A Review of Multilingualism in and for Ontologies_2

2025-04-27 0 0 805.78KB 22 页 10玖币

侵权投诉

A Review of Multilingualism in and for Ontologies

Frances Gillis-Webbera, C. Maria Keeta

aDepartment of Computer Science, University of Cape Town, Cape Town, 7701, South Africa

Abstract

The Multilingual Semantic Web has been in focus for over a decade. Multilingualism in Linked Data and RDF has shown

substantial adoption, but this is unclear for ontologies since the last review 15 years ago. One of the design goals for OWL was

internationalisation, with the aim that an ontology is usable across languages and cultures. Much research to improve on

multilingual ontologies has taken place in the meantime, and presumably multilingual linked data could use multilingual

ontologies. Therefore, this review seeks to (i) elucidate and compare the modelling options for multilingual ontologies, (ii)

examine extant ontologies for their multilingualism, and (iii) evaluate ontology editors for their ability to manage a multilingual

ontology.

Nine diﬀerent principal approaches for modelling multilinguality in ontologies were identiﬁed, which fall into either of the

following approaches: using multilingual labels, linguistic models, or a mapping-based approach. They are compared on design by

means of an ad hoc visualisation mode of modelling multilingual information for ontologies, shortcomings, and what issues they

aim to solve. For the ontologies, we extracted production-level and accessible ontologies from BioPortal and the LOV repositories,

which had, at best, 6.77% and 15.74% multilingual ontologies, respectively, where most of them have only partial translations

and they all use a labels-based approach only. Based on a set of nine tool requirements for managing multilingual ontologies, the

assessment of seven relevant ontology editors showed that there are signiﬁcant gaps in tooling support, with VocBench 3 nearest

of meeting them all. This stock-taking may function as a new baseline and motivate new research directions for multilingual

ontologies.

Keywords: Multilingual ontologies, Multilingual ontology management, OWL

1. Introduction

The Semantic Web was envisaged as an extension of the

World Wide Web [1]. To support this vision, several languages

were developed, notably Resource Description Framework

(RDF) and Web Ontology Language (OWL). We focus on the

latter and in particular its internationalisation goal [2, 3], with

multilingualism as a common component in an interconnected

world.

The Semantic Web landscape has grown substantially since

the 2000’s, and its multilinguality since circa 2010 with the

‘Multilingual Semantic Web’ (MSW) and ‘Multilingual Web

of Data’ (MWD) being two widely used terms in the literature

[4, 5, 6, 7]. Indexed research literature illustrate this; e.g., a

Google Scholar search on the date range 2001–2005 returned

14 hits for MSW and zero for MWD, for 2006–2010, it

increased to 27 hits for MSW and still with zero for MWD, for

2011–2015 it comparatively exploded to 296 for both terms

combined, with another 298 for the period 2016–2020. This

increase is not as evident for the various stock-takings of

multilingual ontologies and related artefacts speciﬁcally

[8, 9, 10], as shown in Figure 1. Those 2004 and 2007 reviews

∗Corresponding author

Email addresses: fgilliswebber@cs.uct.ac.za (Frances

Gillis-Webber), mkeet@cs.uct.ac.za (C. Maria Keet)

on ontologies [8, 9] focussed on language-tagged strings,

which was practically the only option available in the early

years of the Semantic Web, and do not mention the extent of

label coverage within an ontology for each natural language,

relative to its total class and property axiom count. The very

recent review for BioPortal-indexed ontologies [10] considered

only domain ontologies with ‘production’ status on limited

metrics.

This raises several questions, including whether the increase

in research on the Multilingual Semantic Web has translated to

an increase in the ratio of multilingual ontologies in a

repository, compared to the latest review in 2007, as it did for

support for natural languages in RDF datasets and knowledge

graphs (elaborated on below in Section 2). With this review we

seek to answer the following main questions (motivated

afterward):

RQ1 What are the available modelling options to develop a

multilingual ontology?

RQ2 Regarding extant ontologies and multilingual ontologies:

RQ2a What are the modelling approach(es) used for

extant multilingual ontologies?

RQ2b What is the percentage of multilingual OWL

ontologies compared to monolingual ones (in

popular repositories)?

RQ2c What is the percentage of natural language

completeness of each multilingual ontology?

Preprint submitted to TBA October 7, 2022

arXiv:2210.02807v1 [cs.AI] 6 Oct 2022

RQ2d How does that compare the the latest stock-taking,

of the 2007 review?

RQ3 What is the status of tooling support for developing

multilingual ontologies?

First, the aforementioned increase in research papers does

include numerous approaches for modelling language

information in or for ontologies well beyond a simple

language-tagged string, which may beneﬁt from a systematic

comparison (RQ1). This is also needed in order to be able to

ﬁnd multilingual ontologies and to assess them (RQ2) on

which multilingual approaches are actually used, and how

many multilingual ontologies there are. Their presence, or

absence, may well be linked to tooling support to be able to

conveniently develop and maintain them, and, hence, RQ3.

To answer the questions, the ﬁrst step was to analyse

proposed and related variant approaches to multilingualism in

ontologies. To facilitate comparison, we devised a

visualisation notation for them. Nine main approaches were

identiﬁed, which can be grouped into the categories of

multilingual labels, linguistic models, and mapping-based

strategies. For the assessment of ontologies, we consulted

NCBO BioPortal1, a repository of biomedical ontologies [11],

since there has been wide adoption of (OWL) ontologies in the

biomedical domain since the early 2000s [12, 13, 14], and

Linked Open Vocabularies (LOV)2, a curated repository of

RDF and OWL vocabularies that is not limited to any one

speciﬁc domain. Further, we looked beyond the mere number

of multilingual ontologies, to also assess how multilingual the

more-than-one-language ontologies are, introducing the notion

of coverage within an ontology and language-speciﬁc

completeness. Only a few of the multilingual modelling

methods are used, principally just in the labels category, and

even less can be considered fully multilingual where there is

more than one primary language. Tooling for developing and

managing multilingual models is also very limited, where of

the several relevant ones, none meets all the requirements

speciﬁed.

In the remainder of this paper, we ﬁrst describe related work

and key reviews of RDF datasets and knowledge graphs in

Section 2. The ways to model multilinguality is described and

compared in Section 3. This is followed by the BioPortal and

LOV evaluations in Section 4, and the tools assessment in

Section 5. A discussion of the reviews and speciﬁc answers to

the research questions is provided in Section 6. The paper

concludes with Section 7.

2. Related Work

As noted in the Introduction, the sampling of counts of

multilingual ontologies has been sparse, with just [8, 9, 10].

We therefore cast the net wider to also include RDF datasets

and knowledge graph literature pertaining to multilingualism.

The ones with such data are included also in Figure 1 and they

are brieﬂy discussed here.

1https://bioportal.bioontology.org/ontologies

2https://lov.linkeddata.es/dataset/lov/

Ell et al. [15] conducted a review on the Billion Triple

Challenge (BTC) 2010 corpus, which contained over 3 billion

triples, including metadata on the source of the resource the

triple was crawled from. Excluding metadata, 1.4 billion

distinct triples remained and of these triples, 0.7% were

identiﬁed to contain two or more language tags [15]. This was

followed-up in 2018 with a cross-sectional study of labels

across seven datasets [16], including, among others, the

BTC 2010 corpus (with the metadata), a 2014 version of BTC

(4 billion triples), and a 2017 version of Wikidata (2 billion

triples). BTC 2014 was found to support 183 languages (up

from 2010’s 55) and the Wikidata dataset was found to support

424 languages [16]. That is: languages other than English

certainly are being used in the Semantic Web. Due to the way

the results were reported, however, it is not possible to

determine an accurate percentage of the number of triples

containing two or more multilingual labels, but it is an

approximate 6% for BTC 2014 and 40% for Wikidata3.

Coverage was measured for languages together, rather than by

language.

Other dataset and knowledge graph assessments also

indicate substantial use of multiple languages. Notably, the

2015-10 version of Wikidata was found to support 395

languages in 749 million triples, the 2015-04 version of

DBpedia supported 13 languages with 412 million triples, and

Yet Another Great Ontology (YAGO), which integrates

statements of the diﬀerent language versions of Wikipedia, and

YAGO 3 (with the 2014 Wikipedia dump) was found to

support 326 languages in 1 billion triples [17]. Completeness

was discussed as a data quality dimension on whether the

knowledge graph was suited to the task at hand, within the

context of its data. The coverage of selected languages

(English, French, German, Spanish, and Italian) was

considered relative to the annotations: German and French had

a coverage of over 30% in Wikidata, and German of over 10%

in YAGO.

Given the increase in data in multiple languages and an

increase in multilingualism over the years, one may expect an

increase in multilingual ontologies as well.

3. Modelling Options for Multilingual Ontologies

Several options to develop multilingual ontologies have been

identiﬁed [5, 18, 19]:

1. Multilingual labels: the ontology vocabulary is annotated

with language-tagged strings.

2. Linguistic models: the ontology is associated with a

linguistic model external to it.

3. Mapping models: common interlingua which can be

speciﬁed between one or more monolingual ontologies.

3The data was reported in groups of 2, 2–5, 5–10, and >10. 2 was ignored

as it is included in 2–5. The sum of the groups for BTC 2014 was 6.3% and

for Wikidata, 4.4%. Both values were rounded down due to the unavoidable

double-counting of 5.

Figure 1: Timeline of the MSW (T1) and MWD (T2) searches, contrasted with reviews pertaining to multilingualism conducted for the period 2001–2021: available

data for ontologies, RDF datasets (collections of triples) and knowledge graphs (RDF graphs) are shown in the blue, red and green bars respectively.

Table 1: Comparison of the Main Approaches for Modelling Multilinguality in Ontologies

Labels Linguistic Models †Mapping Models

Possible OWL proﬁles All All All

Can account for cross-lingual terms where there is no 1-1

mapping

No Within the

annotation only

Yes

Can support inﬂectional languages No Limited No

Annotations reusable by other resources ±(not if the annotation

is a data literal)

Yes No

Example axiom count for two natural languages in an ontology ‡2≥14 7

Annotations contained within the OWL ﬁle Yes No No

Number of ﬁles to keep in sync 1 Minimum 3 Minimum 2

Can be managed by an ontology editor Yes Limited support Limited support

†Values are for OntoLex-Lemon only for the inﬂectional languages, axiom count, and ontology editor.

‡The axiom count was determined for a class in an ontology for two natural languages. Language-speciﬁc grammatical features were excluded.

OntoLex-Lemon was used for the linguistic model and the simplest method was used, that is, associating two lexical entries with a class. For

the mapping model, OWL 2 was used.

A brief comparison of these approaches is included in Table 1.

We will elaborate, compare, and discuss each approach in turn

in the subsections below.

In order to represent the options and sub-variants

schematically to facilitate comparisons, we use an ad hoc

visual language. This consists of a set of primitives. First, an

ontology can be visualised as two layers, which separate the

semantic and the linguistic components, in line with [20]. The

semantic layer, which contains the classes, object properties

and other axioms, is indicated with a dark grey box that has

further information, such as the type of identiﬁer (opaque or

descriptive). The linguistic layer, which typically contains the

annotations used to associate language information with an

ontology, is indicated by a light grey box, and contains the

various options for the language annotations. An example

visualisation is shown in Figure 2 for a monolingual ontology

with descriptive identiﬁers, i.e., naming the vocabulary

elements with a string that is in a natural language or natural

language-like (indicated with Ln) and meaningful to a human,

such as naming a class Person or RumAndRaisinIcecream

instead of ABC:0000012. The knowledge represented in the

TBox of the ontology may have a natural language-speciﬁc

vocabulary either in or attached to the ontology as a separate

ﬁle. It is possible that two monolingual ontologies in two

diﬀerent natural languages may share the same

conceptualisation [21].

To show the properties and cardinality constraints between

the two layers and their objects, the crow’s feet notation from

Entity-Relationship modelling is used. The notation is

extended to indicate the modelling of language-speciﬁc label

instances. We also introduce a relation that is used to indicate

when two elements must have structural equivalence. This

pertains speciﬁcally to the diﬀerent annotation values available

in OWL 2. See Figure 2 for a legend of the diﬀerent relations.

We now proceed to the three groups of options, going from

the basic to more elaborate options.

3.1. Multilingual Labels

Multilingual labels are the ‘low-hanging fruit’, being the

easiest way to develop a multilingual ontology. An entity in an

ontology (i.e., a class, object property, data property, or

individual) can have multilingual labels by using rdfs:label,

Figure 2: Demonstration of the language to visualise the interaction between

ontologies, any identiﬁers, and natural language aspects, shown here for a

monolingual ontology that uses descriptive identiﬁers. The semantic and

linguistic layers of the ontology are indicated by the grey boxes; the names

of the classes (and other vocabulary elements) are in a natural language Ln.

The possible relations with their constraints are shown in the legend.

which takes as domain rdfs:resource and as range an

rdfs:Literal. Each such label is language-tagged. The

language tag typically consists of an ISO 639 language code4.

However, as ISO 639’s Parts 1-3 [22] do not provide language

codes for all the world’s languages, the language tag can be

extended to capture other information, as long as it conforms

to IETF’s BCP 47 [23, 24]. When this is insuﬃcient [25],

other language models, such as MoLA [26], may be added to

specify additional languages or lects.

An example OWL fragment of this labelling approach is

shown in Listing 1 in Functional-Style syntax (FSS). (Unless

explicitly stated otherwise, all further examples are in FSS.) In

this example, the class Person has two annotations: one in

English and the other in Dutch. Both annotations are a

language-tagged string.

1SubClassOf (: Person owl : Thi ng )

2A nn o ta ti o nA s se rt i on ( r df s : l abe l : P ers on " P er son " @ en )

3A nn o ta ti o nA s se rt i on ( r df s : l abe l : P ers on " P er so on " @ nl )

Listing 1: OWL fragment illustrating multilingual labels

4https://www.iso.org/iso-639-language-codes.html

The class :Person is a named entity, where Person is the

identiﬁer portion of the URI. An identiﬁer of a URI can be

opaque or descriptive [27, 28, 29]. An opaque identiﬁer is

semantic-free (non-meaningful) and used in a number of

ontologies; e.g., [27, 30, 31]. If an opaque identiﬁer is used as

part of the URI, then an additional sign in the form of an

rdfs:label annotation should be added for both humans and

machines to interpret the URI. For a descriptive identiﬁer,

there is a direct relationship between the natural language term

used as an identiﬁer and its semantics, e.g., [30, 32]. The

descriptive identiﬁer is typically in the primary language of the

ontology.

Figure 3 shows a schematic representation for the TBox for

three ways of realising multilingual ontologies with this

approach. Consider the TBox of O

LI , which presents the

sub-variant where a class has an opaque identiﬁer and at least

one natural language label, with each label a diﬀerent

language. An example OWL fragment is given in Listing 2.

1SubClassOf ( :49 3 Dk ow l : Th ing )

2Annot a t i o nAsse r t i o n ( rdfs : label :493 Dk " Per s on " @ en )

Listing 2: OWL fragment with an opaque identiﬁer

An entity can have any number of labels, where iis a label and

L is the set of languages used for an entity: (L1... Ln) with

1≤i≤n. For a primary language ontology, an entity can either

have an opaque or descriptive identiﬁer in Li.

Because meaning can be surmised by a human from a

descriptive identiﬁer, a label is not required. However, as

shown in O

PLD in Figure 3, if the ontology is multilingual, all

entities may have labels in one language, with the language tag

sometimes omitted. Where there are labels for other

languages, these are translations from the primary language.

PLO (Figure 3) has an opaque identiﬁer, but since all entities

have labels in one language and to a limited extent, labels for

other languages, we consider it to be a primary language

multilingual ontology.

The adaptation of a monolingual ontology to another

language is called ontology localisation [33], although this is

used for diﬀerent contexts as well, such as a culture or a

geo-political environment for which there is typically a shared

language [34, 35, 36]. If an ontology is localised using labels,

then only the linguistic layer of the ontology is aﬀected, and

the concept space of the original ontology remains unchanged.

Despite the diﬀerences with the labels and identiﬁers

between each of the TBoxes in Figure 3, O

LI ,O

PLD and O

PLO

have commonality in that they share the same so-called

‘concept spaces’, which can be seen as a ﬁgurative area of a

concept, from a ﬁne-grained concept to a category of concepts,

where a concept space applies to each OWL class and object

property; e.g., the notion of house and the colour blue. Natural

languages may divide up the same concept space diﬀerently,

which may indicate underlying ontological distinctions. One

language may underspecify it compared to that of another. For

instance, there is only one notion of ‘River’ in English,

whereas the French equivalent distinguishes between two

types of river (Rivi`ere and Fleuve). Another example is the

representation of part-whole relations in isiZulu when

Figure 3: Three principal options for multilingual labels in the TBox of the ontology: fully language independent (indicated with L0) with a mandatory requirement

for at least one label in some language (any one of L1. . . Ln) (left), a primary language for the ontology (Ln), using a descriptive identiﬁer and optional labels for

other languages (centre), and mostly language-independent (L0in the semantic later) but with the requirement to have at least one label in one speciﬁed language

(right)

compared to English, which also revealed ontological

distinctions, in that the list of ‘universal’ part-whole relations

required both generalisation and reﬁnements in order to

accurately represent them [37].

Thus the use of labels has its limitations. In addition, not all

grammatical features for a natural language can be dealt with

by a label, such as inﬂected forms, concordial agreement, and

agglutination (where words are formed with the addition of

aﬃxes to a word root or stem). Inﬂectional forms on nouns

include the grammatical categories number and gender, and

tense, aspect and number on verbs, which are typically used

for naming classes and properties, respectively. For each of

these grammatical categories, the word formation may need to

change. An example is the Organization Ontology [38], which

provides support for English, Spanish, French, Italian and

Japanese. For instance, the property org:changedBy with the

English label changed by has labels for most of the supported

languages, but there are two for Spanish: es modiﬁcada por

and es modiﬁcado por. This is to account for grammatical

gender which is determined by the gender of the noun of the

name of the class that is in the position of the domain in the

axiom. Similarly for plural terms, the concordial agreement in

a language such as isiXhosa (a Niger-Congo B (‘Bantu’)

language spoken in South Africa), combined with its

agglutination, will alter the surface realisation of the

annotation depending on its domain and range. Indeed, as

highlighted by Keet and Khumalo, highly agglutinative

languages can represent a challenge in ontologies, where it is

possible that no single human-readable label can be prescribed

to a property, due to the use of context-dependent aﬃxes that

modify the entity’s name or their label [37].

It may be the case that more than one label for a single

language has to be added, fo which Simple Knowledge

Organization System (SKOS) can be used. SKOS has the

prefLabel and altLabel properties (subsumed by

rdfs:label) where labels can be identiﬁed as preferred or

alternative, respectively. rdfs:label does not have

cardinality restrictions, but skos:prefLabel can only be

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AReviewofMultilingualisminandforOntologiesFrancesGillis-Webbera,C.MariaKeetaaDepartmentofComputerScience,UniversityofCapeTown,CapeTown,7701,SouthAfricaAbstractTheMultilingualSemanticWebhasbeeninfocusforoveradecade.MultilingualisminLinkedDataandRDFhasshownsubstantialadoption,butthisisunclearforontolo...

展开>> 收起<<

A Review of Multilingualism in and for Ontologies_2.pdf

共22页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Review of Multilingualism in and for Ontologies_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: