304 ◾ Yan Tang, and Robert Meersman, and Jan Vanthienen
produce an accurate or nonfuzzy decision output. e imprecision inherent in
decision making is not noticed.” An FDT includes fuzziness in the condition and/or
action/decision segments. For example, in Table11.3, the condition entries for short,
about average, long, a little and the action stubs (possible blood deciency and possible
Qi deciency) are described with fuzzy words. e condition and/or action congura-
tion in an FDT is a value in [0, 1]. is approach brings fuzzy logic to decision tables.
11.3 Semantic Decision Tables
e semantic decision table (SDT) concept (Tang and Meersman 2007, Tang and
Meersman 2008, Tang et al. 2008) is introduced to deal with certain problems in
designing decision tables within a community (decision group):
1. Ambiguity in the information representation of the condition stubs and
action stubs. For example, Angelica X in Table11.1 is ambiguous because
Table11.2 Example of an SODT
Condition
Symptoms BD, QD, IN BD, QD BD, IN QD, IN BD QD IN
Decision
Donkey-hide
gelatin
Yes Yes Yes
Angelica X Yes Yes Yes Yes Yes Yes
Eel Yes Yes Yes Yes
BD = blood deficiency; QD = Q
i
deficiency; IN = indigestion.
Table11.3 Example of a Fuzzy Decision Table
Condition
Headache No A little
Sleep Good
About
average Bad Good
About
average Bad
Decision
Possible blood deficiency * *
Possible Qi deficiency * *
Adding Semantics to Decision Tables ◾ 305
Angelica can refer to many species such as Angelica atropurpurea and
Angelica arguta.
2. Conceptual duplication that occurs between conditions, actions, and decision
rules. For example, in Table11.1, the indigestion symptom may be included
among the symptoms of the Qi denition.
3. Uncertainty in the condition entries, e.g., possible blood deciency in
Table11.3 contains a certain level of uncertainty.
4. Diculties in managing large tables (also known as the structural scalability
problem).
5. Specication of hidden or meta decision rules. A hidden decision rule in
Table11.1 might be, “Donkey-hide gelatin is not easily digested; therefore, the
doctor should not prescribe it for a patient who has indigestion problems.”
We use domain ontologies to store semantics from an SDT. With OE technologies,
we can easily disambiguate the decision items, check the conceptual duplications
in a decision table, and specify the uncertainties of the condition entries. Just as an
ideal ontology is extensible and scalable, an SDT is easy and ready to be extended.
While constructing SDTs, hidden and implicit (meta) decision rules are specied
and stored. Since the SDTs are formal commitments based on agreement within a
decision group, they are easily shared.
An SDT is modeled in a three-layer format: (1) decision binary fact types called
SDT lexons; (2) a commitment layer containing constraints and axioms of the fact
types; and (3) decision tasks or applications. e three-layer format was based on the
principles of developing ontology-grounded methods and applications (DOGMA,
Meersman 2001, Spyns et al. 2002) and has served as the main research topic at the
VUB STARLab for 10 years.
An SDT lexon is a quintuple like (γ, disease, has, is of, symptom), in which γ is
the context identier and points to the documents in which disease and symptom are
dened; has and is of represent two roles. A lexon is a fact type.
An SDT commitment corresponds to an explicit instance of an intentional
interpretation for a decision task. Semantic decision rules (including ontological
constraints cited in some OE papers) are stored at this level. We primarily consider
uniqueness, mandatory nature, subset, equality, exclusion, subtype, occurrence,
and frequency when using SDT for data management.
An SDT is the result of annotating a decision table with ontologies. During the anno-
tation process, the decision makers must specify all the hidden rules, for example, one with
a mandatory constraint stating that “EACH disease has AT LEAST ONE symptom.”
is rule is written in semantic decision rule language (SDRule-L, Tang and Meersman
2009) as an SDT commitment: P1 = [disease, has, is of, symptom]: UNIQ (P1).
With these “extra” rules, we can benet from the reasoning advantages brought
forward by OE.
Among the many other interesting use cases of SDTs, one is to embed an SDT in
a process and separate decision rules from the process to improve system tractability
306 ◾ Yan Tang, and Robert Meersman, and Jan Vanthienen
(Tang and Meersman 2007). An important feasibility provided by SDTs is visual-
izing the results and decision rules in the form of decision tables when the decisions
are taken in every decision process. SDTs also increase exibility at the system
management level because the knowledge engineers can create dierent algorithms
and decision rules based on their needs.
Another interesting use is applying SDTs to visualize decision tables at dierent lev-
els that contain instance data at the lowest level and (meta) decision rules at higher levels
(Tang et al. 2008). e idea is to illustrate the meta decision rules of a decision table.
11.3.1 SDTs for Supporting Ontology-Based Data Matching
In the EC Prolix project,* an ontology-based data matching framework was required
to enhance a matching engine in the domains of human resource management and
e-learning. Note that ontology-based data matching is not the same problem as ontol-
ogy matching. e goal of the latter is to resolve semantic inconsistency while inte-
grating more than two ontologies. e goal of the former (the scope of our work) is
to nd the similarities between two data sets, each of which corresponds to one part
of the ontology. ere is only one ontology in that problem.
By considering this ontology as a connected network, the issue of ontology-
based data matching (ODM) can then be viewed as nding connections between
two subnetworks (Korn et al. 1996, Venkateswaran et al. 2006) and the related
work of nding and measuring such connections. Venkateswaran et al. (2006)
focus on nding data objects or web pages belonging to the same (or similar) con-
texts. Korn et al. (1996) used a two-phase method to discover data based on feature
distance. Barrett et al. (2008) illustrate an algorithm of nding the shortest path
between two nodes in a labeled and weighted network.
Our work does not focus on how to draw a boundary between searching
spaces like that of Venkateswaran et al. (2006). We calculate similarity scores
between two subnetworks (or nodes) based on the shortest path between the sub-
networks. We use SDTs to grant semantics to the labeled arcs and weights of the
ontology graph to provide an extra layer of conceptual analysis to a network. In
the next subsection, we illustrate an example of semantic matching in e-learning
and human resource management.
11.3.2 Human Resource Management Use Case
A human resource management (HRM) department uses text descriptions of com-
pany values (such as trustworthy, straightforward, and heart) to evaluate its employ-
ees. e results are recorded on evaluation forms. e training department uses
competency notations of skills and abilities to categorize utilization of learning
courses and materials. Each department uses its own supporting tools and terms.
*
http://www.prolixproject.org/
Adding Semantics to Decision Tables ◾ 307
To achieve better collaboration between departments and enhance the interopera-
bility of the applications across departments, we developed an HRM ontology with
which we annotate the company values and learning components. Each company
value or learning material corresponds to one subgraph in the same ontology.
Figure11.1 illustrates the use case by considering an ontology as a graph and
an annotation result set as a subgraph. e subgraph indicated by straightforward
& heart is the union of the annotation sets of straightforward and heart. e course
ITIl1 subgraph is the annotation set of ITIL1. Note that in our problem settings,
every item (company values and learning) is annotated with the domain ontology.
If a knowledge engineer cannot nd a proper dened concept, then he or she must
dene the new concept by executing the ontology creation methods listed earlier.
We now illustrate our approach based on this use case.
11.3.3 SDTs and GRASIM for Semantic Matching
Our matching strategy consists of the following steps: (1) label the arcs on an ontol-
ogy graph; (2) choose an algorithm for calculating the shortest paths; and (3) con-
vert the values of the shortest paths into a similarity score. SDTs are used to propose
weights to end users for labeling the arcs in step (1). e default weights are calcu-
lated based on the decision rules stored in the SDTs.
After an engineer generates default weights based on the SDTs, he can update
the weights if he is not satised with them. We also use the semantics in these SDTs
to restrict the boundaries of the weights modied by the engineers.
Table11.4 shows an SDT with which our semantic matching engine automati-
cally assigns weights on the ontology graph. Figure 11.2 shows an example for
applying the decision rule specied by column 2 of Table11.4. l
1
=〈γ, CV, describe,
is described by, Network engineer〉, l
2
=〈γ, Network engineer, is a supertype of, engi-
neer〉, l
1
∈
L ⊏ Ω, l
2
∈ Ω and l
2
∉L. According to the decision rule, the weight range
is [0, 50]. e engine can choose three dierent ways (pessimistic, optimistic, and
moderate) to assign the exact values as the weights.
If it takes the pessimistic action, the largest permitted values will be assigned
(see Table11.5). If it takes the optimistic action, the smallest permitted values
will be assigned (see Table11.6). e moderate action will lead to a set of aver-
ages from the two situations above. As noted, a user can update the weights if
he is unsatised with them. Table11.7 shows an SDT containing user-specied
weights. Note that not all the conditions from Table11.4 are used in Table11.7.
Users only need to specify the weights based on the conditions they consider
necessary to specify.
After the weights are assigned, we may choose any shortest path algorithm to
nd the shortest path for one node from one annotation subgraph (suppose it is G
1
)
to another node in the other annotation subgraph (suppose it is G
2
). It is executed
in step (2). In step (3), we convert the shortest path values into a similarity score
that must fall in the range of [0, 1].
308 ◾ Yan Tang, and Robert Meersman, and Jan Vanthienen
Straightforward & Heart
person
Interact with
context
Respect
/is respected by
Empathize
/is empathized with by
emotion
manage
/is managed by
communication
clarity
Simplicity
Has characteristic of
/is char. of
Has characteristic of
/is char. of
Course ITIL1
employee
Practice
service
function
role
Process model
Characteristic of process
Is listed
/by list
describe
/is described by
define
/is defined by
distinguish
/is distinguished by
define
/is defined by
define
is defined by
describe
/is described by
Deal with
/is dealt with by
describe
/is described by
Figure 11.1 Company values (straightforward and heart) from the HRM department and learning course (ITIL1) from training/e-
learning department correspond to two sub-graphs of HRM ontology.
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.