways of doing this. For example:
(14)
a) 

love

noun
;
b) 
love

.
The number of possible different value types is comparatively small, and the advantages of handling values
generically rather than specifically seem less persuasive. The approach currently taken is therefore to define
different element types for the different kinds of value (for example, string to enclose a string, symbol to
denote a symbol, binary to denote a binary value, fs to denote a nested feature structure).
Representing feature structures in XML notation is thus possible and rather straightforward. For example, the
following is the XML representation for the feature structure given in (6) and (8) above.
(15) Feature structure in XML notation


love







Despite the apparent difference between the representations (6), (8) and (15), the information contained in
each is identical. For example, the arc labels in graph notation like ORTH and SYNTAX are now represented
by tags with appropriate names. Furthermore, the XML representation can be processed by general
purpose software tools and thus can be converted automatically to a variety of other formats.
NOTE A more detailed discussion including ways to simplify the representation in (15) is provided in Clause 5.
4.5 Structure sharing
A feature structure representation may need to represent shared components. For example, in representing
the structure of a noun phrase (“la pomme") which includes a determiner (article) “la”, it may be desired to
show that the feature NUMBER is shared between the determiner and the noun. This sharing can be directly
represented in the graphic notation as follows:
(16) Merging paths in graph notation

where POS again stands for part of speech, ORTH for orthography and AGR for agreement feature.
10 © ISO 2006 – All rights reserved

Here, the two SPECIFIER.AGR and HEAD.AGR paths merge on the node NUMBER, indicating that they
share one and the same feature structure as their value.
In AVM notation, structure sharing is represented by providing a boxed integer label at the point where
structures are shared, as in the following example:
(17) Structure sharing in AVM notation

In a similar way, in the XML notation the element var is used to provide a name for the sharing point. It
supplies a label for the point which can then be referenced elsewhere in the structure, as in the following
example:
(18) Structure sharing in XML notation




















Re-entrancy is symmetric and so there is no reason to distinguish amongst the various occurrences of a
shared node. In particular, this implies that a value may be attached to any or all occurrences of a shared
label:
(19) Specifying shared values




















A particularly significant case of structure sharing arises when the aim is to express that two features have the
same value, even though that value is not known or is underspecified.
4.6 Collections as c
...