Relevancy Scoring for Knowledge-based Recommender Systems

Robert David

and Trineke Kamerling

Semantic Web Company, Austria

Rijksmuseum Amsterdam, The Nederlands

Keywords:

Cultural Heritage, Knowledge Representation, Semantic Web, Information Retrieval, Recommender,

Relevancy.

Abstract:

Knowledge-based recommender systems are well suited for users to explore complex knowledge domains like

iconography without having domain knowledge. To help them understand and make decisions for navigation

in the information space, we can show how important speciﬁc concept annotations are for the description of

an item in a collection. We present an approach to automatically determine relevancy scores for concepts of a

domain model. These scores represent the importance for item descriptions as part of knowledge-based rec-

ommender systems. In this paper we focus on the knowledge domain of iconography, which is quite complex,

difﬁcult to understand and not commonly known. The use case for a knowledge-based recommender system

in this knowledge domain is the exploration of a museum collection of historical artworks. The relevancy

scores for the concepts of an artwork should help the user to understand the iconographic interpretation and to

navigate the collection based on personal interests.

1 INTRODUCTION

Complex knowledge domains like iconography are

difﬁcult to understand. Non-experts, who do not have

domain knowledge, cannot interpret the symbols to

determine the meaning of artwork content. Even for

art historians, the analysis of the meaning of an art-

work is a difﬁcult task for complex scenes. To illus-

trate the problem, we present the motivating scenario

of museum visitors who want to understand the con-

tent of artworks and want to navigate museum collec-

tions based on personal interests. However, visitors

do not have domain knowledge. We call such a per-

son a non-expert. Having no understanding about the

artworks, visitors have to use the limited information

that is provided by the museum. Usually, this is a

description in natural language and provided as text

or as an audio guide. The natural language descrip-

tion makes it difﬁcult for non-experts to understand

the details of the scene interpretation, because generic

concepts, like religion or mythology, are mixed with

speciﬁc concepts, like symbolic objects. However,

this detail information helps to understand an artwork

as well as related artworks. We present an approach

to describe visual artwork content using concepts of

multiple interlinked vocabularies. The concepts are

automatically scored based on relevancy regarding the

artworks description. By looking at the concept an-

notations and the associated relevancy scores, visitors

get a better understanding because the most important

concepts of the content are pointed out to them. As

part of a knowledge-based recommender, these con-

cepts can be selectively used by visitors to navigate

artwork collections.

1.1 Knowledge-based Recommender

Systems

Knowledge-based recommender systems are well

suited for complex knowledge domains (Felfernig and

Burke, 2008). The basis of the recommendations is

a knowledge model representing domain knowledge

that is used for similarity-based retrieval. Therefore

knowledge-based recommender systems are not de-

pendent on data from user decisions like other rec-

ommendation approaches. Users with only little do-

main knowledge can still use the system for recom-

mendations based on their interests. They only need

to have general knowledge about the set of collec-

tion items and an informal knowledge of their needs

(Burke, 2000). Because the options for navigation

are based on the model describing the items, this ap-

proach can be used to help users to better understand

the knowledge domain. When navigating the infor-

David, R. and Kamerling, T.

Relevancy Scor ing for Knowledge-based Recommender Systems.

DOI: 10.5220/0008068602330239

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 233-239

ISBN: 978-989-758-382-7

233

mation space, two navigation decisions can be taken.

First, the user selects from the results of similarity-

based retrieval (Burke et al., 1997), which will recom-

mend items that are similar to the current item based

on the knowledge model. Second, the user can tweak

(Burke et al., 1997) the selection based on informa-

tion needs to alter the resulting recommended items.

Knowledge-based recommender systems help users

explore and understand an information space (Burke,

2000). Users have to make decisions for retrieval and

tweaking based on the model to navigate the informa-

tion space. Although knowledge-based recommender

systems only require general knowledge of the do-

main, users still might not have enough understanding

to decide how to tweak the recommendations to sat-

isfy their information needs. To help users understand

and make these tweaking decisions, we can show how

signiﬁcant speciﬁc parts of the description are for an

item. Knowing the signiﬁcance is also helpful for

understanding the meaning of items in the informa-

tion space. The signiﬁcance is expressed as a rele-

vancy score. We focus on scoring concept annotations

which are actually instances of concepts (Gicquel and

Lenne, 2013) assigned to the items. They represent

the semantic context of an item. Showing the signif-

icance of concepts therein helps the users make deci-

sions for retrieval to satisfy information needs.

1.2 Information Content

The notion of information content is described in

(Resnik, 1995). It uses a hierarchic taxonomy and a

corpus where the taxonomy concepts occur to deter-

mine the informativeness of concepts. The taxonomic

hierarchy represents how general or speciﬁc a con-

cept is and therefore can be used to determine how

informative it is regarding a description. As prob-

ability to encounter a concept in the taxonomy in-

creases, informativeness decreases, so the more ab-

stract a concept, the lower its information content. If

there is a unique top concept, its information content

is 0 (Resnik, 1995). The information content score is

not only based on the taxonomy, but also uses corpus-

based statistical information. The frequency of con-

cepts in a corpus is taken into account and is part of

the calculation of the probability to encounter a con-

cept. If a concept does not occur in the corpus, its in-

formation content is 0. Concepts that do not occur in

the corpus do not contribute to the information con-

tent score. The information content of concepts can

be used to measure the semantic similarity between

two concepts based on the taxonomy and the corpus

(Resnik, 1995).

2 RELEVANCY SCORING BASED

ON INTERLINKING

The calculation of relevancy scores for concepts we

propose is based on the notion of information content.

We adapt the approach to use interlinked taxonomies

instead of an annotated textual corpus to calculate in-

formativeness and use it as a relevancy score. The ba-

sis is the use of two taxonomies to describe the collec-

tion items. These taxonomies must have different se-

mantic expressivity regarding the knowledge domain

they describe to provide an advantage to non-experts.

Higher expressivity means a more concrete descrip-

tion requiring a higher level of domain knowledge to

understand. Lower expressivity means a more com-

mon description that also non-experts understand. We

use the taxonomy with higher expressivity as a basis

to calculate scores for the concepts of the taxonomy

with lower expressivity. That means we interpret the

higher taxonomy as the annotated corpus according

to information content and calculate the scores based

on the concept occurrences. These occurrences are

represented by the links between concepts of the two

taxonomies. The concepts of the lower taxonomy are

used to describe the concepts of the higher taxonomy.

They therefore provide a description of these concepts

that is understood by non-experts. The ﬁrst step of the

method is to provide these taxonomy interlinks. The

second is the calculation of the information content

scores which represent the relevancy.

2.1 Vocabulary Interlinking

Interlinking multiple vocabularies that are used to an-

notate items has several advantages. It automatically

extends descriptions by adding additional linked con-

cepts and it improves search by improving recall.

Using multiple vocabularies, we can describe items

on different vertical semantic levels (Hyvönen et al.,

2007). Users with different views or different knowl-

edge can use different levels for describing the same

item. The levels are represented by vocabularies with

different semantic expressivity or in general a differ-

ent point of view on the domain. Interlinking vocabu-

laries also provides different approaches to data inte-

gration. (Hollink et al., 2003) shows how to improve

semantic search by interlinking concepts of different

ontologies that are used to annotate artworks. This

increases the recall of the search. (Hyvönen et al.,

2007) uses three vertical levels of detail with increas-

ing semantic expressivity for different kinds of con-

tent annotation. It can be used to integrate content

annotated at different levels of granularity. Each new

level of annotation granularity adds new information

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

234

with respect to the previous level. This means that

semantically richer representations can be easily in-

terpreted at the lower level. These examples show the

advantages provided by vocabulary interlinking.

2.2 Relevancy Scoring

The interlinks between concepts of multiple vocabu-

laries represent descriptions in vertical semantic di-

rections. To explain concepts of a higher semantic

level, and therefore items of a collection annotated

with these concepts, we provide relevancy scores for

the lower level concepts based on the notion of in-

formation content. This can only be done for vo-

cabularies with a taxonomic structure because the hi-

erarchy is needed to calculate informativeness. The

scores represent the importance a concept has for the

annotated item in the scope of the interlinked vocab-

ularies. Relevancy scoring can be done automatically

and independent of any collection that the vocabular-

ies are applied to. The following example explains

the method. Let taxonomy A and taxonomy B be

taxonomies used to annotate items. Taxonomy A is

of higher semantic expressivity than taxonomy B. By

using the interlinks we calculate the relevancy of con-

cepts of taxonomy B based on their occurrences in

taxonomy A. If a concept from taxonomy B is not

interlinked to concepts of taxonomy A, its informa-

tion content is 0. Concepts that are not interlinked do

not contribute to the information content score. The

relevancy for a concept of taxonomy B is the infor-

mation content score based on the hierarchy of taxon-

omy B and the number of interlinks with concepts of

taxonomy A. This means we interpret taxonomy A as

a corpus for taxonomy B and the frequency of inter-

links is taken into account to calculate the probability

to encounter a concept. In ﬁgure 1, we see several

interlinks going from concepts of taxonomy B to con-

Figure 1: Taxonomy interlinking.

cepts of taxonomy A. Concept b2 has two links (a3,

a4), b3 has one link (a3) and b4 also has one link (a2).

The total number of interlinks is 4. We calculate the

information content scores (ic) for concepts of taxon-

omy B. The information content score for a concept

of taxonomy B is the negative log of the likelihood to

encounter this concept, including the child concepts.

ic(b1) = − log(

) = 0

ic(b2) = − log(

) = 0.125

ic(b3) = − log

= 0.6

ic(b4) = − log

= 0.6

The information content scores represent the rele-

vancy that each concept of taxonomy B has for de-

scribing concepts of taxonomy A. The scoring is done

automatically and is not dependent on the item collec-

tion the two taxonomies describe.

3 ART HISTORIC

IMPLEMENTATION

3.1 Vocabularies for Art Description

In this paper we focus on semantic artwork descrip-

tions to be used by a knowledge-based recommender

system for understanding and exploring an artwork

collection. We select two vocabularies for doing

the descriptions. Both are organized as hierarchic

taxonomies. They are available as Semantic Web

(Berners-Lee et al., 2001) data models using the Sim-

ple Knowledge Organization System (SKOS) (Alis-

tair Miles, 2009). Both vocabularies were designed to

describe artworks, so it makes sense to apply them

both to artwork items for a semantic description.

They represent information at different vertical levels

of semantic expressivity, so they qualify for interlink-

ing and relevancy score calculation.

Iconclass. Iconclass (Van de Waal, 1974) is a clas-

siﬁcation system for iconography. It is structured

as a hierarchy of concepts representing themes of

the iconographic tradition of art history. Therefore

it is mainly intended for historical artworks of Eu-

ropean origin. The concepts within Iconclass in-

clude common artwork themes throughout history

and mythology. It is not intended to describe single

parts of an artwork, but contains concepts that can be

used to describe the scene as a whole. Historic and

Relevancy Scoring for Knowledge-based Recommender Systems

235

mythological persons and stories are included. Even

single scenes from stories are represented by con-

cepts in Iconclass. For example, the concept http:

//iconclass.org/71U4273, titled "Holofernes be-

headed by Judith with his own sword; the maidser-

vant may be keeping watch", represents the scene of

the Old Testament where Judith beheads Holofernes.

This example shows the high level of detail the con-

cepts of Iconclass have. With just one concept the

whole content of an artwork can be described. Icon-

class is intended to be used by domain experts to iden-

tify and annotate iconographic scenes for historical

artworks.

AAT. The Getty Art and Architecture Thesaurus

(Petersen, 1990) includes a broad range of concepts

that can be used to describe art and architectural

works. It has many concepts for simple items and

physical objects, concepts for materials that can be

used to describe the physical characteristics of works,

and also concepts for immaterial things like emotions

and seasons. AAT concepts can be used to describe

the elements shown by an artwork. Therefore AAT

can be used for the same purpose as Iconclass, al-

though on a different semantic level because it does

not contain iconographic concepts. Because both vo-

cabularies have the same purpose, but different views

on artworks, they are suitable for the use case evalua-

tion.

3.2 Relevancy Scoring for Art History

Iconclass, as an expert taxonomy of the art histori-

cal domain, is difﬁcult to understand for non-experts.

In contrast, the AAT represents a more general tax-

onomy and is therefore easier to understand for non-

experts. To provide a better understanding for the

Iconlass concepts, we will relate them to associated

AAT concepts. This is done by ﬁrst interlinking the

AAT concepts to Iconclass concepts and then calcu-

lating the information content scores for the AAT con-

cepts based on the interlinks. These scores show the

relevancy of the AAT concepts for iconographic de-

scriptions in the context of the two vocabularies.

Interlinking Iconclass and AAT. For interlinking

Iconclass and AAT concepts we have to ﬁnd a way to

describe the Iconclass concepts using the AAT con-

cepts. The hierarchic structure cannot be used for this

because the two taxonomies, although intended for

a similar purpose, represent different knowledge do-

mains. An alignment based on the structure is there-

fore not possible. However, both taxonomies provide

labels as textual descriptions of the concepts that can

be used as a basis for matching. For Iconclass con-

cepts, besides the labels, there is a context description

for each concept that is done using a controlled vocab-

ulary. Although this vocabulary does not have a struc-

ture, there are no ambiguous terms in it. It also repre-

sents a more general description of a concept which is

something that is similar to what AAT provides. For

the AAT, we also use the provided labels as a basis

for interlinking. We match the labels of the AAT con-

cepts to the subject terms of the Iconclass concepts

to create an initial mapping. This is done by using a

regex-based approach that removes terms in brackets

and then directly matches the results. Although both

labels and subjects are not ambiguous within each vo-

cabulary, the mapping result contains many ambigu-

ities. To disambiguate, we take the taxonomic hier-

archy of AAT and try to resolve the match based on

the labels of the path to the taxonomic root. If the

subjects of an Iconclass concept are found among the

labels of the AAT concepts on the path to the root, we

identify this AAT concept as a correct match. After

resolving the ambiguities using this method, we have

an automatically generated interlink set for Iconclass

and AAT.

Relevancy Scoring. The relevancy scoring is done

for AAT concepts based on the result of the interlink-

ing. We compute the information content values as

described above. The result are the relevancy scores

for all interlinked AAT concepts that represent how

signiﬁcant a concept is for the description of art his-

torical items.

4 USE CASE EVALUATION

For evaluating the approach we apply it to an art his-

torical collection of items. Evaluation is done regard-

ing the correctness of the interlinking and scoring, as

well as the art historical correctness of the results. The

higher scored concepts are more signiﬁcant than the

lower scored concepts regarding an item description.

That means that the order of the concept annotations

of an item, created by the relevancy scores, has to be

veriﬁed for correctness from an art historical point

of view. The evaluation is presented in three parts.

First, we describe the collections we used as a basis

for the evaluation. Second, we show three example

items from the collections to illustrate the approach.

We discuss the correctness of the interlinking and the

concept scores of these items. We then show how the

concepts relate the different items with each other and

how the relevancy scores represent the signiﬁcance of

these relations. Third, we describe how we evaluated

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

236

the approach on a larger scale by creating sample sets

and performing an art historical analysis of the scored

concept annotations.

4.1 Artwork Collection

For the collection of artworks used in the evaluation

we decided to use the Linked Data representation (Di-

jkshoorn et al., 2014) of the Rijksmuseum collections,

which is available as Linked Open Data. This dataset

contains over half a million objects, including detailed

descriptions and high-quality images released under a

public domain license. The museum uses the Icon-

class vocabulary to describe subject matter. We chose

two collections from the Rijksmuseum dataset which

contain visual artworks. The ﬁrst collection contains

paintings, while the second collection contains pot-

tery. We evaluate the approach by ﬁrst concluding

AAT concepts based on the asserted Iconclass con-

cepts for an item and then verifying the interlinks and

the relevancy scores regarding the iconographic inter-

pretation.

4.2 3 Examples

We present 3 examples of artworks that show differ-

ent scenes. The Iconclass annotations were added by

the Rijksmuseum and are asserted in the published

dataset. Most of the annotations of the three artworks

are located in different parts of the Iconclass taxon-

omy. There is not much relation between them re-

garding the iconographic description. However, by

expanding to AAT concepts based on interlinking, we

can conclude shared concepts among them. We show

how the artworks relate and how signiﬁcant the rela-

tion between them is based on the relevancy scores.

Example 1: Time (http://hdl.handle.net/

10934/RM0001.COLLECT.7192) The ﬁrst artwork

shows time as an old man, naked and with wings on

his back. He is kneeling on his right knee and holds

up a cloth with his right hand. On the ﬂoor lies a

golden scale between leaves. Time, being an abstract

concept, is represented by the physical form of a per-

son. Such symbolic representations are called alle-

gories. The item is annotated with one Iconclass con-

cept which describes the scene, as shown in table 1.

Table 1: Time.

Iconclass Concept Title

http://iconclass.org/23A1 Father Time, man with wings and scythe

Example 2: Autumn (Conversation) (http://

hdl.handle.net/10934/RM0001.COLLECT.6396)

The second artwork shows a landscape in the season

of autumn. We can see a road on the banks of a water

with some ﬁshing boats. On the road there is a man

who speaks with a woman. The artwork is part of a

series about the four seasons. The item is annotated

with three Iconclass concepts which describe the

scene, as shown in table 2.

Table 2: Autumn (Conversation).

Iconclass Concept Title

http://iconclass.org/46C22 harbouring

http://iconclass.org/23F44

autumn landscape; landscape symbolizing autumn

(the four seasons of the year)

http://iconclass.org/33A35 conversation, dialogue; conversation piece

Example 3: Still Life with Fruit and Oys-

ters (http://hdl.handle.net/10934/RM0001.

COLLECT.9119) The third artwork shows a still life

with fruits, oysters and a watch. The item is anno-

tated with four Iconclass concepts which describe the

scene, as shown in table 3.

Table 3: Still Life with Fruit and Oysters.

Iconclass Concept Title

http://iconclass.org/23U25 watch

http://iconclass.org/25F72(OYSTER) molluscs: oyster

http://iconclass.org/41C653 fruit

http://iconclass.org/41C38 laid table as still life

Interlinks and Relevancy Scores. Using the inter-

links we conclude the following AAT concepts with

relevancy scores for the items shown in table 4, ta-

ble 5 and table 6. These concepts do not only repre-

sent what scenes are shown on the artworks, but also

what the meaning of these scenes is. They represent

iconographic interpretations using general concepts

that non-experts understand more easily compared to

the iconographic concepts of Iconclass. When we

Table 4: Time.

AAT Concept Title Score

http://vocab.getty.edu/aat/300024371 scythes 10.3707

http://vocab.getty.edu/aat/300375053 wings (animal components) 8.9653

http://vocab.getty.edu/aat/300133089 time 6.8398

http://vocab.getty.edu/aat/300202507 allegories (document genre) 5.3514

http://vocab.getty.edu/aat/300055866 allegory (artistic device) 5.3514

http://vocab.getty.edu/aat/300179372 Nature 5.1355

look at the concluded AAT concepts, we can see some

of them are incorrect for the description of the art-

work. These false positives were produced because

the disambiguation did not have sufﬁcient context in-

formation in these cases. We show them in the ta-

bles marked in red. They are incorrect and therefore

will not be considered any further. Then we look at

the relevancy scores and the order of concepts that

Relevancy Scoring for Knowledge-based Recommender Systems

237

Table 5: Autumn (Conversation).

AAT Concept Title Score

http://vocab.getty.edu/aat/300178227 conversation pieces (portraits) 12.2425

http://vocab.getty.edu/aat/300026185 dialogues 10.8562

http://vocab.getty.edu/aat/300133093 autumn 10.7384

http://vocab.getty.edu/aat/300008678 harbors 10.1631

http://vocab.getty.edu/aat/300018959 Company (style) 9.7576

http://vocab.getty.edu/aat/300160084 companies 9.2221

http://vocab.getty.edu/aat/300133091 seasons 7.2487

http://vocab.getty.edu/aat/300055250 trafﬁc 6.8398

http://vocab.getty.edu/aat/300133089 time 6.8398

http://vocab.getty.edu/aat/300082981 ships 6.6367

http://vocab.getty.edu/aat/300053892 transporting 6.2485

http://vocab.getty.edu/aat/300179372 Nature 5.1355

http://vocab.getty.edu/aat/300008626 landscapes (environments) 4.7171

http://vocab.getty.edu/aat/300265711 Homo sapiens (species) 4.3210

http://vocab.getty.edu/aat/300055806 civilization 4.1848

http://vocab.getty.edu/aat/300026009 societies 4.1838

http://vocab.getty.edu/aat/300055768 culture 4.1259

Table 6: Still Life with Fruit and Oysters.

AAT Concept Title Score

http://vocab.getty.edu/aat/300041615 watches 12.9356

http://vocab.getty.edu/aat/300310174 oysters 11.5493

http://vocab.getty.edu/aat/300380162 vegetables 9.9399

http://vocab.getty.edu/aat/300265736 tables (architectural elements) 9.8446

http://vocab.getty.edu/aat/300027364 tables (documents) 9.8446

http://vocab.getty.edu/aat/300216559 bellies 9.6034

http://vocab.getty.edu/aat/300039548 tables (support furniture) 9.1290

http://vocab.getty.edu/aat/300053578 measuring 8.4813

http://vocab.getty.edu/aat/300266510 Mollusca (phylum) 8.1650

http://vocab.getty.edu/aat/300024838 instruments 8.0681

http://vocab.getty.edu/aat/300015638 still lifes 7.3983

http://vocab.getty.edu/aat/300254496 food 7.0114

http://vocab.getty.edu/aat/300011868 fruit 6.8789

http://vocab.getty.edu/aat/300133089 time 6.8398

http://vocab.getty.edu/aat/300011734 earth (soil) 5.2898

http://vocab.getty.edu/aat/300311363 earth (color) 5.2898

http://vocab.getty.edu/aat/300179372 Nature 5.1355

http://vocab.getty.edu/aat/300055806 civilization 4.1848

http://vocab.getty.edu/aat/300026009 societies 4.1838

http://vocab.getty.edu/aat/300055768 culture 4.1259

http://vocab.getty.edu/aat/300249395 Animalia (kingdom) 3.0796

they create for the artwork description. We can see

that the order is valid regarding the depicted scenes

and the iconographic interpretations. It goes from the

more speciﬁc concepts to the more generic ones and

provides additional detail information to non-experts

compared to the Iconclass concepts. We also see that

there are concepts added that have thematically to do

with the artwork, but are not actually shown. An ex-

ample is the concept for scythes for the artwork Time

(table 1), which is the top scoring concept (table 4).

This is an often used symbol of the allegorical rep-

resentation of time as an old man, although it is not

shown on this speciﬁc artwork example. Still we

would want to use this concept when retrieving re-

lated artworks.

Relations between the Artworks. We showed

three examples of artworks with their interlinked and

scored AAT concepts. We can see the signiﬁcant con-

cepts as a description of what the artwork scene rep-

resents. Now we can compare the three artworks to

see how they are related. This comparison can be

used by a knowledge-based recommender system to

retrieve navigation options. The three artworks show

different scenes and have different iconographic in-

terpretations. Still they have some concepts in com-

mon. These concepts, time and nature, are shown

in table 7. The ﬁrst artwork, Time, is an anthropo-

morphic representation of time. This is an art his-

torical interpretation that requires domain knowledge.

The second artwork, Autumn (Conversation), shows

a scene in autumn. Although not clearly visible from

the content, the title of the artwork indicates this to

the viewer. Autumn, as a season, is one of the four

time periods in a year. The third artwork, Still Life

with Fruit and Oysters, is a still life, which is a genre

of artworks showing mostly inanimate subject mat-

ter. Still lives often have symbolic elements. This

artwork depicts a watch as part of the scene, which is

a symbolic representation of the brevity of life. We

can see that the three artworks have the concept of

time in common. Also, they have the concept of na-

ture in common, although this is less signiﬁcant. Ta-

ble 7 shows the scores for both concepts. We can see

Table 7: Shared concepts.

AAT Concept ID Title Score

http://vocab.getty.edu/aat/300133089 time 6.8398

http://vocab.getty.edu/aat/300179372 Nature 5.1355

that the concept of time is in the upper score range of

the scored concepts, while nature has a lower score.

The relevancy scores of these concepts represent how

strongly the three artworks are related. There is a rela-

tion, although not very obvious. This is also reﬂected

in the scores. It is difﬁcult for a non-expert to under-

stand that the concept of time is part of the meaning

of the three scenes. The ﬁrst artwork has time in the

title, while the second artwork mentions autumn, and

the third artwork depicts a watch. By providing the

interlinked and scored AAT concepts, we can show

these shared concepts and the signiﬁcance regarding

the three artwork scenes. The three examples illus-

trate how the scored concepts represent a detailed de-

scription of artwork content. The concept order based

on the scoring shows how signiﬁcant each concept is

for the description of the artwork content. By com-

paring the concepts and the relevancy scores, we can

show relations between artworks as well as how sig-

niﬁcant these relations are.

4.3 Art Historical Evaluation

The evaluation was done on samples from the selected

two collections of the Rijksmuseum dataset. We gen-

erated 3 sets of item samples from these collections

containing 100 items each. As part of this research a

prototypical web application was developed to calcu-

late and retrieve scored AAT concepts for the items.

KEOD 2019 - 11th International Conference on Knowledge Engineering and Ontology Development

238

The resulting item information for the sample sets was

published as a report. This report was then evaluated

by experts of the art historical domain. The evalua-

tion had to verify the correctness of the interlinks as

well as the correctness of the order of the concept an-

notations. As already seen in the presented examples,

the interlinks contain incorrectly disambiguated con-

cepts. This not only created incorrect concept anno-

tations, but also inﬂuenced the relevancy scores. We

ignored the incorrect concepts from the artwork de-

scriptions and evaluated the remaining correct con-

cept annotations regarding the correctness of the or-

der. It showed that the correct concept annotations

provide a valid description of the details of the content

regarding the relevance. The concluded concepts au-

tomatically extend the asserted descriptions and help

non-experts to understand the details using the rele-

vancy scoring. They can be used as a basis for re-

trieval in knowledge-based recommender systems.

5 CONCLUSION

In this paper we presented an approach for relevancy

scoring based on the interlinking of vocabularies. It

is not dependent on instance data and does not have

drawbacks regarding the size of the collection. It

will work for small collections well as for large ones.

However, it is dependent on the vocabularies selected

for interlinking and the quality of the interlinks. To

help users to better understand the meaning of art-

works, the two vocabularies must have different se-

mantic expressivity, so the concluded information is

represented using a view that is understandable to

non-experts. The data model can be generated au-

tomatically by vocabulary interlinking and calculat-

ing information content scores for the interlinked con-

cepts. We performed a interdisciplinary evaluation to

verify the approach. It showed that the interlinking re-

sulted in false positive concept annotations that were

incorrectly disambiguated. We ignored these false

positives for the further evaluation and only consid-

ered the correct ones for relevancy scoring. It showed

that the order of the correct concept annotations for

the sample artworks was valid. However, the in-

correct interlinks inﬂuence the relevancy scores and

might result in a changed order for some artwork de-

scriptions regarding correct concepts. This, as well

as improving the disambiguation, has to be evaluated

in future work. The next step is the extension of the

implemented prototype for recommendations where

the options are given by the scored concept annota-

tions. We can then evaluate how these artwork de-

scriptions help visitors to navigate the museum col-

lections based on their interests by using the scored

concepts for retrieval and as tweaking options.

REFERENCES

Alistair Miles, S. B. (2009). Skos simple knowledge organi-

zation system reference. http://www.w3.org/TR/skos-

reference/. [Online; accessed 31-March-2015].

Berners-Lee, T., Hendler, J., Lassila, O., et al. (2001). The

semantic web. Scientiﬁc american, 284(5):28–37.

Burke, R. (2000). Knowledge-based recommender systems.

In Encyclopedia of Library and Information Systems,

page 2000. Marcel Dekker.

Burke, R. D., Hammond, K. J., and Young, B. C. (1997).

The ﬁndme approach to assisted browsing. IEEE

Expert: Intelligent Systems and Their Applications,

12:32–40.

Dijkshoorn, C., Jongma, L., Aroyo, L., van Ossenbruggen,

J., Schreiber, G., ter Weele, W., and Wielemaker, J.

(2014). The rijksmuseum collection as linked data.

Semantic Web, (Preprint):1–10.

Felfernig, A. and Burke, R. (2008). Constraint-based rec-

ommender systems: Technologies and research is-

sues. In Proceedings of the 10th International Confer-

ence on Electronic Commerce, ICEC ’08, pages 3:1–

3:10, New York, NY, USA. ACM.

Gicquel, P.-Y. and Lenne, D. (2013). Semantic and Con-

textual Proximities for Informal Learning: The Case

Study of Museum Visits. International Journal of

Computer Applications, 69(26):12134–8410.

Hollink, L., Schreiber, G., Wielemaker, J., and Wielinga, B.

(2003). Semantic annotation of image collections. In

Second International Conference for Knowledge Cap-

ture (KCAP2003) Workshop on Knowledge Markup

and Semantic Annotation.

Hyvönen, E., Ruotsalo, T., Haggström, T., Salminen,

M., Junnila, M., Virkkilä, M., Haaramo, M.,

Mäkelä, E., Kauppinen, T., and Viljanen, K. (2007).

CultureSampo–Finnish Culture on the Semantic Web:

The Vision and First Results, pages 33–58. LIT Ver-

lag, Berlin.

Petersen, T. (1990). Art & architecture thesaurus. Ox-

ford University Press. http://www.getty.edu/research/

tools/vocabulary/aat/.

Resnik, P. (1995). Using information content to evaluate se-

mantic similarity in a taxonomy. In Proceedings of the

14th International Joint Conference on Artiﬁcial In-

telligence - Volume 1, IJCAI’95, pages 448–453, San

Francisco, CA, USA. Morgan Kaufmann Publishers

Inc.

Van de Waal, H. (1974). Iconclass: an iconographic classi-

ﬁcation system, volume 2. North-Holland.

Relevancy Scoring for Knowledge-based Recommender Systems

239