=Paper=
{{Paper
|id=Vol-1293/paper11
|storemode=property
|title=Using Semantic Lifting for Improving Educational Process Models Discovery and Analysis
|pdfUrl=https://ceur-ws.org/Vol-1293/paper11.pdf
|volume=Vol-1293
|dblpUrl=https://dblp.org/rec/conf/simpda/CairnsOGFSJK14
}}
==Using Semantic Lifting for Improving Educational Process Models Discovery and Analysis==
https://ceur-ws.org/Vol-1293/paper11.pdf
Using Semantic Lifting for Improving Educational
Process Models Discovery and Analysis
Awatef Hicheur Cairns1, Joseph Assu Ondo1, Billel Gueni1, Mehdi Fhima1, Marcel
Schwarcfeld1, Christian Joubert1, Nasser Khelifa2,
ALTRAN Research, 2 ALTRAN Institute
1
Vélizy-Villacoublay, France
{awatef.hicheurcairns, joseph.assu, billel.gueni,
mehdi.fhima, marcel.schwarcfeld, christian.joubert,
nasser.khelifa}@altran.com
Abstract. Educational process mining is an emerging field in the educational
data mining (EDM) discipline, concerned with discovering, analyzing, and
improving educational processes based on information hidden in datasets and
logs. These data are recorded by educational systems in different forms and at
different levels of granularity. Often, process discovery and analysis techniques
applied in the educational field have relied exclusively on the syntax of labels
in databases. Such techniques are very sensitive to data heterogeneity, label-
name variation and their frequent changes. Consequently, large educational
process models are discovered without any hierarchy or structuring. In this
paper we show how by linking labels in event logs to their underlying
semantics, we can bring educational processes discovery to the conceptual
level. In this way, more accurate and compact educational processes can be
mined and analyzed at different levels of abstraction. We have tested this
approach using the process mining Framework ProM 5.2.
Keywords: Semantic Process Mining, Educational Process Mining, Ontology,
Semantic Matching, ProM.
1 Introduction
Nowadays, education and training centers promote personalized curriculums where
students are free to choose the skills they want to develop (from beginner to specialist),
the way they want to learn (theoretical or practical aspects) and the time they want to
spend. This tendency is reinforced by the emergence of "e-learning” which represents
an increasing proportion of the in-company trainings. Educational systems support a
large volume of data, coming from multiple sources and stored in various formats and
at different granularity levels [6], [16]. These data can be exploited by instructors to
understand students’ learning habits, the factors influencing their performance and
their target skills. To answer these questions, there is an increasing research interest in
using process mining in education [6],[10], [15], [16]. The idea of process mining [1]
is to discover, monitor and improve real processes (i.e., not assumed processes) by
extracting knowledge from event logs (recorded by an information system). However,
the proposed approaches for process models extraction in the education field are
somewhat limited because they rely on classical process mining techniques which are
purely syntax oriented i.e. based on the labels in event logs [2]. For instance, we have
encountered a massive professional training dataset of a worldwide consulting
company where depending on the country and the region involved different names
were used for the same training. So, the actual semantics behind the trainings’ labels
remain in the head of education management people (e.g. teachers, carrier advisors,
etc.) who have to interpret them. To handle this question, semantic annotations on
event logs could be used to prevent such interpretation efforts [2], [3]. To benefit
from the actual semantics behind these labels, semantic process mining techniques
were introduced in [2], [3], [4], leveraging mining and analysis techniques to the
conceptual level. In this paper, we show how semantic process mining ideas may help
to discover simplified educational process models and to extract more knowledge
about their properties. For the first time, to our knowledge, a professional training
dataset of a consulting company is taken as a case study to extract and analyze
training paths annotated with semantic information. Also, we propose a
(semi)automatic procedure used to associate semantics to training labels. The
remainder of this paper is organized as follows. Section 2 summaries educational
process mining techniques. Section 3 presents the semantic process mining core idea.
Section 4 explains our approach to extract educational process models annotated with
semantic information. Finally, section 5 concludes the paper.
2 Process Mining in the Educational Field
Process mining is a relatively new technology which emerged from the information
technology and management science [1]. It focuses on the development of automated
techniques to extract process-related knowledge from event logs. An event log
corresponds to a set of process instances (i.e. traces) following a business process.
Each recorded event refers to an activity and is related to a particular process instance.
An event can have a timestamp and a performer (i.e. a person or a device executing or
initiating an activity). Educational Process Mining (EPM) refers to the application of
process mining techniques in the education domain [16]. Educational event logs may
include students’ registration procedures, student’s examination traces or activity logs
in e-learning environments. The three major types of process mining techniques are
(cf. Fig. 1): Process model discovery takes an event log and produces a complete
process model able to reproduce the behavior observed in this log. Conformance
checking aims at monitoring deviations between observed behaviors in event logs and
process models or predefined business rules and constraints. Process model extension
aims to improve a given process model based on information (e.g., time, performance,
case attributes, decision rules…etc.) extracted from an event log related to the same
process. Regarding available process mining tools, the ProM Framework is the most
complete and powerful one aimed at process analysis and discovery from all
perspectives (process, organizational and case perspective) [8]. It is implemented as
an open-source Java application with an extendable pluggable architecture.
Fig 1. Process mining concepts
ProM supports a wide range of techniques for process discovery, conformance
analysis and model extension, as well as many other tools like conversion, import and
export plug-ins. The de facto standard for storing and exchanging events logs are the
MXML (Mining eXtensible Markup Language) format or more recently the XES
(eXtensible Event Stream) format. In practice, however, ProM presents certain issues
of flexibility and scalability which limit its effectiveness in handling large logs from
complex industrial applications [13]. We may get over these limitations by using the
service oriented architecture of the ProM 6 framework. Theoretically, such
architecture may allow the distribution of ProM’s plugins over multiple computers
(e.g., grid computing). We are recently testing such a construction in the development
of an interactive and distributed platform tailored for educational process discovery
and analysis. Let us note that, lately, educational process mining has emerged as a
promising and active research field [6], [15], [16]. However, the application of
process discovery techniques presents some challenges given the huge volume and the
traces’ heterogeneity often encountered in educational datasets. In fact, when
analyzing event logs containing a lot of distinct traces, traditional process discovery
techniques generate highly complex models (i.e. spaghetti models) [13]. In this case,
the adoption of filtering, abstraction or clustering techniques may help reduce the
complexity of the discovered process models [14], [17]. For instance, a clustering
technique was proposed in [6] to improve both the performance and readability of the
mined students’ behavior models in the context of e-learning. In our previous work
[10], we proposed a two-step clustering approach for partitioning training processes
depending on an employability indicator. We think that semantic process mining
techniques seem to be a promising area to explore in order to handle the issue of
traces’ heterogeneity and so to extract simplified process models.
3 Semantic Process Mining
The semantic process mining techniques, introduced in [2], [3] aim to analyze and
extract process-related knowledge from event logs, at the conceptual (semantic) level
[4]. The challenges for mining and monitoring processes from a semantics perspective
have been studied in the context of the European project SUPER [9]. The concept of
semantic log purging was proposed in [12], taking a case study in the higher education
domain. In [5], the authors proposed a combination of standard process mining
techniques with semantic lifting procedures on the event logs in order to mine more
precise process models. The core idea of semantic process mining is to explicitly
annotate elements in event logs with the concepts that they represent. These concepts
are formalized in generic or domain specific ontologies. Hence, semantic process
mining techniques are built on the following three basic elements: ontologies, ontology
reasoners, and references from elements in logs/models to concepts in ontologies [2].
First, ontologies define and formalize a set of concepts shared by (a group of) people to
refer to things in the world and the relationships among these concepts. Second, the
reasoner provides reasoning over the ontologies in order to derive new knowledge,
e.g., subsumption, equivalence, etc. Finally, the references associate meanings to labels
(i.e., strings) in event logs and/or models by pointing to concepts defined in ontologies.
The discovery, conformance checking, and extension techniques rely on subsumption
relations induced by these ontologies to raise the level of abstraction from the
syntactical level to the semantical level. Thus, these techniques can be applied without
requiring any modification of models or logs if the elements in different logs and
models link to the same concepts (or super/sub concepts of these concepts). Let us note
that all semantic plug-ins developed in ProM are based on the following concrete
formats for the basic building blocks: Event logs are in the SA-MXML (i.e. Semantic
Annotated Mining eXtensible Markup Language) file format. SA-MXML is a
semantically annotated version of the MXML format which incorporates the model
references (between elements in logs and concepts in ontologies). Ontologies are
defined in WSML (Web Service Modeling Language) [7], [11]. The WSML 2
Reasoner Framework [18] is used to perform all the necessary reasoning over the
ontologies.
4 Case Study: Leveraging Educational Process Mining
Techniques at the Semantic Level
Our motivating example is based on real-world training databases from a
worldwide consulting company. This company has around 6 000 employees that are
free, during their careers, to take different trainings aligned with their profiles. These
trainings are provided by internal or external organizations. The data collected for
analysis includes the employees’ profiles (demographics data), their careers (i.e. the
jobs/missions they did) and their training paths (the set of trainings taken during the
past three years) (cf. Table 1). In what follows, we apply a process model discovery
algorithm (e.g. the heuristic miner [8]) on a fragment of the training event log (cf.
table 1), containing 1000 traces, 2419 events and 280 originators. We can see that the
obtained result is an unreadable spaghetti like process model (cf. Fig. 2). This result
can be explained by the heterogeneity in employees’ training paths and the great
number of different trainings’ labels. Let us note that depending on the organization,
the country and the region involved, different labels (i.e. string) were used for the
same training. Moreover, some training courses can be seen as special cases of other
trainings. For instance, the trainings “Collective English”, “Collective Face to Face
English”, “English In Group” are in fact the same training which is given different
names following data sources. Moreover “Collective Face to Face English” is a
variant of “Face to Face English”, which is a special type of the “English” training.
Table 1. Example of an educational event log
Fig 2. Fragment of a spaghetti process describing all trainings followed by the consulting
company’s employees during the last three years. The process model was extracted using the
Heuristic Miner plug-in of ProM.
To handle this issue, we need to link different trainings which are variants or
synonyms of the same training to a unique concept in a training ontology. Usually,
there are two ways to achieve this. We can manually create all the necessary
ontologies and annotate the necessary elements in educational event logs with
ontologies’ concepts. It is also possible to use tools to (semi)automatically discover
ontologies based on the elements in these logs [4]. The discovered ontologies can be
manually improved in a second step. Let us note that semantic process mining tools
can also play a role in ontologies’ extraction and enhancement from event logs. The
ontology depicted in Fig. 3 is used to formalize the concepts for trainings in our
example. It contains 42 concepts and 129 instances. We built this ontology manually
taking as starting point the semantic description of trainings provided in training
organizations’ catalogues. We distinguished five super-concepts related to the training
domain: Communication, Staff Management, Project Management, Audit and
Control, Information Technologies.
Fig 3. Fragment of the “Training ontology”: only some instances (i.e. training labels) are represented
These concepts are subdivided into sub-concepts which are in their turn subdivided
into lower sub-concepts (cf. Fig 3). Trainings’ labels are the instances of this ontology
and each label is associated to one concept or sub-concept. To simplify the ontology
depicted in Fig 3, we only represented one instance (training label) per concept. We
used the tool WSMT (Web Service Modeling Toolkit) to implement the training
ontology in the WSML format since it is supported by the ProM 5.2 framework.
Moreover, the semantic process mining plug-ins existing in ProM 5.2 expect log
elements to be connected with process ontologies (i.e., to be in the SA-MXML
logging format). So to enrich the educational log of our example with semantic
annotations from the Training Ontology, we implement a conversion plug-in in ProM
5.2. The latter takes as input the original educational log (in MXML format) and the
Training Ontology (in WSML format) and produces the corresponding semantically
annotated event log (in SA-MXML format).
4.1 Semantic Matching Between Training Labels and Concepts
In order to help end users in the comprehension of the underlying semantics of
training courses, we develop a (semi)automatic procedure, which can be used to
associate a concept (of the training ontology) to a training label. The association used
is based on the importance of the words in a label or in a concept. We assume that
each word of a label L plays the same semantic role and hence has the same
importance as well as the other words constituting L. We also suppose that at least
one of the words characterizing a concept, or one of their synonyms, appears in all the
labels associated to it. Therefore, there is an intersection between the set of the words
of a label and the set of the words characterizing its associated concept. To build our
technique we develop the following modelling: consider W = {w1,…,wn} a set of
words, we consider a training label TLi as succession of wj, noted TLi= w1ҍ+… ҍ+ wm,
where wj∈W and the symbol ҍ represents blanks and all articles, pronouns, etc. For
instance the label “Introduction to Information Systems” contains the set of words W
= {Introduction, Information, Systems, Management} separated with three blanks and
the preposition ‘to’. We consider Li the set of the words that contains TLi, so Li =
{w1,…,wm} and in our case we assume that card(Li) represents the length of the label
TLi (we note Len(TLi)), for example Len(Introduction to Information Systems) = 3.
We also consider Cj = {w′1,…,w′k} as the set of the words characterizing a concept Cj.
Word importance: is a metric, or a weight, reflecting the importance of a word in a
label according to our hypothesis given below. As each word plays the same role in a
label we compute its importance wp as follow: wp(w) = 1/ Len(TL) where w ∈ L. for
the label TL = “Management in Information Systems”, Len(TL)=3 and
wp(Management)=1/3. This wp reflects clearly the relation between the length of a
label and the importance of its word. A small label, like ones using only one or two
words, gives a great semantic importance to its word that are considered like keys,
whereas long labels use lot of words for their description giving its words a small
semantic role.
Word concept weight: the weight of a word w in a concept C, noted cw(w),
corresponds to the sum of all word importance of w, or one of its synonyms, in all the
labels associated to the concept C: cw(w)= ∑ wpTLi(x), where i∈{1,…,h} and TLi is
associated to C. For instance, consider the concept characterized by the following
words (“management”, “project”). If “management” appears three times in the labels
with the following wp: ½, ½ and ⅓ therefore cw(“management”)= ½+½+⅓ =1.3.
This metric establishes a monotone relation between the frequency of the word in the
labels and its importance, and it is clear that more a word is used, more it is important
and more it will be used to characterize a concept.
Concept matching: to generate automatically the concept C associated to a label TL
we create first a word weight table as follow:
1. We compute the set of all the words of all the labels contained in the
training catalogue. We note this set as LW.
2. We create a matrix M = (ai,j 1≤i≤n, 1≤i≤m) ai,j is the wp(i) in the label j,
n = card(LW) and m is the number of all training Labels.
3. For each word w in LW, we sum its wp(w) computed in the previous step
and we store the result in the returned table.
After constructing this table, for a label TL we compute the semantic intersection
between L and C as follow: L ∩ C = {wj, wj ∈ L ∧ wj∈! C}. wj∈! C means that wj or a
synonym of it is included in C. Then we compute the score of matching between L
and C, noted SC(L,C) as the sum of the concept weight of each element of L ∩ C. We
repeat this operation for all the concepts we have and then we associate L with the
concept having the high score. If we have the concepts C1,…,Cn then L will be
associated to C if SC(L,C) = Max(SC(L, Cn)). The semantic importance we use in our
matching is simplified compared with approaches doing deep semantic analysis using
sophisticated techniques because we do a significant human effort to define the
Ontology with different level, and we stress on the concepts of the level 2 to enrich
them with words that are generally and mostly used to define the labels associated to
each concept of this level. We remark that if we have two or more concepts having
the same Max(SC(L, Cn)) we infer a conflict and in this case we need a user’s
intervention to choose what concept to associate to the label. We have tested this
matching technique on Altran catalogue containing 128 labels and 35 concepts. Fig 4
depicts the obtained results. Let us note that in these tests we have identified some
cases where we have not identified matching between labels and concepts.
Fig 4. The number of labels (ordinate) associated to each one of the 35 concepts (absciss) of our
case study
This is due to the use of some abbreviations that are hard to decrypt. In these tests,
concepts contain only words that we find in labels and we do not need in this case to
search synonyms. We plan in the future to use a dictionary in order to enhance the
identification of synonyms.
4.2 Educational Process Models Mining at the Conceptual Level
After constructing a semantically annotated educational log, we specify the level
of abstraction (i.e. concepts in the training ontology) used as a base for the mining and
the analysis of training processes. To achieve this, we use the filter plug-in “Ontology
Abstract Filter” implemented in Prom 5.2, which allows us to choose the required
level of abstraction [8]. The Ontology Abstract Filter plug-in takes as input a
semantically annotated event log (in SA-MXML format) and produces as output
another event log where the names of tasks (i.e. trainings) are replaced by the names
of the chosen concepts. The produced log can also be exported as an SA-MXML log.
After this step, we may apply a control-flow mining algorithm (e.g. the Heuristic
Miner plug-in) to extract the educational process model relaying on the concepts
chosen in the previous step. We may choose concepts at different level of
abstractions. When we use only the concepts at level 2 of the Training Ontology tree
(i.e., the concepts “Communication”, “Language”, “Testing”, “Audit_And_Control”,
“IT_Service_Management”…etc.”), a process model like the one in Fig. 5 could be
discovered.
Fig 5. Training process model mined using the heuristic miner plug-in where only the concepts at
the level 2 of the tree for the ontology “TrainingOntology” (cf. Fig.3) are considered.
It contains 18 events (nodes) and 30 arcs which is more compact than the model
extracted before the semantic abstraction (cf. Fig 2). Let us note that during the
abstraction phase we deliberately replace the labels of the different kind of English
trainings by their concept at level 1 (i.e. English). We can see that the mined model in
this case is more compact (i.e., has a higher abstraction level) than the one in Fig 2. In
this model we can see that trainings associated to the concept “Management” are
taken 444 times. Also, there are seven trainees who took an “English” training after a
“Management” training. The frequency associated to this relation in the educational
log is 0,889.
4.3 Educational Process Analysis at the Conceptual Level
In our case study, process mining advantages are not limited to the discovery of
employees’ training processes. In fact, training advisors and directors of training
organizations often need to check (off-line or on-line) whether trainees’ paths
conform to established career paths, trainings’ prerequisites or business rules. The
semantic LTL checker plug-in of ProM 5.2 is the perfect tool for auditing educational
processes at the conceptual level [2]. This tool can be used to verify the same formula
(e.g. generic formula such as prerequisite) on a set of different event logs as long as
the arguments of this formula and the elements in these logs link the same concepts
(or super/sub concepts of these concepts). There is a set of predefined formulas in the
semantic LTL model checker plug-in. It is also possible to tailor the semantic LTL
checker plug-in to express specific types of constraints encountered in the educational
domain [16]. All these properties can be easily coded using the LTL language and
imported into the user interface of the plug-in. In what follows we want to check if the
rule “A Project Management training must be taken before a Project Management
Professional Certification (PMP) can be taken” was always respected (prerequisite
check). We define this property in a LTL file as follows:
formula c2_is_a_prerequisite_of_c1 ( c1 : ate.WorkflowModelElement, c2 :
ate.WorkflowModelElement) :=
{ Is the training C2 a prerequisite for the training C1?
}
( <> (activity == c2) /\ (activity != c2 _U activity == c1) ) ;
Fig 6. The results returned by the semantic LTL Checker plug-in while verifying the PMP
prerequisite
Fig 6 shows the result displayed when this property is checked. We can see that there
are 26 trainees who took the PMP training while they didn’t take the Project
Management training before (i.e., incorrect case instances). There are also 718 trainees
that satisfy this property (i.e. they took the “PMP” training after a “Project
Management” training).
5 Conclusion
In this paper we showed how by associating semantic annotations to educational
event logs, more accurate and compact educational processes can be extracted and
analyzed at different levels of abstraction. Also we developed a semantic matching
procedure allowing to link training labels to the right concepts of a training ontology,
in a (semi)automatic way. In future works, we will investigate how concepts from
ontologies can be associated to training providers. We can then benefit from these
semantic annotations in mining social networks and organizational models between
training providers [1], [10], at the conceptual level. We plan also to conduct a case
study in an on-line education setting that would illustrate the benefit of process
mining approaches, at the syntactic and semantic levels, to mine and understand
students’ behaviors. Another important step in our works is to develop new clustering
and classification techniques which take into account semantic annotations on event
logs [14], [17]. For instance, trace clustering techniques [14] can be extended to
partition event logs depending on trace similarities at the conceptual level. To
implement our approach, we are currently developing an interactive and distributed
platform tailored for educational process discovery and analysis. This platform will
allow different education centers and institutions to load their data and access
advanced data mining and process mining services [10]. Moreover, in order to
optimize and enhance platform response time, our platform will allow distributing
heavy analysis computations on many processing nodes.
Acknowledgments. This ongoing work is being carried out by Altran Research and
Altran Institute within the context of the PHIDIAS project.
References
1. Aalst, W. M. P., and al.: Process Mining Manifesto, In BPM 2011 Workshops Part I,
LNBIP 99, pp. 169-194 (2012).
2. Alves de Medeiros, A.K., van der Aalst, W.M.P., Pedrinaci, C: Semantic Process
Mining Tools: Core Building Blocks. In 16th European Conference on Information
Systems, pp. 1953-1964. Galway, Ireland (2008).
3. Alves de Medeiros, A. K., . van der Aalst, W. M. P: Process Mining towards
Semantics. Advances in Web Semantics I, LNCS 4891, pp 35-80 (2009).
4. Alves de Medeiros, A. K. and al: An Outlook on Semantic Business Process Mining
and Monitoring. In R. Meersman, Z. Tari & P. Herrero (Eds.), The Confederated
International Conferences On the Move to Meaningful Internet Systems, LNCS, vol.
4806, pp. 1244-1255, Springer-Verlag (2007).
5. Azzini, A., Braghin, C., Damiani, E., Zavatarelli, F: Using Semantic Lifting for
improving Process Mining: a Data Loss Prevention System case study. In the 3rd
International Symposium on Data-driven Process Discovery and Analysis, CEUR-
WS.org, pp. 62-73 (2013).
6. Bogarín, A., Romero, C., Cerezo, R., Sánchez-Santillán, M: Clustering for improving
educational process mining. In Proceedings of the Fourth International Conference on
Learning Analytics And Knowledge. ACM, New York, NY, USA, pp. 11-15 (2014).
7. de Bruijn, J., Lausen, H., Polleres, A., Fensel, D.: The Web Service Modeling
Language WSML: An Overview. In Y. Sure and J. Domingue, editors, ESWC,
LNCS, vol. 4011, pp. 590-604. Springer (2006).
8. van Dongen, B. F., de Medeiros, A. K. A., Verbeek, H. M. W., Weijters, A. van der
Aalst, W. M. P: The Prom Framework: a New Era in Process Mining Tool Support.
In ICATPN'05, Gianfranco Ciardo and Philippe Darondeau (Eds.). LNCS, vol. 3536,
pp. 444-454, Springer-Verlag, Heidelberg (2005).
9. European Project SUPER - Semantics Utilised for Process Management withing and
between Enterprises. http://www.ip-super.org/
10. Hicheur Cairns A. and al: Custom-Designed Professional Training Contents and
Curriculums through Educational Process Mining. In The Fourth International
Conference on Advances in Information Mining and Management, pp 53-58 (2014).
11. Lausen, H., de Bruijn, J., Polleres, A., Fensel, D: The WSML Rule Languages for the
Semantic Web. W3C Workshop on Rule Languages for Interoperability, W3C
(2005).
12. Ly, L.T, Indiono, C., Mangler, J., Rinderle-Ma, S.: Data transformation and semantic
log purging for process mining. In Proceedings of the 24th international conference
on Advanced Information Systems Engineering, J. Ralyté and al. (Eds.).LNCS, vol.
7328, Springer-Verlag (2012)
13. Reichert, M.: Visualizing Large Business Process Models: Challenges, Techniques,
Applications. In 1st Int’l Workshop on Theory and Applications of Process
Visualization Presented at the BPM 2012, LNBIP, vol. 132, pp. 725-736, Springer-
Verlag (2013).
14. Song, M., Günther, C.W., van der Aalst, W.M.P.: Trace Clustering in Process
Mining. In: Ardagna, D., Mecella, M., Yang, J. (eds.) BPM 2008. LNBIP, vol. 17,
Springer, Heidelberg, pp. 109–120 (2009).
15. Trcka, N., Pechenizkiy, M: From Local Patterns to Global Models: Towards Domain
Driven Educational Process Mining. In Proc. 9th International Conference on In
telligent Systems Design and Applications, pp. 1114-1119. IEEE Computer Society
(2009).
16. Trčka, N., Pechenizkiy, M., van der Aalst, W.: Process Mining from Educational
Data (Chapter 9). In Handbook of Educational Data Mining. pp. 123-142. CRC Press
(2010).
17. Veiga, G.M., Ferreira, D.R.: Understanding Spaghetti Models with Sequence
Clustering for ProM. In: Rinderle-Ma, S., Sadiq, S., Leymann, F. (eds.) BPM 2009.
LNBIP, vol. 43, pp. 92–103, Springer, Heidelberg (2010).
18. WSML 2 Reasoner Framework (WSML2Reasoner). http://tools.deri.org/