Feature-based Ontology Mapping from an Information

Receivers’ Viewpoint

Fumiko Kano Glückstad

and Morten Mørup

Copenhagen Business School, Dept. of International Business Communication,

Dalgas Have 15, DK-2000 Frederiksberg, Denmark

Technical University of Denmark, DTU Informatics, Section for Cognitive Systems,

Richard Pedersens Plads, DK-2800 Kgs Lyngby, Denmark

Abstract. This paper compares four algorithms for computing feature-based

similarities between concepts respectively possessing a distinctive set of

features. The eventual purpose of comparing these feature-based similarity

algorithms is to identify a candidate term in a Target Language (TL) that can

optimally convey the original meaning of a culturally-specific Source Language

(SL) concept to a TL audience by aligning two culturally-dependent domain-

specific ontologies. The results indicate that the Bayesian Model of

Generalization [1] performs best, not only for identifying candidate translation

terms, but also for computing probabilities that an information receiver

successfully infers the meaning of an SL concept from a given TL translation.

1 Introduction

Accelerated by the recent internet revolution with its fast-paced globalization, cross-

cultural communication, e.g. between an Asian and a European, becomes inherently

challenging due to the lack of sufficient linguistic resources directly bridging remote

languages. This challenge is not only caused by the lack of linguistic resources, but

also by differences in human perception of similar concepts existing in diverse socio-

cultural communities. The MONNET on Multilingual Ontologies for Networked

Knowledge project [2] and the KYOTO project on Knowledge-Yielding Ontologies

for Transition-based Organization [3] are some typical major projects that deal with

such multilingual issues based on ontological methodologies. The approaches taken in

these major research projects are thoroughly analyzed in [4] based on three

dimensions: international (standardized) vs. culturally-influenced domains; functional

(conceptual) vs. documental (lexical) localization; and finally interoperable vs.

independent ontology. The work presented here challenges this multilingual issue by

mapping independent ontologies from a culturally-influenced domain in a functional

manner. The work is part of an overall framework for investigating how background

knowledge possessed by a Source Language (SL) communicator and a Target

Language (TL) reader should be represented and linked in light of various cognitive

processes involved in cross-cultural communication. Background knowledge is

considered as the average domain knowledge possessed by average citizens in a

Kano Glückstad F. and Mørup M..

Feature-based Ontology Mapping from an Information Receivers’ Viewpoint.

DOI: 10.5220/0004088200340043

In Proceedings of the 9th International Workshop on Natural Language Processing and Cognitive Science (NLPCS-2012), pages 34-43

ISBN: 978-989-8565-16-7

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

specific socio-cultural community and assumed to be represented as domain ontology.

We employ a knowledge representation method known as Terminological Ontology

(TO) [5] by constructing two culturally-dependent TOs respectively representing the

Danish- and the German educational systems. A specific purpose is to identify the

most optimal algorithm of mapping culturally-influenced domain knowledge existing

in two cultures using taxonomically organized hierarchical feature-structures obtained

from these TOs. A candidate algorithm is the so-called Bayesian Model of

Generalization (BMG) [1], a novel cognitive model that considers the hierarchical

feature-structure as prior knowledge of an SL communicator or a TL audience,

depending on the assignment of variables to be explained in next section. More

specifically, the BMG computes asymmetric (uni-directional) similarities based on

feature values either from an SL communicator- or a TL audience’s viewpoint by

considering the prior knowledge as cultural bias. The asymmetric coordination in

communication is also well illustrated in the Relevance Theory of Communication

[6]. Accordingly, the BMG is compared against Tversky’s set-theoretic model [7] that

has previously been tested in [8].

In Section 2, the similarity measures applied here are further explained in detail.

Section 3 describes an experiment applying four different feature-based similarity

measures to data-sets obtained from the TOs, respectively representing concepts in

the educational systems in Denmark and Germany. Section 4 discusses the analysis of

the results followed by conclusions in Section 5.

2 Feature-based Similarity Algorithms

The first three similarity measures are derived from Tversky´s ratio model. This

model is defined as follows [7]:



(

,

)

=1/[1+

∗



(

−

)

+∗



(

−

)



(

∩

)

]

(1)

In equation (1), X and Y, respectively, represent feature sets of objects x and y, and 

is considered as additive function.(Y⋂X) represents common features present in both

Y and X, (Y-X) denotes distinctive features existing in Y but not in X, and (X-Y)

denotes distinctive features in X but not in Y. α and β are free parameters which

enables one to compute an asymmetric similarity relationship between object x and y.

Accordingly, three combinations of parameter values are assigned in the previous

study [8]: A) α=1 and

=1: which corresponds to the Jaccard Similarity Coefficient

[9] that represents a symmetric similarity relationship between object x and y; B) α=1

and

=0: which only computes distinctive features present in Y, not in X; and C) α=0

and

=1: which only computes distinctive features present in X, not in Y.

As briefly stated above, a key point is to clarify which variable is defined as a

concept in an SL- or a TL culture. According to Tversky [7], if sim(y,x) is interpreted

as the degree to which y is similar to x, then y is the subject to the comparison and x

is the referent. This definition should be applied to all three parameter settings

defined in here. Keeping this definition in mind, an additional key point is that

Tenenbaum & Griffiths [1] demonstrate that Tversky’s model C) is formally

corresponding to the following equation which forms the basis of the BMG explained

below. Equation (2) computes the conditional probability that y falls under C

(Consequential region) given the observation of the example x [1]. The consequential

region C in our work indicates the categorical region where a subject y belongs.



(

∈



)

=1/[1+

∑

(ℎ,)

:∈,∉

∑

(ℎ,)

:,∈

]

(2)

In equation (2), a hypothesized subset h is defined as the region where a concept

belongs to h, if and only if, it possesses feature k [1]. Thus, P(h, x) = P(x|h)P(h) in

equation (2) represents the weight assigned to the consequential subset h in terms of

the example x. Accordingly, the BMG - algorithm D) - is considered as a model

where the weight P(h, x) is - based on the strong sampling scheme defined in [1] -

specifically assigned to Tversky’s model C). The weight is defined as follows [1]:



(





ℎ)=

|ℎ|



∈ℎ

0ℎ

(3)

Here, |h| indicates the size of the region h [1]. In our work, the number of objects

possessing the k

feature in the referent ontology is considered as the size of the

region h. [1] explains that the prior P(h) is not constrained in their analysis so that it

can accommodate arbitrary flexibility across contexts. Hence in this work, we set

P(h) = 1. In the following experiments the BMG is compared against the three

parameter settings defined for Tversky’s Ratio model.

3 Experiment

3.1 Data Preparation

Data Source: The data-sets used in [8] have been used as original data sources. They

are based on document corpora obtainable from the Eurydice web-site published by

the Education, Audiovisual and Culture Executive Agency under the EU commission.

These documents describe the German- and Danish educational systems both in

English and in the original languages based on the ISCED classification. Hence, it is

feasible to identify terminological expressions in the original language from these

documents and eventually identify translation equivalences linking between German

and Danish. Hence, language-dependent terms and their definitions describing the

educational systems in the two cultures have manually been extracted from the

respective English corpora for developing TOs. The reason that these non-remote

European languages are employed in this work is that these documents are written in

accordance with the standardized template defined by the Eurydice, which may better

provide for a well controlled experiment for assessing the similarity measures.

One of the key principles for developing the TOs is that a concept automatically

inherits all feature specifications of its super-ordinate concepts [5]. A dimension of a

concept is an attribute occurring in a non-inherited feature specification of one or

more of its sub-ordinate concepts. The values of the dimension allow a distinction

among sub-concepts of the concept in question. For example, a dimension of the

concept “pre-primary education” is [AGE] whose values are [0-3 | 3-6]. These

dimension values distinguish the sub-concepts: “nursery” and “kindergarten”. The

dimension can only occur on sister concepts and a given value can only appear on one

of these sister concepts. In this way, a concept must be distinguished from each of its

nearest super-ordinate concepts as well as from each of its sister concepts by at least

one feature specification [5]. These principles enable us to generate well-structured

feature sets that are assumed to be useful for the feature-based similarity

computations. Tables 1 and 2 show examples of the expressed feature structures.

Table 1. Example of german data source (terms and feature sets).

ID Term Feature-values

G2 preschool education {ISCED97, children & young, ISCED0}

G5 kindergärten {ISCED97, children & young, ISCED0, child welfare, 3-6y.o.}

G7 schulkindergärten & vorklassen {ISCED97, children & young, ISCED0, preparation}

G10 primary education {ISCED97, children & young, ISCED1}

G11 primary school {ISCED97, children & young, ISCED1, <6-10y.o.<}

G13 secondary education { ISCED97, children & young, ISCED2+3}

G14 lower secondary level {ISCED97, children & young, ISCED2+3, <10-16y.o.<}

G15 school offering one single course {ISCED97, children & young, ISCED2+3, <10-16y.o.< , single}

G16 hauptschule {ISCED97, children & young, ISCED2+3, <10-16y.o.< , single , general basic, 5-9

grade}

G18 gymnasium {ISCED97, children & young, ISCED2+3, <10-16y.o.< , single, intensified, 5-12/13

grade}

G19 schools offering several courses {ISCED97, children & young, ISCED2+3, <10-16y.o.< , several}

Table 2. Example of danish data source (terms and feature sets).

ID Term Feature-values

D2 pre primary {ISCED97, children & young, ISCED0}

D4 kindergarten {ISCED97, children & young, ISCED0, 3-6y.o.}

D6 single structure {ISCED97, children & young, ISCED1+2}

D7 alternative structure {ISCED97, children & young, ISCED1+2, alternative}

D8 home tuition { ISCED97, children & young, ISCED1+2, alternative, compulsory, 6-16y.o}

D9 efterskole or youth school {ISCED97, children & young, ISCED1+2, alternative, compulsory, <14-18y.o.<}

D10 efterskole {ISCED97, children & young, ISCED1+2, alternative, compulsory, <14-18y.o.<, boarding school, approved

by state}

D11 youth school {ISCED97, children & young, ISCED1+2, alternative, compulsory, <14-18y.o.<, day-to-day, public

municipal council}

D14 municipal school {ISCED97, children & young, ISCED1+2, formal teaching, municipality}

D16 0-9

form {ISCED97, children & young, ISCED1+2, compulsory}

D17 0

form {ISCED97, children & young, ISCED1+2, compulsory, preparation}

D18 1-9

form {ISCED97, children & young, ISCED1+2, compulsory, general basic}

D19 10

form {ISCED97, children & young, ISCED1+2, optional}

Creation of Feature-term Matrices: In order to compute similarities, matrices

referring to the German- and Danish educational systems which, respectively, consist

of 58 and 52 terms are manually generated from the feature sets. Feature value

columns are defined in the following way:

1. All feature values existing in the Danish and German data sources are registered in

both matrices.

2. If feature values in the Danish and German matrices are completely overlapping

(e.g. “ISCED0-pre-primary” in DK and “ISCED0-pre-primary” in GE), the feature

columns in question should be merged into one column.

3. If a feature is possessed by a term, the numeric value should be “1”, otherwise “0”

in the matrices.

4. If a feature value in one matrix is completely included in a feature value in the

other matrix (e.g. “ISCED1+2” in DK and “ISCED1” in GE), a term possessing the

feature that includes the other feature (e.g. Danish “ISCED1+2”) should have numeric

value “1” in both feature columns (e.g. “ISCED1+2” in DK and “ISCED1” in GE). It

means that a term possessing a feature value that is included in the other feature (e.g.

German “ISCED1”) should have numeric value “1” only in the feature column in

question.

5. If feature values in the Danish and German matrices are partly overlapping (e.g.

“ISCED1+2” in DK and “ISCED2+3” in GE), a dummy column referring to the exact

overlapping feature value (e.g. “ISCED2” for both DK and GE) is created. In this

example, a Danish term possessing a feature “ISCED 1+2” should have numeric value

“1” in both “ISCED 1+2” and “ISCED2” columns, but not in the “ISCED2+3”

column.

In this way, we create the German matrix consisting of 58 terms x 117 feature values

and the Danish matrix consisting of 52 terms x 117 feature values.

3.2 Similarity Computation

The basic idea of similarity computation here is to identify a translation candidate

from concepts existing in a TL culture. Assuming that SL communicators and TL

information receivers have general conceptualization of culturally-dependent domains

- in this case the educational system in each country - all combinations of similarities

between TL- and SL terms are computed. When computing similarities based on the

three settings of Tversky’s model and the BMG described in Section 2, the variables:

“terms subject to comparison” and “referent terms” are consistently defined across the

four feature-based similarity algorithms: A) Tversky: α=1 and

=0 (Jaccard); B)

Tversky: α=1 and

=0; C) Tversky: α=0 and

=1; and D) the BMG.

3.3 Results

Figure 1 shows the most typical patterns of similarity scores obtained from the

aforementioned four algorithms from top to bottom: A), B), C), and D). Figure 1-a

indicates that algorithms A) and B) are relatively identical, showing that rather

general German terms, such as G2 and G46, score higher similarities. On the other

hand, algorithms C) and D) show that the terms from G2 to G9, all of which are

within the category of preschool education in Germany, score the highest. Especially,

the BMG clearly identifies the series of German preschool educations, all of which

are targeted for children under the school age categorized as ISCED0. Since - in the

simplified formulae of the BMG - the sum of distinctive features possessed by

referent (variable: x) but not subject to comparison (variable: y) and common features

possessed by both x and y become denominator, the eventual score results in the value

“1”. From a communicator’s viewpoint, it is reasonable to consider that, based on

prior knowledge of Danish SL concepts, all the German TL terms that possess feature

ISCED0 targeted for children under the school age are categorized as objects

belonging to D2.

In Figure 1-b, algorithm B) identifies general German terms such as G1, G2 and

G46 as the most similar terms to D4: kindergarten. On the other hand, all other

algorithms indicate that term G5: “kindergärten” has the highest similarity in terms of

D4. Especially, the BMG clearly points out this implication selecting G5 as the most

similar concept to D4, because the size principle weights the feature value, “3-6 years

old”, which is possessed only by D4 and G5, heavier than other features.

Fig. 1. Similarity scores: German as variant, Danish as referent.

Figure 1-c for the algorithms C) and D) show that the series of terms referring to

the German primary- and secondary education have slightly higher similarity scores

than other terms. However, the scores in the BMG are particularly low. When

inspecting the feature values of D14, it becomes clear that D14 contains two

distinctive features (“formal teaching” and “municipality”) that are not possessed by

any German terms. In addition, the fully- and partly overlapping common features are

possessed by many terms in both German and Danish, which result in assigning lower

feature weights due to the size principle of the BMG [1]. It should be noticed that,

when contrasting the feature set of D14 with the definition from the text corpus: a

comprehensive school covering both primary and lower secondary education, i.e. one

year of pre-school class, the first (grade 1 to 6) and second (grade 7-9/10) stage basic

education, or in other words it caters for the 6-16/17-year-olds, it turns out that no

decisive features (age, grade etc.) that describe D14 are included in the feature set.

Hence, the result in Figure 1-c could potentially be significantly improved if the

quality of data source is reconsidered.

As described in Section 2, the BMG can, by exchanging assignment of variables x

and y, also compute probabilities that a TL audience generalizes a source concept

from a stimulus presented by an SL communicator. Hence, in Figure 2, Danish SL

concepts are defined as subjects to comparison and German TL terms as referent. This

computes probabilities, from a German TL reader´s viewpoint, that he/she possibly

infers the meanings of Danish SL concepts based on his/her prior knowledge of the

German educational concepts when a German TL term is given as translation.

Although Figure 2 shows that all four algorithms scored the highest for D4, it

demonstrates that, due to the assigned feature weights, the BMG clearly indicates that

a German TL audience will, from the given TL stimulus G5, likely infer D4. Another

noteworthy point is that similarity relations between D4 and G5 are not symmetrical,

e.g. the BMG result in Figure 2 is 82.3%, while it is 100% in Figure 1- b.

Fig. 2. Generalization probabilities: Danish as variant, German as referent.

4 Discussions

By inspecting similarity scores of all combinations between Danish and German

concepts, the results obtained from the BMG seem to reasonably identify optimally

specific translation candidates if the structured feature sets are properly prepared.

For further analyzing the performance of the BMG, Figures 3 outlines

corresponding relationships between the Danish SL concepts and the German TL

terms from a Danish communicator’s viewpoint. The corresponding relationships are

depicted by the three patterns: 1) solid thick arrows, when the probability scores are

70% or higher; 2) transparent thick arrows, when the probability scores are 40% or

higher and below 70 %; and 3) thin arrows, when the probability scores are 20% or

higher and below 40%.

Fig. 3. Ontology mapping overview: from a Danish communicator’s viewpoint.

Figure 3 also indicates that a communicator who has prior-knowledge of the

Danish educational system (gray filled square box) observes each German TL concept

as translation candidate and assess whether each German TL concept falls under the

class of each Danish SL concept. A more concrete and imaginable picture would be

that a communicator whose mother tongue is Danish seeks for a translation candidate

in his/her non-native language (German). For example, in a situation where a Danish

communicator looks for a German translation candidate for a concept D2, all of the

German terms within the relevant transparent square box, from G2 to G9, respectively

falls under the class D2 with the probability range [70 ; 100].

On the other hand, Figure 4 illustrates that a German TL information receiver

possibly generalizes the meanings of Danish SL concepts from a given German TL

translation as stimulus based on his/her prior knowledge of the German educational

concepts (gray filled square box). For instance, if he/she observes a German stimulus,

G3, he/she will likely infer some of the Danish source concepts within the relevant

transparent square box, from D2 to D5, with the probability range [40 ; 70[ that is

lower than the case of the German stimulus, G2 with the probability range [70 ; 100].

Although the BMG [1] can be quite useful as algorithm for linking multilingual

culturally-specific concepts existing for two cultures, there are still some

unsatisfactory results that have been identified in this study. For example, in both

Figures 3 and 4, the German concept, G11, has relations with D8. According to our

intuitive assessment based on the basic domain-knowledge, G11 should rather be

relevant to some of the concepts among D13-D18. When inspecting the feature sets of

G11 and D8/ D13-D18, it becomes obvious that, while G11 contains a feature “[10-16

y.o.]”, D13-D18 which refers to the Danish formal primary education for children 6-

16 years old does not contain features referring to age range. Instead, D8 contains the

important definitional feature about the age. This problem has been caused, not by the

BMG, but by the particularly strict principles for constructing TOs which may risk

causing the elimination of important features. This issue indicates that, if some

decisive features are lacking or some irrelevant features are included, the results

obtained from the BMG can immediately be affected. Hence, a future attempt would

be to investigate how to generate appropriate feature sets, that is, a more flexible

taxonomic organization of feature structures based on terms and definitions identified

in domain-specific parallel corpora. This may improve the mapping of culturally-

specific concepts applying the BMG. Another key point is that the analysis performed

here is a rather subjective assessment. Hence, for future undertakings, it is necessary

to identify an appropriate method based on assessments using human subjects.

Fig. 4. Ontology mapping overview: from a German audience’s viewpoint.

5 Conclusions

In this work, the Bayesian Model of Generalization [1] and Tversky’s set-theoretic

model [7] have been applied to data-sets consisting of culturally-specific concepts and

of features extracted from data sources based on Terminological Ontologies [5]. The

results indicate that, if input data-sets consisting of culturally-specific concepts and of

feature-values in two cultures are properly prepared, the BMG [1] can be uniquely

used not only for identifying a TL translation candidate, but also for estimating

probabilities of how a TL information receiver generalizes an SL concept from a

given TL translation. To successfully promote the next step for an overall framework,

a human based assessment of concept mappings as well as an improvement of the

method to create highly appropriate feature sets, will be required.

References

1. Tenenbaum, J. B., Griffiths, T. L., 2001. Generalization, Similarity, and Bayesian

Inference. In: Behavioral and Brain Sciences 24, 629-640.

2. Declerck, T. et al., 2010. Ontology-based multilingual access to financial reports for

sharing business knowledge across Europe. In: Rooz, J., Ivanyos, J. (Eds.) Internal

Financial Control Assessment Applying Multilingual Ontology Framework, Budapest:

HVG Press, 67-76.

3. Vossen, P. et al., 2008. KYOTO: A system for mining, structuring and distributing

knowledge across languages and cultures. In: Proc. The 6

International Conference on

Language Resources and Evaluation, Morocco, 1462-1469.

4. Cimiano, P. et al., 2010. A Note on Ontology Localization. In: Journal of Applied Ontology

Vol. 5, No. 2, IOS Press, 127-137.

5. Madsen, B. N., Thomsen, H. E., Vikner, C., 2004a. Principles of a system for

terminological concept modelling. In: Proc. The 4

International Conference on Language

Resources and Evaluation, ELRA, 15-19.

6. Tversky, A., 1977. Features of Similarity. In: Psychological Review 84, 327-352.

7. Sperber, D., Wilson, D., 1986. Relevance: Communication and Cognition. Oxford:

Blackwell

8. Glückstad, F. K., 2011. Asymmetric Similarity and Cross-Cultural Communication Process.

In: 9th International Conference on Terminology and Artificial Intelligence: Proceedings of

the Conference. 8-10 November 2011, Paris, France. Paris : Institut National des Langues et

Civilisations Orientales , 2011. p. 59-65.

9. Jaccard, P., 1901. Étude comparative de la distribution florale dans une portion des Alpes et

des Jura. Bulletin de la Société Vaudoise des Sciences Naturelles 37: 547–579.