Consistency and Interoperability on Dublin Core Element Values in
Collections Harvested using the Open Archive Initiative Protocol for
Metadata Harvesting
Sarantos Kapidakis
a
Department of Archival, Library and Information Studies, University of West Attica,
Keywords:
Dublin Core, Metadata, OAI-PMH, Harvesting, Language, Controlled Vocabularies, Controlled Terms,
Repeated Values, Linked Open Data, Dendrogram.
Abstract:
When resource descriptions use the exact same value for an entity, this value is easier parsed, identified and
utilized by automatic procedures. The use of controlled values, even when it is common and very useful, it
is usually not enforced during the data entry. In this paper we study the use of the controlled values in many
harvested collections and we study all Dublin Core elements and also their similarity. We mainly focus in the
element language, as there is a lot of standardization on how to denote language values, followed by other
elements that normally use controlled values. We discovered values that are repeated many times and in many
collections and many more values that are used only once! The lack of coordination among collections during
their creation results to many variations for each value, even when the value is used consistently and many
times inside a collection. The study uses dendrogram to reveal the current usage of the Dublin Core elements
inside and among active collections by clustering the collections with similar values and helps adopting better
guidelines, designing better tools and improving the effectiveness of the collections.
1 INTRODUCTION AND
RELATED WORK
In many cases, metadata elements of many resources
may share the same value. E.g. some resources may
have the same creator, who should always be denoted
with the exact same value, so that all these resources
will be retrieved on each appropriate search request.
Many elements can take controlled values and of-
ten most of them do. The type element takes con-
trolled values most often, and has many value repe-
titions, and less variety of values, been followed by
the element language. But the values of element lan-
guage are more inter-operable, as there are more stan-
dards and good practices defining them.
Elements expressing language and type of the re-
source often take controlled values, forming groups of
resources with the same values. Many metadata ele-
ments can use controlled values and the library guide-
lines define which metadata elements should only use
controlled values, and how these values should be se-
lected.
a
https://orcid.org/0000-0002-8723-0276
When controlled values are not used, metadata
creation may be easier and natural language process-
ing tools can be used later to identify the denoted en-
tities. Although this approach provides a huge im-
provement in the retrieval procedure, the errors that
can be introduced are not acceptable in many cases
by the library tradition, and libraries continue to use
controlled values.
While libraries concentrate on improving their
own collections, there are issues that arise when
searching across different services, and Maltese
in (Maltese, 2018) described the approach that the
University of Trento followed. Libraries develop and
use various controlled vocabularies for years, and in
(Harper and Tillett, 2007) Harper at al. explains
how such tools that were developed by the Library
of Congress can be used in the development of robust
Web services and Semantic Web technologies.
The Linked Open Data
1
(LOD) defines a mod-
ern trend to universally adopt controlled terms, where
each entity of interest (e.g. person, subject, geo-
graphic location) has a unique URL that should be
1
http://linkeddata.org
Kapidakis, S.
Consistency and Interoperability on Dublin Core Element Values in Collections Harvested using the Open Archive Initiative Protocol for Metadata Harvesting.
DOI: 10.5220/0010112001810188
In Proceedings of the 12th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2020) - Volume 2: KEOD, pages 181-188
ISBN: 978-989-758-474-9
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
181
adopted and used to represented it, eliminating any
ambiguities and permitting easy correlation and nav-
igation to resources through their common entities.
Similar issues apply to the whole Cultural Heritage
Information, and in (Baca, 2003) Baca deals with
practical issues in applying metadata schemes and
controlled vocabularies. For scientific data, the FAIR
Guiding Principles, as described by ”Wilkinson et al.
in (Wilkinson, 2016), also put specific emphasis on
enhancing the ability of machines to automatically
find and use the data, which includes unique identi-
fiers, such as controlled values.
An approach to metadata quality evaluation is
applied to the open language archives community
(OLAC) in (Hughes, 2005) by Hughes that is using
many OLAC controlled vocabularies.
The evaluation and quality of metadata is exam-
ined as one dimension of the digital library evalua-
tion frameworks and systems in the related literature,
like (Moreira et al., 2009; Zhang, 2010). Addition-
ally, Fuhr et al. in (Fuhr et al., 2007) and Vullo et al.
in (Vullo et al., 2010) propose a quality framework
for digital libraries that deal with quality parameters.
Kir
´
aly at al, in (Kir
´
aly et al., 2019), examine the data
quality in Europeana, for the purpose of multilingual-
ity, and they consider the consistently of the Dublin
Core language element in these records.
In (Kapidakis, 2018) Kapidakis studies the pres-
ence and the repetitions of the values of the meta-
data elements from many harvested metadata, ex-
amining the number of their statements and the text
they contain. He also studied how they are evolv-
ing over time. In (Harper, 2016) Harper processes
the items of the Digital Public Library of America
(DPLA) and demonstrates the “metadata fingerprints”
(D3 Star Plots) to visualize the metadata characteris-
tics. He used them to summarize the number of meta-
data statements per item from different providers,
across multiple fields, and also to compare the sig-
natures of the items versus those with at least 1 hit,
using google analytics. He also used these “finger-
prints” to visualize the word counts, to comparatively
study the different Dublin Core elements. He did not
try to process, examine the consistency or find pat-
terns in repeated or controlled values. In (Kapidakis,
2019) Kapidakis studies the presence and the repeti-
tion of the values of the 15 Dublin Core elements, and
comments on individual unexpected values, using a
1000 record sample from each collection. He found
that most Dublin Core elements are using controlled
values in some collections, and are using free values
in many others and that the values in a collection need
better conforming to common rules.
In this paper we study the values of each Dublin
Core element in many collections, where we mainly
focus in the elements that their values have many
repetitions, such as the language element followed
by type, format, rights, coverage and publisher. We
show how the effort to conform to known practices or
standards form groups of values with similar patterns.
We examine records from publicly available official
collections that will normally be used in combination
with other collections.
In order to study the Dublin Core metadata we first
have to collect them from the services that provide
them. We run a harvesting task for each service, ask-
ing for all its records. In many cases such tasks fail,
due to permanent or temporary errors. The records
that were returned constitute a collection, even an in-
complete one, in case of an error in the communica-
tion resulting to less records.
The rest of the paper is organized as follows: In
section 2 we describe our methodology and how we
selected our services and used the software we made
to create our data-set, and we examine general char-
acteristics of the harvested metadata. In section 3 we
study the individual values for the Dublin Core ele-
ment language in all metadata records that the har-
vesting tasks returned for all collections. We also
study how the controlled values are unique or com-
mon over many collections, when all metadata are
used as one collection. In section 4 we examine
the sets of values for the Dublin Core element lan-
guage used in each collection and we present a den-
drogram showing their similarity, to reveal the differ-
ences among different services. Similar analysis is
performed for the other Dublin Core elements, and in
section 5 we present dendrograms for some other el-
ements, to demonstrate the differences in the patterns
of their values. Finally on section 6 we conclude and
present issues for further research.
2 HARVESTING
METHODOLOGY AND
CHARACTERISTICS OF THE
DATA
To harvest and study the provided metadata, we cre-
ated an OAI-PMH client using the oaipy library and
used it to ask each service from a list of OAI-PMH
services to provide all its metadata records, and to
process them. Harvesting tasks are common for the
OAI-PMH services, which periodically satisfy har-
vesting requests for the new or updated records, and
involve the exchange of many OAI-PMH requests and
responses, starting from the negotiation phase for the
KEOD 2020 - 12th International Conference on Knowledge Engineering and Ontology Development
182
Figure 1: The number of records retrieved per collection for
the 2251 collections.
supported OAI-PMH features of the two sides.
The sources listed on March of 2020 in the offi-
cial OAI-PMH Registered Data Providers
2
site were
used as the list of services to harvest and contained
4627 entries. Sometimes the tasks time out resulting
to abnormal termination of the task. Instead of using
a flat timeout deadline, which would be inappropriate
to harvest a large number of records, we set partial
timeout deadlines of 15 minutes for each task, which
would be extended many times (up to 1 day - we never
reached this hard deadline) as long as the task did re-
turn any new records. We interrupted any incomplete
task when it stopped delivering records and moved on
to the next task.
A significant part of the conducted Open Archive
Initiative harvesting services did not respond. As a re-
sult, only 2251 services returned records. We did not
try to recover from harvesting errors or to restart any
failing harvesting tasks, but we processed the records
we received. Overall, our sequential execution of the
record harvesting tasks from their services took more
fifteen days to complete.
In Fig. 1 we can see the distribution of the num-
ber of returned records for each service. Most collec-
tion are small, but there are much bigger ones among
them. In Fig. 2 we depict the number of statements
per collection, sorted in decreasing order, for each dc
element. We can see similar and even overlapping
curves for the elements, where most collections have
very few statements and few collections have most of
the statements.
To better observe the distribution, in Fig. 3 we
only show the number of language statements, sorted
in decreasing order. Only 2062 collections contained
language statements and 2010 of them contained at
2
https://www.openarchives.org/Register/BrowseSites
Figure 2: The number of statements for the 15 dc elements,
in the 2251 collections.
Figure 3: The number of language statements, for the 2062
collections.
least 10 language statements.
A big challenge was to harvest the metadata
needed for such a study. Enough data have to be
collected despite of temporary or permanent errors.
The data we used are all publicly available, but they
change slightly over time. Getting as many records as
possible is essential to get safe conclusions. Overall,
the amount of information harvested from these ser-
vices was huge, and as it was difficult to store, process
and manage, we kept part of it, only the one that was
needed for our study. In fact, we stored and processed
(in memory) about 28GB.
Our harvested metadata consists of 19713013
records with 356105991 statements. In Table 1 we
show, for each Dublin Core element (column ele-
ment), the number of collections that include such
statements (column collections), and the number of
statements in all these collections (column state-
Consistency and Interoperability on Dublin Core Element Values in Collections Harvested using the Open Archive Initiative Protocol for
Metadata Harvesting
183
Table 1: Summary data for the occurrences of each DC el-
ement in the 2251 collections. The elements with most re-
peated values are first.
element collections statements values across
language 2062 16840201 15328 1029
type 2229 25921548 13089 1982
format 2064 13494470 604885 28353
rights 1972 15024492 364138 71249
subject 1896 54073813 7493705 1383596
relation 1919 13139458 7365138 366647
coverage 403 5850572 474233 17530
publisher 2204 14638984 690294 85598
creator 2241 62918155 11107239 2053622
contributor 1167 9415885 2063538 117905
title 2249 21656612 17879231 1149977
identifier 2250 40712655 31833633 1295127
description 2232 21995791 13787533 684706
source 1858 11120103 2626893 126302
date 2232 29303252 3510478 339012
ments). Column values counts the number of distinct
(after normalization) values in these statements that
reveals the degree of repetition of the values. Column
values contains the number of these values that appear
on more than one collection and reveals if the values
have a broader (than inside a collection) scope. As an
example, 16840201 are language statements and con-
tain 15328 distinct values (after normalizing 16059
values), of which 1029 appear on more than one col-
lection.
There are many different values representing the
same quantity, such as a language, and we tried to
reduce them in an unattended, automatic, element-
independent way, by normalizing the values and using
only the normalized values afterwards. Our normal-
ization ignores letter case, punctuation and spacing.
Our data reveal that all elements do have both
many and few repetitions on different collections. The
number of collections with many and few repetitions
is different, and language, type, rights and format are
mostly repeated elements, followed by relation, sub-
ject, coverage and publisher, both inside a collection
and also across collections.
3 INDIVIDUAL LANGUAGE
VALUES
In Fig. 4 we depict the number of collections that
share each dc element value, sorted in decreasing or-
der. We can see similar distribution with extreme val-
ues for all elements. The values for element subject
are higher than the other elements. This can be ex-
plained partly by the fact that there are more subject
statements, and also by adopting more common sub-
ject schemes in the different collections.
Figure 4: The number of collections that share each dc ele-
ment value, for the values that are found in more than one
collection.
Figure 5: The number of collections that share each lan-
guage value, for the 1029 values that are found in more than
one collection (from the 15328 values in total).
To observe the distribution better, Fig. 5 shows the
number of collections, sorted in decreasing order, that
share each language value, for the 1029 values that
are present in more than one of the 2062 collections
(from the 15328 values in total). The fact that only
1029 language values appear on more than one col-
lection indicate that most other values are most prob-
ably ill-formed, as it is very unlikely that so many lan-
guages are actually in use. The language statements
included 15328 distinct values (after normalization).
Compared to other resource properties that are
also often repeated (e.g., keywords or resource for-
mat), the language description depends less on the
type or topic of the resources, its semantics is well
understood by most people and there exist standards
and good practices describing how to denote any lan-
guage. Additionally, the description of the resource
language is mostly repeated: many resources contain
KEOD 2020 - 12th International Conference on Knowledge Engineering and Ontology Development
184
eng 1472
en 563
english 127
en_us 89
ingles 10
e 7
enm 5
united states 5
english old ca 450 1100 3
english language 2
Ingilizce 1
egnlish 1
Figure 6: Some values of the element language denot-
ing English only, with the number of collections they are
present.
more than one language, most resources in a collec-
tion will be in a few languages, and different collec-
tion often have in common some of their languages.
The most common values found are “eng”, “en”
and “spa”, while values like “English”, “en US” and
“en-US” are also common. The actual languages
of the resources are not so many, but the languages
are, unfortunately, mostly specified in many different
ways, which makes automatic processing of such val-
ues much harder.
Some of the issues in the language specifications
are:
No standard or good practice is followed. E.g.,
for specifying English alone, all of the values in
Fig. 6 have been used (only those that are present
on more than one collections are shown). The
most common value, eng, is used 3185233 times
in 1472 collections, but the other values are used
considerably less.
Instead of adopting a bare control value, the
value includes a verbose description. E.g. title
in Arabic at head of title, preface and
afterword in German, text in Yiddish
script, predominately English some
Hebrew and Chinese. These additional details
should normally be included in a description
element.
More than one value is stored inside a single
statement. E.g. text in Yiddish; prayers
in Hebrew with Yiddish translation
and elaboration, primarily in Russian
with some Hebrew, mostly German; some
English and French. Normally, multiple
statement should be used for multiple values.
There are spelling errors in the value. E.g.
Englsih, Englishx. The value content on the
original system should provide a choice of the le-
gal values, to avoid such errors
The value is completely irrelevant, probably in-
serted by mistake. E.g. born digital, 39
x 53 cm, 1946-1951 in english, <--please
select language-->.
The value uses local language or a non standard
character set encoding. E.g. Anglais. There is
no established provision on specifying a character
encoding on specific metadata elements, and most
good practices define values in ASCII (which are
identical in UTF-8). Other encodings should be
avoided for global use.
Similar remarks apply for the values of the other
Dublin Core elements.
Figure 7: The number of language distinct values in each of
the 2062 collections.
Figure 8: The number of language distinct values in each of
the 2062 collections, that are found in two or more collec-
tions.
The collections we used were created indepen-
dently and probably have different description guide-
Consistency and Interoperability on Dublin Core Element Values in Collections Harvested using the Open Archive Initiative Protocol for
Metadata Harvesting
185
lines, but are registered to the OAI-PMH registry, to
allow metadata aggregation and use in an aggregated
context, resembling one big collection. There are no
agreed rules among those collections, but we expect
them to mostly adopt common practices.
Fig. 7 demonstrates how the same values are
found inside a collection and Fig. 8 in different from
our 2062 collections, not taking into account how
many times a value appears in each collection.
4 CLUSTERING LANGUAGE
VALUES
One would expect to see as values for the element lan-
guage either the three-letter language codes (e.g. eng,
spa, jpn, . . . ) or the two-letter language codes (e.g.
en, fr, de, . . . ). Values like these really are the most
common ones.
But other values are also reasonable, such as val-
ues containing a dialect specification (e.g. en US,
en ZA, pt BR, . . . ) or using language names in En-
glish (e.g. English, Korean, German, . . . ), as they
could be automatically converted to more appropriate
values.
To cluster the sets of values used, and reveal their
similarity, we used as a clustering metric the eu-
clidean distance of the value vectors with their fre-
quencies. A collection joins a cluster if its similarity
(inverse distance) to any of the collections in the clus-
ter is above a given threshold.
Thus, if a collections specifies only language eng
and another one specifies only language fre, they
have no similarity. When another collections includes
both languages eng and fre, its similarity to both the
previous collections is established, although its ex-
act value depends on the frequency of each language
value, and the clustering algorithm uses this collec-
tion to bridge the distance between the collections and
to (eventually) form a cluster with the collections with
common language names.
On the other hand, if a collection specifies lan-
guages eng and fre and another collection en and
fr, these collections have no similarity, even though
they are specifying the same real languages, because
they are using different values for them. If all other
collections use either the three letter or the two letter
specification (but no both) - or another (uncommon)
value schema, the above collections will not become
part of the same cluster.
A dendrogram is a diagram representing a tree
performing hierarchical clustering on data and show-
ing the resulting tree. The top of the U-link indicates
a cluster merge. The two legs of the U-link indicate
Figure 9: Dendrogram for language from 2010 collections.
which clusters were merged. The length of the two
legs of the U-link represents the distance between the
child clusters. We use dendrograms to show the possi-
ble division of the collections to sub-clusters and clus-
ters (depending on where we should set the similarity
threshold) and to explore the similarity among them.
In the dendrogram of Fig. 9, we can see the cluster
distances from the 2010 of our 2062 collections, ex-
cluding the 52 tiny ones, with less than 10 language
statements. We observe that we have 3 big clusters,
and we can see the smaller differences inside each
cluster.
When two clusters contain mostly different val-
ues and any common values are incidental, their dis-
tance is high. The high distance of the child clusters
also shows that even though the collections use some
common values (to justify some similarity), they also
contain many different ones, possibly some of them
by mistake or from failing to strictly follow rules or
practices.
5 OTHER REPEATED
ELEMENTS
We also examined the value patterns for other Dublin
Core elements. The clusters for the elements type,
format, rights, coverage and publisher, which are the
elements with the smallest number of distinct val-
ues in more than one collection are also demonstrated
through dendrograms.
In the dendrogram of Fig. 10, we can see the clus-
ter distances from the 2211 of our 2229 collections,
excluding the 18 tiny ones, with less than 10 type
statements. We observe that we have 4 big clusters
that are quite distant apart (heterogeneous), which are
subdivided into further heterogeneous clusters.
In the dendrogram of Fig. 11, we can see the clus-
KEOD 2020 - 12th International Conference on Knowledge Engineering and Ontology Development
186
Figure 10: Dendrogram for type from 2211 collections.
Figure 11: Dendrogram for format from 2013 collections.
Figure 12: Dendrogram for rights from 1873 collections.
ter distances from the 2013 of our 2064 collections,
excluding the 51 tiny ones, with less than 10 format
statements. We observe that we have 2 very different
big clusters, and the cluster to the right has many more
collections splitted in groups - with similar values.
In the dendrogram of Fig. 12, we can see the clus-
ter distances from the 1873 of our 1972 collections,
Figure 13: Dendrogram for coverage from 319 collections.
Figure 14: Dendrogram for publisher from 2154 collec-
tions.
excluding the 99 tiny ones, with less than 10 rights
statements. We observe that we have 4 very different
clusters, and the cluster to the right has 70% of the
collections and half of them are very similar, while
the other half are form many sub-clusters.
The element that is present in the fewest collec-
tions is coverage. In the dendrogram of Fig. 13, we
can see the cluster distances from the 319 of our 403
collections, excluding the 84 tiny ones, with less than
10 from the few collections including coverage state-
ments. We observe that we have 7 very different big
clusters, and two of them include most of the collec-
tions.
The element publisher is commonly found in
many collections, but its values are in many cases
highly repeated but also vendor specific: the collec-
tions declare their own publisher values. In the den-
drogram of Fig. 14, we can see the cluster distances
from the 2154 of our 2204 collections, excluding the
50 tiny ones, with less than 10 publisher statements.
We observe that we have 1 big cluster, with most col-
lections, another cluster with few collections and few
Consistency and Interoperability on Dublin Core Element Values in Collections Harvested using the Open Archive Initiative Protocol for
Metadata Harvesting
187
more clusters with one collection each.
6 CONCLUSIONS AND FUTURE
WORK
A good practice on metadata descriptions consists of
using specific values to identify discrete entities. This
is the same approach that leads to the Linked Open
Data, when specifying entities from other, open, col-
lections. Additionally, quite often these entities are
much fewer than the collection items (e.g. the num-
ber of languages of the items, or their type), leading
to many value repetitions. This occurs more on some
Dublin Core elements (like language and type), than
on some others (like identifier and date). But it does
not occur as often as we would expect: Quite often,
there are many values corresponding to the same en-
tity.
We started by studying the language element.
Even though the language specification information
is handled in standard ways in many computer appli-
cations, on many harvested metadata the Dublin Core
element language lacks any standardization.
We examined the language values found and their
frequencies. In many cases we found illegal or prob-
lematic values and we classified them into categories.
We used dendrograms to show the similarity of the
language values among collections.
Nevertheless, there are more common understand-
ing (and also standards) on what the language entities
are and how to denote them. And still, many items
do not adopt the same values, and provide many “un-
usual” values.
We repeated the procedure for the type, format,
relation, coverage and publisher elements. The sit-
uation on all these elements was similar. We could
observe the value clustering differences by using den-
drograms.
We conclude that we need more standardization
on values, more so on language values, so that state-
ments across collections follow some good practices.
The situation is similar to the other Dublin Core ele-
ments, although not identical.
To collect all possible records, we adapted the har-
vested procedure to handle both reasonable timeouts
and large number or records, using a repeated harvest-
ing procedure for many small timeout intervals.
In the future, we could study unique and low rep-
etition values that are similar to other values, on el-
ements with high usual repetition and also repeated
values on elements with low usual repetition in order
to derive rules and guidelines for automatically creat-
ing value mappings.
REFERENCES
Baca, M. (2003). Practical issues in applying metadata
schemas and controlled vocabularies to cultural her-
itage information. Cataloging & Classification Quar-
terly, 36(3–4):47–55.
Fuhr, N., Tsakonas, G., Aalberg, T., Agosti, M., Hansen,
P., Kapidakis, S., Klas, C.-P., Kov
´
acs, L., Landoni,
M., Micsik, A., Papatheodorou, C., Peters, C., and
Sølvberg, I. (2007). Evaluation of digital libraries. In-
ternational Journal on Digital Libraries, 8(1):21–38.
Harper, C. A. (2016). Metadata analytics, visualization, and
optimization: Experiments in statistical analysis of the
digital public library of america (dpla). Code4Lib, 33.
Harper, C. A. and Tillett, B. B. (2007). Library of congress
controlled vocabularies and their application to the se-
mantic web. Cataloging & Classification Quarterly,
43(3–4):47–68.
Hughes, B. (2005). Metadata quality evaluation: experience
from the open language archives community. Lecture
Notes in Computer Science, 3334.
Kapidakis, S. (2018). Metadata Synthesis and Updates on
Collections Harvested using the Open Archive Initia-
tive Protocol for Metadata Harvesting. 22nd Interna-
tional Conference on Theory and Practice of Digital
Libraries, TPDL 2018, LNCS 10450, Springer, pages
16–31.
Kapidakis, S. (2019). Repeated values on Collections
Harvested using the Open Archive Initiative Protocol
for Metadata Harvesting. 11th International Confer-
ence on Management of Digital EcoSystems, MEDES
2019, November 12–14, 2019, Limassol, Cyprus,
ACM 2019, ISBN 978-1-4503-6238-2.
Kir
´
aly, P., Stiller, J., Charles, V., Bailer, W., and Freire, N.
(2019). Evaluating data quality in europeana: Met-
rics for multilinguality. In Garoufallou, E., Sartori,
F., Siatri, R., and Zervas, M., editors, Metadata and
Semantic Research, pages 199–211, Cham. Springer
International Publishing.
Maltese, V. (2018). Digital transformation challenges for
universities: Ensuring information consistency across
digital services. Cataloging & Classification Quar-
terly, 56(7):592–606.
Moreira, B. L., Gonc¸alves, M. A., Laender, A. H., and Fox,
E. A. (2009). Automatic evaluation of digital libraries
with 5squal. Journal of Informetrics, 3(2):102–123.
Vullo, G., Clavel, G., Ferro, N., Higgins, S., van Horik, R.,
Horstmann, W., and Kapidakis, S. (2010). : Quality
interoperability within digital libraries: the DL. org
perspective. In: 2nd DL. org Workshop in conjunction
with ECDL, 2010:9–10.
Wilkinson, M. D. e. a. (2016). The fair guiding principles
for scientific data management and stewardship. Sci.
Data, 3(160018).
Zhang, Y. (2010). Developing a holistic model for digital
library evaluation. Journal of the American Society for
Information Science and Technology, 61(1):88–110.
KEOD 2020 - 12th International Conference on Knowledge Engineering and Ontology Development
188