Computer Science and Information Systems 12(3):9791008 DOI: 10.

2298/CSIS141103036S

Towards Analogy-Based Reasoning in Semantic Network

Milos Stojkovic, Milan Trifunovic, Dragan Misic, and Miodrag Manic

Faculty of Mechanical Engineering in Nis, Aleksandra Medvedeva 14,

18000 Nis, Serbia
{milos.stojkovich, milantrifunovic, misicdr, miodrag.manich}@gmail.com

Abstract. In this paper an approach in realization of analogy-based reasoning in

semantic networks is presented. New semantic model, called Active Semantic
Model (ASM), was used. Core of the process is performed by ASMs association
(semantic relation) plexus upgrading procedure based on recognition and
determining similarity between association plexuses. Determining similarity
between association plexuses is performed by recognition of topological analogy
between association plexuses. ASM responds to unpredicted input by upgrading
new association plexus modeled on remainder of the context whose subset is
recognized as topologically analogous association plexus.
Keywords: Artificial Intelligence, Analogy-Based Reasoning, Active Semantic
Model, Cognitive Data Processing, Semantic Features

1. Introduction

Example: Angelina, a preschool girl, enters the office of schools psychologist who
assesses cognitive abilities of newcoming first class pupils. After welcoming Angelina,
psychologist asks the girl to seat and look the paper laying on the table. Psychologist
begins to explain the task to Angelina: Here you have two pictures in the first row, and
your task is to choose the missing picture in the second row (Fig. 1).

Fig. 1. Angelinas task

Angelinas mother, IT expert, also present in the office, intrigued by the task, tries to
come up with computational procedure, method or approach that already exists, and
which can be employed for solving Angelinas task. She also contemplates about efforts
needed to force such system to bring proper conclusion without input pre-planning.
First its necessary to describe the pictures in a way so that system would be able to
980 Milos Stojkovic et al.

provide reasonable response. To ensure such a response, system should be prepared to

interpret pictures and their spatial relations semantically, as well the task itself, i.e. to
identify what it is expected to return. Finally, the system should be able to generate a
kind of unique (previously never defined) and valuable response based on a very small
portion of domain knowledge already modeled within it. Additionally, that system
should not be custom-made for Ravens progressive matrices solving. While mother
was thinking about the system she would need, Angelina chose the correct picture.
Thats my girl!, she thought.
Semantic interpretation of data represents one of the biggest challenges faced by
modern information technologies. In fact, this problem is closely related to the ability of
computer applications to attach certain meaning to data which is being processed. The
motive for solving this problem lies in ever increasing need to enable software
applications to provide meaningful answers when it is not possible to predict the input,
and consequently the code by which a meaningful response is programmed.

2. Related Work

2.1. Issues with Reasoning and Learning in DL Based Ontologies

The most exploited current model of knowledge representation in form of semantic

network is ontology. Actually, ontology in IT and AI contexts usually refers to a kind of
vocabulary of terms (concepts) and relations among the terms codified in a description
logic (DL) formalism that should enable a computer application to interpret meaning of
the terms from the vocabulary [1]. Aiming the ultimate objective to enable computer to
draw meaningful response on unpredicted input, which is essentially related to its ability
to interpret semantics of the input data, IT experts in last two decades were focusing
mainly on development of DL-based ontologies that utilize first-principles reasoning for
semantic interpretation. First-principles reasoning ensures logical deduction applying
logical inference rules on axioms (mainly employing First-Order Logic FOL) that are
related to closed domain of richly axiomatized discourse [2]. In particular, almost all
DL oriented reasoners are based on tableau-based decision algorithms or resolution-
based decision procedures [3], [4], [5]. Besides the core set of logical rules, these
semantic models usually also provide a production rules dialect [6] which allows creator
of vocabulary to define domain specific rules of inference to ensure correct entailments
for the case of not logically derivable semantic interpretation.
By insisting on richly axiomatized ontologies current semantic models are struggling
to stay in domain of strongly structured knowledge which allows application of first-
principles reasoning approach. The remarkable effort has been carried out in last two
decades to generate a large set of different ontologies that demonstrate undoubted power
of DL and FOL. However, the most common real situations where computer-
autonomous semantic interpretation is highly required are related to data sets whose
semantics is not consistently and precisely logically modeled. Besides, constructing
richly axiomatized ontologies for real world knowledge intensive applications is a time
consuming and difficult task [7], [8], which often results in incompleteness of ontology.
 Towards Analogy-Based Reasoning in Semantic Network 981

Due to lack of explicit specification of vocabulary terms intended meaning or

insufficiently structured data semantics, it is hard for current semantic reasoners to
provide relevant entailments [9], [10]. The same cause generates challenges related to
ontology learning, assertions populating, ontology enrichment and evolution, as well as
ontology matching and mapping [11], [12]. To extract ontological elements from an
input and learn about new ontology elements or to enrich existing ontology from that
input autonomously (by computer application), it is necessary for the set of input
elements to be already formalized in a way that automated inference will yield the
expected results [11], [13]. To match and align two different ontologies, even
semantically very close, but created by different domain experts, first and inevitable
step is identifying common concepts [14], [15], [16], [17], [18]. The key question is
how to measure commonality, i.e. similarity between differently conceptualized and
hence differently described concepts; even more, how to measure similarity between
concepts of semantically very distant vocabularies (completely disjointed) [19]. Existing
research in the domain of ontology alignment has developed several semi-automatic
approaches for measuring concept similarity [14], [17], like lexical similarity between
concepts, structural similarity of concepts in ontologies (similarity of
ancestors/descendant, depth and length of path in the tree) and similarity of concept
instances/annotations. Nevertheless, in most of the real word cases, ontology alignment
process mainly relies on human interventions (i.e. similarity assessments) [14], [19].
After recognition of similarity between concepts there are two more activities to
perform before engaging semantic reasoners to produce entailments autonomously.
First, there is a need to formalize and/or harmonize semantics of concepts with DL to
ensure alignment and secondly, one should identify mapping rules, which is very
complex and nonautomated task [15]. Aforementioned activities (needed to learn, enrich
and/or align onotologies codified in DL) can only be performed if domain experts
collaborate with skilled ontology engineers familiar with the theory and practice of
knowledge representation [11].
It seems that strongly structured knowledge approach within DL-founded ontologies
is certainly powerful tool for deduction in the local field of semantics, but at the same
time it reduces capability to infer autonomously and flexibly. Actually, semantic
reasoners designed to work with DL-founded ontologies showed themselves weak in
making relevant entailments beyond the predefined and embedded logical formalism of
deduction. The similar is also true for reasoning flexibility ability to make a relevant,
but quite different entailments about the same concept for semantically distant or
different contexts (vocabularies) with a single set of logical inference rules and axioms
on disposal. Finally, having analytical ability to autonomously dissolve a portion of
knowledge about one concept or group of concepts from one context and apply it to
quite different (semantically distant) concept or a group of concepts that are inherent to
equally different context is something which appears not as strong side of richly
axiomatized ontologies which rely on first-principles reasoning approach.
Often emphasized, essential weakness of ontologies, as well as all semantic models,
is reflected in the fact that an increase in level of detail of the meaning to be described,
significantly increases complexity and time required to create the ontology [20].
Therefore, much more ontologies that do not require a high level of semantic detail
were developed.
982 Milos Stojkovic et al.

2.2. Why to develop new semantic model?

In order to overcome obvious weaknesses of DL ontologies it becomes more attractive

to explore whether the cognitive process of analogy making or analogical reasoning can
serve as the basis for ontology learning and alignment processes [21], [22], [23]. In
addition, growing exploration within the field of DL ontologies similarity indicates that
focus is moving to analogy-based identification of semantic correspondence between
ontologies [14], [17], [24]. As Forbus et al. state in [2] analogy based reasoning is a
method of last resort for DL ontologies issues, something to use only when other
forms of inference have failed.
Familiarizing with research [25] on how humans derive knowledge from analogies
between symbols, concepts, situations and events inspired us to try to create our own
semantic model with better capacity to perform analogy based reasoning. We found that
links structure in a semantic network needed to be changed slightly in order to
determine and categorize similarities between concepts, situations and events more
efficiently.
However, before we get down to describe in-house developed semantic model
featured by original structure of links, we need to give a short survey on Analogies and
Analogy-Based Reasoning.

2.3. Analogies

Achieving autonomy, flexibility and analyticity of semantic interpretation is considered

a major current goal of all artificial intelligence methods and models, including
ontologies [1]. In pursuit for solution, interest for approach where semantic
interpretation of data is based on analogies reappears [26]. Research in cognitive
psychology often indicate that use of analogies represents the core of cognitive process,
and may be considered as primary process of cognition and communication [27].
Traditional logic distinguishes three forms of reasoning: deductive, inductive, and
analogy-based reasoning (ABR). Examples of heuristics most commonly used for
solving problems are determination of partial goals and reliance on analogies [28]. In
the latter case, known procedure, which proved to be successful in solving previous
related (similar) problems is used to solve new problem. Precondition for success of this
strategy is recognition of analogy between two problems and recalling the solution
applied earlier. One of the reasons why sometimes it is difficult to recognize analogy
between two problems is the fact that their elements are different, although relations are
the same [25].
ABR is often used to characterize methods that solve new problems based on past
cases from a different domain, while typical case-based reasoning (CBR) methods focus
on indexing and matching strategies for single-domain cases [29]. In general, analogy
involves several subtasks including retrieving from memory the source case most
similar to the target problem, mapping (or aligning) the elements of target and the
source, transferring knowledge from the source to the target, evaluating what was
transferred in the context of the target, and storing the target in memory [30].
Three major types of case representation are feature vector cases, structured cases,
and textual cases [31]. Feature vector approach represents a case as a vector of attribute-
 Towards Analogy-Based Reasoning in Semantic Network 983

value pairs, while structured approach as clusters of relations between the kinds of
elementary objects that comprise it [32].
Case representation and the way similarity is assessed during retrieval are strongly
related to each other. In some applications of ABR, similarity of stored cases is assessed
in terms of their surface features, which are parts of their description typically
represented using attribute-value pairs. Various methods exist: k-nearest neighbor (k-
NN) based on Euclidean distance, mix neural networks [33], fuzzy logic [34], and
genetic algorithms [35]. Structured cases often require knowledge intensive matching
algorithms to assess structural similarity. Experiments confirmed that both surface and
structural similarity assessment are necessary for sound retrieval [36], [37]. Structural
features, however, have a greater impact than surface features on a problem solvers
ability to use an analogue once its relevance has been pointed out [37]. Retrieval based
solely on similarity has limitations. Thats why similarity is increasingly being
combined with other criteria to guide the retrieval process, such as adaptability of the
retrieved case [38], [39].
ABR is considered the most flexible and analytical approach within the corpus of
CBR systems [29], [40]. MARVIN [40] is interesting and very expressive example. It is
a system for general knowledge representation in form of analogies, and
graphical/tabular visualization and searching for analogies, i.e. analogy-based
reasoning. Whats special about this example is that the system uses XML syntax for
representing and visualizing analogies. Searching for analogies in this system is based
on the so-called superficial similarity of analogies (full or partial match of nodes names
in the structure). Additional structural search mechanism traces synonyms, hypernyms
relations and other k-level generalizations in order to extend the set of potential
analogies which could be used for reasoning. This approach of structural mapping, i.e.
searching graphs of semantic or functional model is dominant in many other earlier or
later realizations. Unfortunately, it doesnt bring, nor demonstrate, full potential and
advantage of ABR approach, but often discredits it as too limited and arbitrary.
Research on analogy reasoning is concerned with mechanisms for identification and
utilization of cross-domain analogies [41], [42]. The major focus has been on finding a
way to transfer (or map) the solution of an identified analogue problem to the present
problem. Analogical mapping has been studied in many theories of analogy, such as
Proteus [30], AMBR [43], [44], LISA [45], SME [46], and ACME [47].
In order to recognize analogy between two problems, it is necessary to have insight
into common elements of the solution which can be applied to new problem. This
insight is actually contained in similarity and/or sameness of relations between these
elements [25]. Realization of this claim is the main objective of Active Semantic Model
(ASM) to embed knowledge in semantic relations and their plexuses (not in the nodes
of the semantic network), and also to try to recognize analogies by determining the
similarity of semantic relations and their plexuses in order to interpret the meaning and
draw conclusions. Structure of ASM [48], and approaches to semantic categorization of
data [49], [50], and recognizing topological analogy in ASM semantic network were
presented so far. This paper proposes to extend ASM for ABR.
ABR brought by ASM allows us to overcome the need for: 1) pre-planned conditions
which have to be fulfilled in order to trigger predefined response, i.e. inference, and 2)
standardization of nomenclature. It provides system with ability to make creative
reactions, and to ensure relevant answer with minimal investment in preparation.
984 Milos Stojkovic et al.

3. Active Semantic Model

ASM is a sort of semantic network model, developed in-house, aiming to capture and
interpret semantics of design features related to manufacturability issues [48]. The most
frequent representation of semantic network is graph notation consisting of nodes and
links (or edges, arcs). Nodes usually represent concepts, objects or situations in a
particular domain, while links usually represent the semantic relations between these
concepts, objects or situations. More complex link structure is what distinguishes ASM
from other semantic network models. Actually, decision to focus on link structure came
from the thesis stating that the knowledge people have about things (visual
representations, objects, situations, etc.) is contained in associations between concepts
that abstractly represent those things [51]. Beside functional relation between concepts,
ASMs link express also its affiliation, accuracy, and significance for specific context,
and for particular instructor (user). In this way semantic link provides chunk of
knowledge which is subjective and context related. Furthermore, each semantic link
bears information about direction and character of associating between concepts (that is,
about the way in which semantic interpretation should be made). This feature of ASMs
link to point out the pathway of inference induced us to use the term association for
the link instead of relation. Here, we will explain ASM in brief.

3.1. Structure

The structure of ASM is built just from associations (links of network). Each
association is characterized by eleven parameters [49] among which two of them are
names of concepts (cpti, cptj). Considering that these parameters can belong to more
than one association, they represent junctions of associations, i.e. virtual nodes of
network. The explicit knowledge related to concepts and their instances is also linked to
these virtual nodes of ASMs network via associations. On the other side of the
association which connects concept of ASM network with a chunk of explicit
knowledge about it is a pointer to a certain local computer or internet depository where
the knowledge is stored. These pointers are named concept bodies, because they point
out to some kind of knowledge embodiments of each instance of the concept. For
example, concept Blue-Color can be embodied by one or plenty of specific values of
color codes and procedure to generate this color on the computer screen in accordance
to its code. Thus, one concept can have several concept bodies, i.e. its real represents.
The parameter name of the concept in an association is used to designate human
abstraction of different level of complexity, from Something concept to very complex
spatial and time and/or abstract contexts. The Something concepts can be tangible
(e.g. pencil) and intangible (e.g. geometrical shape, circle) objects, attributes (e.g. blue),
activities (e.g. cause, use), or abstract ideas (e.g. number, below). The spatial and time
contexts can be different situations and events. The abstract context can be e.g.
differential equation. There can be only one concept with a given name, but there can be
many associations belonging to different contexts associating it with other concepts.
Beside two different names of concept, an association in ASM is defined by
additional three sets of parameters:
 Towards Analogy-Based Reasoning in Semantic Network 985

Topological parameters:
roles (ri, rj) of concepts within association denote what functional role each of these
concepts have in their mutual association (e.g. Photo associates to Photo-Album as a
part to assembly).
type (t) denotes type of associating (e.g. affiliation in the aforementioned association
between Photo and Photo-Album). Actually, roles and type of associating make a
determinate triplet. For semantically unsymmetrical types of associations (e.g.
affiliation) roles indicate the default orientation of association.
direction (d) of associating (, , ) denotes whether both concepts linked by
association associate to each other or just one of them associates to another and not vice
versa (e.g. in the association between Triangle and Geometric-Shape, Triangle almost
always associates to Geometric-Shape, but Geometric-Shape doesnt associate
necessarily to Triangle). It should be noted that direction is not a parameter which
indicates the direction of deduction process (graph routing).
character (c) of associating (+, -) denotes how both concepts in an association
associate to each other (positive character denotes that concepts associate to each other
affirmatively, like Ball and Oval when we think about the Ball we think that it is Oval,
while negative character denotes that concepts associate to each other, but negatively,
like Ball and Cubical Ball is something which is not Cubical).
Weight parameters:
accuracy (h) of an association for the given context (0; 0.25; 0.5; 0.75; 1) denotes
how accurate are the values of all other parameters. 0 accuracy denotes that association
(values of the other parameters) is untrue and 1 denotes that instructor (user) is
convinced in absolute correctness of other parameters values of association (of course
for the given context). Untrue associations are important because these associations
indicate misapprehensions which are, sometime, very important for inference process.
These associations indicate what segments of network should be ignore by inference
engine.
significance (s) of an association for the given context (0; 0.25; 0.5; 0.75; 1) denotes
how significant this association is for semantic interpretation of related concepts in a
given context. For example, in a context related to the Ravens problem shown in Fig. 1
the associations which can exist and are used to describe that Circle has no Corners are
not as significant as the associations which are used to describe Circle spatial location
with regard to other geometric shapes. This parameter of association can help ASM
inference engine to categorize Circle from two pictures in the first row as similar to
Triangle from the picture in the second row, even though the associations that are used
to describe geometric features of Triangle (e.g. it has Corners and it is not Oval) are
quite opposite to the associations which describe geometric features of Circle.
Affiliation parameters:
context id is a parameter which denotes to which context an association belongs, i.e.
in regard to which context the values of associations parameters are valid. This feature
allows instructor (user) to describe and ASMs inference engine to interpret the very
same concept in semantically quite different way for different contexts.
instructor id or user id is a parameter which denotes who has created the association
(Fig. 2). Like association affiliation to context, the origin of an association with regard
to its creator allows instructors (users) to describe and, later, AMSs inference engine to
interpret the semantics of the same concept in a different way. That is how ASM
provides possibility to add subjectivity to semantics of a concept. After all, we should
986 Milos Stojkovic et al.

never forget that values for aforementioned parameters are assigned based on the
instructors (users) subjective assessment.
Beside associations (links) as the basic structure of ASMs network, the association
plexuses have very important role in inference process that ASM carries out. Actually,
association plexus (PLX) is a term used to denote a set of at least two associations
connected by the mutual virtual node(s) (name(s) of concepts). In general, association
plexus doesnt need to have specific abstract meaning, neither instructor (user) needs to
define it (like an association). However, perceiving the semantic network of ASM
segregated not just by its basic elements associations, but also by plexuses facilitates
the identification of similarity or analogy of topology between different segments of the
semantic network which drives analogy-based reasoning process in the core of inference
engine of the ASM. Besides the association plexuses which are not created by instructor
and may or do not have any abstract meaning, there is a possibility for instructor (user)
to create (usually) more complex plexuses that serve to describe the semantics of
complex concepts (like a concept that represents an activity, e.g. Cause), situations
(time independent) or events (time dependent). These kind of plexuses is designated as
contexts (CTX) just to emphasize the difference between plexuses with and without
abstract meaning. Each context is defined by its name and its creator (instructor (user))
and is used to define affiliation of each association in ASM network, that is, its
relevancy. General context is defined and built in ASM structure independently of the
instructor (user), while other particular contexts are created by the user. All the
associations from particular contexts are assigned to the general one, but usually with
different parameters.

Implant-Design context
Free-Form
CAD-Model
{cpti=Implant,ri=concept,t=attributive,d=, {cpti=Implant,ri=sub-type,t=classifying,
c=+,h=0.75,s=0.75,rj=attribute,cptj=Free-Form} d=,c=+,h=1,s=1,rj=type,cptj=CAD-Model}

{cpti=Implant,ri=product,t=product-activity,d=,c=+, Implant
h=1,s=1,rj=activity,cptj=Implant-Design-Procedure}
{cpti=Implant,ri=assembly,t=affiliation,
d=,c=+,h=1,s=1,rj=part,cptj=Implant-Extension}
Implant-Design-
Procedure Implant-Extension

Fig. 2. ASM association structure: Several associations with specified parameters belonging to a
context

ASM structure is not domain-specific and can be used for knowledge representation
in diverse fields. Knowledge from specific domain should be represented through
context(s), while associations as semantic relations between contexts allow knowledge
from one context to be applicable to others.

4. Topologically Analogous Association Plexuses

The most common and probably the most significant case of semantic content similarity
between different association plexuses or contexts is called topological analogy
(similarity) (Fig. 3). Topologically analogous association plexuses or contexts have the
 Towards Analogy-Based Reasoning in Semantic Network 987

same type of topology (combination of appropriate values of topological parameters of

associations) and the same structure. Associations belonging to two different association
plexuses or contexts, with similar values of weight parameters and the same values of
topological parameters are called topologically correspondent associations (TCA)
(associations represented by the same type of line in Fig. 3). Concepts belonging to
TCA-s of two different association plexuses or contexts, which have the same role in
these TCA-s are called topologically correspondent concepts (TCC) (concepts
represented by the same background pattern in Fig. 3). Two types of topologically
analogous association plexuses or contexts are distinguished: semantically distant
(association plexuses or contexts do not share concepts, nor are their concepts similar,
synonyms or connected over series of up to four associations) and semantically close
(association plexuses or contexts share one or more concepts, or have concepts which
are similar, synonyms or connected over series of up to four associations).

CPT54
CPT25

R3 R3 R6
CPT5 CPT35 CPT33
R6
R6 R5
CPT4 CPT3
R6 R3
R2 R5 CPT31 PLXN
R1 R3 R1
PLXX R1 R4
CPT1 R2
R1 CPT32
R4 CPT51
R2 R2
CPT2
CPT34
CPT22

CPT21
CTXX CPT57
CTXN

Fig. 3. Association plexuses PLXX and PLXN are topologically analogous

4.1. Analogy and Similarities in ASM (Multi-level recognition of similarity)

In the core of its process of data semantics interpretation ASM employs specially
developed algorithms for recognizing topological similarity, i.e. analogy between
different parts of the semantic network. Analogy of semantics that can be recognized
between semantically (more or less) different concepts is essentially related to the
similarity between topology of subgraphs built by links (associations) of these concepts
in the semantic network. Depending on scope of focus in the process of topology
similarity recognizing, ASM uses two main algorithms for recognizing topological
similarity:
1. Contiguous, algorithm for determination of similarity between associations of two
concepts that are not directly connected, but over one layer of intermediate concepts
[50], and
988 Milos Stojkovic et al.

2. Wide, algorithm for determination of similarity between plexuses (subgraphs) of

associations (Determining the topological analogy between association plexuses
aimed for semantically distant concepts).
Actually, both algorithms are designed to determine degree and class of semantic
similarity, i.e. semantic correspondence between two concepts and both of them are
based on determination of similarity of topological parameters of associations. Weight
parameters are used to refine associations that have to be considered in the similarity
determination process by their semantic relevance regarding particular context,
instructor/user and his motivation. Contiguous algorithm is simpler and more explicit
(clearly defined) in similarity determination, but its application is limited to
semantically close concepts. Wide algorithm is aimed for determination of semantic
correspondence between two semantically distant concepts and for that it performs
determination of similarity between topology of different plexuses (subgraphs) of
associations.
To create inference autonomously, ASM employs algorithm for association plexus
upgrading which is executed in three attempts (see section 5). Actually, the paper is
mainly focused on describing this algorithm.
ASM encompasses additional two self-learning algorithms: 1) algorithms for creation
of heuristics, and 2) algorithm for knowledge crystallization (weakly structured
knowledge is crystallized into strongly structured knowledge, i.e. logic formalisms).
Both algorithms are aimed to provide ASM with capability to learn, i.e. to formalize
knowledge gained from experience with analogy-based reasoning and human
interventions. In relation to learning (reasoning in acquiring new knowledge)
capabilities of ASM, it should be mentioned that ASM is designed to perform data
semantic interpretation in regard to user motivation context as a reference framework.
Fulfillment of user actual motivation appears as the most important criteria for learning
analogies from experience. However, describing these algorithms is out of the scope of
this paper.
In ASM there are also CASE procedures (a sort of simplified set of predicate logic
rules) that can be regarded as generic logic formalism that will be employed if the case
of semantic graph is recognized to be suitable for triggering predefined logic rule. This
set of rules can change (i.e. in case learning process) enabling ASM to learn and
improve its reasoning performance over time. If case of unknown association plexus
(a sort of an ontology portion) is recognized as very similar or same as one from
Experience Set, then it is more efficient to apply strict logic formalism which is very
likely to be truthful for that kind of association plexus topology.

5. Association Plexus Upgrading Procedure

Every association plexus can be observed as a part of the semantic network connected to
other parts of the semantic network by associations involving other concepts. In general,
it is very difficult to distinguish where one association plexus ends, and where others
begin. User introduces new association plexus (which represents new or unknown
situation) to ASM, usually by creating associations between concepts of which some or
all are known to ASM, i.e. were added to ASM semantic network earlier (Fig. 4).
 Towards Analogy-Based Reasoning in Semantic Network 989

ASM responds to input by recognizing topological analogy between new and known
association plexuses (from the narrowed semantic network space) and upgrading new
association plexus modeled on remainder of the context (whose subset is recognized as
topologically analogous association plexus). The response is being formulated through
creating new associations between concepts from new association plexus and known
concepts in the network.
Association plexus upgrading procedure is based on similarity between new and
known association plexuses. New association plexus concepts will be connected
modeled on their TCC-s in similar association plexuses.

CPT3 ASM space

X1,3

CPT1
Associations between
PLXX X1,2 known concept CPT2
and other concepts
CPT2 from the ASM
semantic network

Fig. 4. Introducing new association plexus PLXX to ASM. Concept CPT2 is known to ASM

In the case when new association plexus PLXX is topologically analogous to certain
known association plexus PLXN (the more TCA-s they have, the better), regardless of
whether they are semantically close or semantically distant, ASM will use the logic of
topologically analogous association plexus upgrading (element denotes topological
correspondence (for associations and concepts) or topological analogy (for contexts and
association plexuses); element denotes association between concepts):
If

A PLX X
i, j A CPT
PLX N
k ,l i
PLX X
CPT k
PLX N
(1)
PLX X PLX N

where AiPLX
,j
X
ACPT
PLX X
i CPT j
, AkPLX
,l
N
ACPT
PLX N
k CPTl
, PLX N CTX N ,
then it is possible that there exists context CTXX, whose subset is new association plexus
PLXX, which is topologically analogous to known context CTXN:
CTX X CTX X PLX X CTX X CTX N (2)

Therefore, new association plexus PLXX should be upgraded to context CTXX,

modeled on the remainder of the known context CTXN (Fig. 5).
990 Milos Stojkovic et al.

PLXX PLXN
CPT3 CPT16
CPT13 N13,16
CPT6
R3
R3
X1,3
N11,13
CPT5 CPT15
N11,15
R1
R1
CPT1 N14,15
CPT11 N11,14
R2
CPT4 R2 CPT14
X1,2
PLXX R4 N11,12
N12,14
CPT2 R4
CTXX PLXN CPT12
CTXN

Fig. 5. Topologically analogous association plexus upgrading logic

Logic of topologically analogous association plexus upgrading is carried out through

three attempts (sub-procedures). First and second attempt have several iterations.
Each iteration for every attempt is followed by iteration of the process of determining
semantic similarity of concepts, which can also result in the creation of association(s)
between concepts. This procedure is presented in detail in [50].

5.1. First Attempt

The first attempt is carried out through several iterations. The procedure for each
iteration is identical. First attempt ends in situation when ASM is not able to add new
association to known association plexus.
The same example will be used independently to illustrate first attempt procedure for
semantically distant and semantically close TCC-s.

Semantically Distant TCC-s. ASM first recognizes semantically distant TCC-s of new
and known association plexus:

CPTi CPTi PLX X CPTj CPTj PLX N (3)

such that:
1. CPTi CPTj
CPTi CPT j , ACPT
CTX 0
i CPT j

CTX 0
2. t ACPTi CPT j

similarity

tA CTX 0
CPTi CPT j synonymous
ACPT
CTX N
j CPT j 1
A CTX
CTX N
CPT j CPT j 1 N
3.
A CTX N
CPT j CPT j 1 PLX , PLX CTX
N N N
 Towards Analogy-Based Reasoning in Semantic Network 991

If such concepts are found, ASM searches for all associations in the semantic
network involving concepts from new association plexus, which are topologically
correspondent to associations from known association plexus involving their TCC-s,
and adds these associations if TCC-s have the same roles in them (Fig. 6):

ACPT
CTX M
i CPTk

ACPT
CTX M
i CPTk

CTX M
CTX M
ACPTi CPTk
ACPT
CTX N
j CPT j 1
, r CPTi r CPT j ;

A CTX X
CPTi CPTk:tA tA
CTX X
i,k
CTX M
i,k (4)

c AiCTX
,k
X
c A ,d A d A
CTX M
i,k
CTX X
i,k
CTX M
i,k
h AiCTX
,k
X
h A , s A s A
CTX M
i,k
CTX X
i,k
CTX M
i,k
where AiCTX
,k
X
ACPT
CTX X
i CPTk
, AiCTX
,k
M
ACPT
CTX M
i CPTk
.

PLXX PLXN
CPT3 CPT16
CPT13 N13,16
CPT6
R3
R3
X1,3
N11,13
CPT5 CPT15
N11,15
R1
R1
CPT1 N14,15
CPT11 N11,14
R2
CPT4 R2 CPT14
X1,2
PLXX R4 N11,12
N12,14
CPT2 R4
CTXX PLXN CPT12
CTXN

Fig. 6. Association plexus upgrading in first attempt (semantically distant TCC-s). TCC-s (CPT1,
CPT11) and (CPT2, CPT12) are semantically distant

If several associations, involving concept CPT1 and topologically correspondent to

association between concepts CPT11 and CPT14, are found in the semantic network,
ASM analyzes if second concept in these associations is involved in the same or similar
associations as concept CPT14 (e.g. concept CPT4 is involved in association with
concept CPT15 or similar concept, which is topologically correspondent to association
between concepts CPT14 and CPT15). ASM finally adds only associations that meet this
condition.
In many situations during the upgrading procedure in first and second attempt,
structure of the known context CTXN has to be taken into account. One example is the
addition of the association between concepts CPT2 and CPT4. In this situation ASM
adds association between concept CPT2 and existing concept CPT4 (it is assumed that
association between concepts CPT1 and CPT4 was previously added) which is
topologically correspondent to association between concepts CPT12 and CPT14.
992 Milos Stojkovic et al.

Semantically Close TCC-s. Recognition of semantically distant TCC-s is followed by

the recognition of semantically close TCC-s of new and known association plexus. First
ASM recognizes TCC-s of new and known association plexus (or contexts of which
they are a subset) which are identical (denoted by ) or are synonyms (fifth class of
similarity (denoted by 5. ): absolute value of the difference of accuracy and
significance for all association pairs connecting these concepts have to be less than 0.25;
all association pairs connecting these concepts (through the same connectional
concepts) have to have the same type of associating (and the same corresponding
concept roles) and the same characters and directions of associating) or similar (fourth
class of similarity (denoted by 4. ): absolute value of the difference of accuracy and
significance for all association pairs connecting these concepts have to be less than 0.5;
all association pairs connecting these concepts have to have the same type of associating
(and the same corresponding concept roles) and the same characters and directions of
associating) in general context (semantically close TCC-s):

CPTi CPTi PLX X CPTj CPTj PLX N (5)

such that:
1. CPTi CPTj
5. 4.
2. CPTi CPTj CPTi CPTj CPTi CPTj in general context

3.
ACPT
CTX N
j CPT j 1

ACPT
CTX N
j CPT j 1

CTX N

A CTX N
CPT j CPT j 1 PLX N , PLX N CTX N

PLXX PLXN
CPT3 CPT16
CPT13 N13,16
CPT6
R3
R3
X1,3
N11,13
CPT5 CPT15
N11,15
R1
R1
CPT1 N14,15
CPT11 N11,14
R2
CPT14 R2 CPT14
X1,2
PLXX R4 N11,12
N12,14
CPT12 R4
CTXX PLXN CPT12
CTXN

Fig. 7. Association plexus upgrading in first attempt (semantically close TCC-s). TCC-s (concept
CPT12) of new and known association plexus are identical

If such concepts are found, ASM adds associations of known association plexus
involving found TCC-s, except that the concept from known association plexus will be
replaced by its TCC in new association plexus (Fig. 7):
 Towards Analogy-Based Reasoning in Semantic Network 993

A CTX X
CPTi CPT j 1 :tA tA
CTX X
i , j 1
CTX N
j , j 1

c AiCTX
, j 1
X

c ACTX CTX X
j , j 1 , d Ai , j 1
N
d ACTX N
j , j 1 (6)
h A h A , s A
CTX X
i , j 1
CTX N
j , j 1
CTX X
i , j 1 sA CTX N
j , j 1
, j 1 ACPTi CPT j 1 , Aj , j 1 ACPT j CPT j 1 .
CTX N CTX N
where AiCTX X CTX X

5.2. Second Attempt

The second attempt is carried out in several iterations. The procedure for each iteration
is identical. Second attempt ends in situation when ASM is not able to add new
association to known association plexus. Complete first attempt is carried out between
second attempt iterations. The second attempt will continue from the situation
illustrated in Fig. 6 (first attempt for semantically distant TCC-s).
ASM searches for concepts in the semantic network which are similar to concepts
from new association plexus in specific context, and are involved in associations which
are topologically correspondent to associations from known association plexus. It is
necessary to find the concepts in the semantic network which are similar to concepts
from new association plexus in at least third class of similarity (denoted by 3. ;
absolute value of the difference of accuracy and significance for all association pairs
connecting these concepts have to be less than 0.5; all association pairs connecting these
concepts have to have the same type of associating (and the same corresponding
concept roles) and the same characters of associating):

CPTi CPTi CTX X CPTj CPTj CTX K (7)

such that:
3.
1. CPTi CPT j

2.
ACPT
CTX K

j CPT j 1
ACPT
CTX K
j CPT j 1
CTX K
CTX K
ACPT j CPT j 1
ACPT
CTX N
k CPTk 1

If such concepts are found, ASM adds associations involving them, which are
topologically correspondent to associations from known association plexus, except that
found concept will be replaced by its similar concept in new association plexus (Fig. 8):

A CTX X
CPTi CPT j 1 :tA tA
CTX X
i , j 1
CTX K
j , j 1

c AiCTX
, j 1
X

c ACTX K CTX X
j , j 1 , d Ai , j 1 d ACTX K
j , j 1 (8)
h A h A , s A
CTX X
i , j 1
CTX K
j , j 1
CTX X
i , j 1 sA CTX K
j , j 1
, j 1 ACPTi CPT j 1 , Aj , j 1 ACPT j CPT j 1 .
where AiCTX X CTX X CTX K CTX K
994 Milos Stojkovic et al.

CPT52

AS R9
CPT1 CPT51
CPT56 CPT5
R10

CPT58 CTXK

PLXX PLXN
CPT3 CPT16
CPT13 N13,16
CPT6
R3
R3
X1,3
R10 N11,13 R10
CPT5 CPT15
N11,15
R1 R9 R11 R11
R1 R9
CPT1 N14,15
R14 CPT11 N11,14 R14
R2
CPT4 R2 CPT14
X1,2
PLXX R4 N11,12
N12,14
CPT2 R4
CTXX PLXN CPT12
CTXN

CPT65

CPT4 AS CPT63 R14

CPT67 CPT5
R11

CPT69 CTXL

Fig. 8. Association plexus upgrading in second attempt. Concepts CPT1 and CPT51 are similar in
context CTXK, while concepts CPT4 and CPT63 are similar in context CTXL

5.3. Third Attempt

Third attempt does not have iterations. After the third attempt is carried out, the user,
depending on whether he is satisfied with the results, decides whether to complete the
upgrading procedure or to carry it out from the beginning (from the first attempt).
The goal of the third attempt is to find candidate concepts in the semantic network
which should be connected with the remaining concepts (concept CPT3) from new
association plexus. Candidate concepts and their corresponding concepts (concept
CPT16) from known association plexus are usually semantically distant. Focus of the
third attempt is the similarity between associations involving candidate concepts and
associations involving their corresponding concepts from known association plexus.
In the third attempt ASM recognizes concepts (concept CPT16) involved in
associations from context whose subset is known association plexus, which do not have
TCC-s in the context whose subset is new association plexus. After that ASM identifies
all association plexuses with associations involving recognized concepts, as well as
their topologically analogous association plexuses. In the last step ASM identifies TCC-
s of the recognized concepts which are involved in the same or similar set of TCA-s in
recognized topologically analogous association plexuses (Fig. 9).
 Towards Analogy-Based Reasoning in Semantic Network 995

CPT111 CPT211 CPT523

PLX2 PLX1
PLX1 CPT6 PLX2 CPT16
CPT125 CPT522

CPT112 CPT212 CPT521 PLX5

PLX2 PLX9
PLX3 PLX5
CPT911
CPT323 CPT723

PLX9 CPT79 PLX3 PLX7

CPT16 CPT322 CPT6 CPT722

CPT912
CPT321 PLX3 CPT721 PLX7

PLX3 PLX8
CPT632
PLX6
CPT823

CPT631 CPT6 CPT633 CPT79 CPT822

CPT821 PLX8
PLX4 PLX6

CPT1032 CPT432
PLX10 PLX4

PLX4 PLX10

CPT1031 CPT125 CPT1033 CPT431 CPT16 CPT433

Fig. 9. Recognizing candidate concept(s) in the semantic network which should be connected
with the concept CPT3. Concepts CPT6 and CPT16 are TCC-s in most cases of identified
topologically analogous association plexuses

PLXX PLXN
CPT3 CPT16
CPT13 N13,16
CPT6
R3
R3
X1,3
N11,13
CPT5 CPT15
N11,15
R1
R1
CPT1 N14,15
CPT11 N11,14
R2
CPT4 R2 CPT14
X1,2
R4 N11,12
N12,14
PLXX CPT2 CTXX PLXN R4
CPT12 CTXN

Fig. 10. Association plexus upgrading in third attempt

If such TCC-s are found ASM adds associations between these concepts (concept
CPT6) and corresponding concepts (concept CPT3) from new association plexus which
will have the same parameters as associations from known association plexus
recognized at the beginning of the attempt (association between concepts CPT13 and
CPT16) (Fig. 10).
996 Milos Stojkovic et al.

6. Case

Association plexus upgrading procedure is demonstrated through solving one Ravens

Progressive Matrix which was Angelinas task (Fig. 1).
Ravens Progressive Matrices (RPM) is a multiple choice test of abstract reasoning
introduced by dr John C. Raven in 1936. They are often used as a test of so-called
general intelligence [52], which was also one of Ravens motives when constructing the
test. Each RPM problem is presented as a 2x2 or 3x3 matrix of pictures following a
pattern. Bottom right position in each matrix is left blank and solvers task is to choose
the missing picture for that cell from provided list with eight possible solutions. The
first and most common set of RPM are the Standard Progressive Matrices (SPM),
consisting of 60 matrices, developed in 1936 [53] and published in 1938.
Why Ravens Progressive Matrices?: The RPM is a type of problem which is
familiar to almost everyone, since most people faced an IQ test and it is very clear what
the task is. Describing data semantic interpretation process is already too complex task
by itself and would be even more difficult to understand if we try to introduce it through
ASM cognitive process applied upon some domain specific problem (e.g. choosing the
most suitable CAD procedure for reverse modeling of a free-form shape like sternum
i.e. chest bone, which actually was our real task to solve). On the other side, introducing
new research product (such as ASM) usually needs more space and details, which is
always an issue.
In addition, RPM can be very useful and simple example for comparing different
approaches in describing and interpreting semantics of figures presented in RPM.
Considering the universality, RPM seems as very appropriate domain to compare
learning capabilities of different not domain-specific knowledge representation models.
Finally, within this kind of domain different reasoning engines can be compared by how
many elementary activities in preparation i.e. customization of reasoner an instructor
should do to make it capable for relevant entailments.
As of today the cognitive and computational characteristics of RPM arent yet well
understood [54] and no general algorithm for solving them in their entirety has been
developed.
In our approach the missing picture (solution) is not chosen from the provided
alternatives, but rather built (generated), making the task much harder than in
conventional RPM solving.
Two pictures in the first row of the matrix presented in Fig. 1 are semantically
described in the context CTX1 (Fig. 11).
One picture in the second row of the matrix presented in Fig. 1 is semantically
described in the context CTX2 (Fig. 12).
 Towards Analogy-Based Reasoning in Semantic Network 997

CTX1
assembly Figure 1 assembly

predecessor

assembly
Figure 2 successor
part
assembly assembly
Empty-Triangle subject
Left determinant
assembly compound concept

part determinant
Empty-Triangle- Empty
compound concept
Lower-Left-Corner
determinant
object activity Comprise
compound concept part
Empty-Circle object
Lower determinant
subject subtype
type
part Circle
Empty-Triangle- object type
compound concept
Lower-Right-Corner subtype object
subject Filled-Circle
part
activity
Right determinant compound concept
Occupy
determinant Filled

Fig. 11. Semantic description of the two pictures (Figure 1 (left) and Figure 2 (right)) in the
first row of the matrix. In first picture we have empty triangle which comprises filled circle, while
filled circle occupies empty triangles lower right corner. In second we have empty triangle which
comprises empty circle, while empty circle occupies empty triangles lower left corner

CTX2
assembly assembly
Figure 3

predecessor
part
Empty-Rectangle subject
Figure 4 successor
assembly compound concept

determinant
Empty activity Comprise
Lower determinant

part Triangle
Empty-Rectangle- object type
compound concept
Lower-Right-Corner subtype object
subject Filled-Triangle
part
Right activity
determinant compound concept
Occupy
determinant Filled

Fig. 12. Semantic description of the picture (Figure 3) in the second row of the matrix. Figure
4 will eventually be the missing picture (solution)
998 Milos Stojkovic et al.

General context contains, among other things, knowledge about the geometric shapes
from the pictures of the matrix (Fig. 13).

determinant Empty-Rectangle- determinant

Left Lower
compound Lower-Left-Corner compound
determinant concept concept determinant
part
compound compound
concept concept
assembly
Empty-Rectangle- assembly part Empty-Rectangle-
Empty-Rectangle
Upper-Left-Corner part assembly Lower-Right-Corner
compound assembly
compound compound
concept
concept concept
part
determinant compound compound determinant
concept Empty-Rectangle- concept
Upper Right
determinant Upper-Right-Corner determinant

determinant
Empty Filled
determinant determinant
Triangle
compound concept compound concept
type
Empty-Triangle subtype subtype Filled-Triangle
General Context

Fig. 13. Knowledge about the geometric shapes from the pictures of the matrix

Association plexuses representing knowledge about the picture in the second row of
the matrix and the first picture in the first row of the matrix are topologically analogous
and semantically close (Fig. 14). TCA of these two association plexuses are represented
by the same type of line, while TCC are represented by the same background pattern.
ASM tries to upgrade new association plexus (representing knowledge about one
picture in the second row of the matrix) through several iterations. In the first iteration
ASM recognizes semantically distant TCC-s of two association plexuses: 1) Empty-
Rectangle and Empty-Triangle, and 2) Triangle and Circle. ASM searches for
associations in the semantic network involving concept Empty-Rectangle which are
topologically correspondent to association between concepts Empty-Triangle and
Empty-Triangle-Lower-Left-Corner. If several of them are found (e.g. in General
context), ASM selects the one whose second concept is involved in the same or similar
associations as concept Empty-Triangle-Lower-Left-Corner (Fig. 15 up). The same
approach is used (applied) for concept Triangle (Fig. 15 up). The next step is
upgrading of new association plexus through creating new associations: 1) between
concepts Empty-Rectangle and Empty-Rectangle-Lower-Left-Corner, and 2)
between concepts Triangle and Empty-Triangle (Fig. 15 down).
 Towards Analogy-Based Reasoning in Semantic Network 999

CTX2

assembly assembly
Figure 3

predecessor
part
Empty-Rectangle subject
Figure 4 successor
assembly compound concept

determinant
activity Comprise
Empty
Lower determinant

part Triangle
Empty-Rectangle- object type
compound concept
Lower-Right-Corner subtype object
subject Filled-Triangle
part
Right activity
determinant compound concept
Occupy
determinant Filled

CTX1

assembly Figure 1 assembly

predecessor

assembly
Figure 2 successor
part
assembly assembly
Empty-Triangle subject
Left determinant
assembly compound concept

part determinant
Empty-Triangle- Empty
compound concept
Lower-Left-Corner
determinant Comprise
activity
object
compound concept part
Empty-Circle object
Lower determinant
subject subtype
type
part Circle
Empty-Triangle- object type
compound concept
Lower-Right-Corner subtype object
subject Filled-Circle
part
activity
Right determinant compound concept
Occupy
determinant Filled

Fig. 14. Recognized topologically analogous and semantically close association plexuses
Subsets of CTX1 and CTX2 contexts. TCA are represented by the same type of line, while TCC are
represented by the same background pattern. Associations from the context CTX1 which dont
belong to recognized topologically analogous association plexus are grayed out
1000 Milos Stojkovic et al.

CTX1 CTX1
assembly Empty-Triangle
determinant Empty
Left
determinant
part compound concept
Empty-Triangle- Empty-Circle
compound concept
Lower-Left-Corner
subtype
type
Circle
Lower determinant

General Context General Context

Empty
determinant Empty-Rectangle- determinant
Left Lower
compound Lower-Left-Corner compound determinant
Triangle
concept part concept
type
assembly compound concept

Empty-Rectangle Empty-Triangle subtype

CTX2

assembly Figure 3 assembly

predecessor

Figure 4 successor
part
assembly Empty-Rectangle subject

assembly compound concept

part determinant
Empty-Rectangle- Empty
Lower-Left-Corner
activity Comprise

Empty-Triangle
Lower determinant
subtype
type
part Triangle
Empty-Rectangle- object type
compound concept
Lower-Right-Corner
subtype object
subject Filled-Triangle
part
activity
Right determinant compound concept
Occupy
determinant Filled

Fig. 15. First iteration of upgrading new association plexus Creation of new associations: 1)
between concepts Empty-Rectangle and Empty-Rectangle-Lower-Left-Corner, and 2)
between concepts Triangle and Empty-Triangle (down). Explanation for choosing the
appropriate associations (up)

In second iteration ASM recognizes semantically close TCC-s of two association

plexuses (concepts Lower, Empty, Occupy, and Comprise). Instead of adding
associations of context CTX1 involving found semantically close TCC-s (like
 Towards Analogy-Based Reasoning in Semantic Network 1001

association between concepts Lower and Empty-Triangle-Lower-Left-Corner),

ASM analyzes the structure of context CTX1, and creates appropriate associations
between existing concepts (association between concepts Lower and Empty-
Rectangle-Lower-Left-Corner) (Fig. 16 up). The same approach is used (applied) for
other recognized semantically close TCC-s. ASM upgrades new association plexus
through creating five new associations (Fig. 16 down).

assembly Empty-Rectangle CTX2

assembly Empty-Rectangle
part
Empty-Rectangle- part
Lower-Left-Corner
compound concept Empty-Rectangle-
Lower-Left-Corner
Lower determinant

Empty-Triangle- Lower determinant

compound concept Lower-Left-Corner

CTX2

assembly Figure 3 assembly

predecessor

Figure 4 successor
part
assembly Empty-Rectangle subject

assembly compound concept

part determinant
Empty-Rectangle- Empty
compound concept
Lower-Left-Corner
determinant Comprise
activity
object
compound concept
Empty-Triangle object
Lower determinant
subject subtype
type
part Triangle
Empty-Rectangle- object type
compound concept
Lower-Right-Corner
subtype object
subject Filled-Triangle
part
activity
Right determinant compound concept
Occupy
determinant Filled

Fig. 16. Second attempt of upgrading new association plexus Creation of five new associations
(down). Decisions are made based on the analysis of the structure of context CTX1 (up)

In third iteration ASM recognizes semantically distant TCC-s of two association

plexuses: 1) Empty-Rectangle-Lower-Left-Corner and Empty-Triangle-Lower-Left-
1002 Milos Stojkovic et al.

Corner, 2) Empty-Rectangle and Empty-Triangle, and 3) Empty-Triangle and

Empty-Circle. Association between concepts Empty-Rectangle-Lower-Left-Corner
and Left was found in General context, and is topologically correspondent to
association between concepts Empty-Triangle-Lower-Left-Corner and Left. As for
the remaining two TCC-s ASM analyzes the structure of context CTX1, and creates two
appropriate associations between existing concepts in context CTX2. ASM finally
upgrades new association plexus through creating three new associations (Fig. 17).

CTX2

assembly Figure 3 assembly

predecessor

assembly
Figure 4 successor
part
assembly assembly
Empty-Rectangle subject
Left determinant
assembly compound concept

part determinant
Empty-Rectangle- Empty
compound concept
Lower-Left-Corner
determinant Comprise
activity
object
compound concept part
Empty-Triangle object
Lower determinant
subject subtype
type
part Triangle
Empty-Rectangle- object type
compound concept
Lower-Right-Corner
subtype object
subject Filled-Triangle
part
activity
Right determinant compound concept
Occupy
determinant Filled

Fig. 17. New association plexus is finally upgraded through creating three new associations

Experimental evaluation of upgrading procedure was done for the following cases:
product quality assessment in the early stages of product design [55]; automation of
choosing and composing manufacturing process for free-form design parts [56];
exception detection in business process management systems [57]. Presented approach
is also being evaluated in the area of digital reconstruction of free-form objects.

7. Implementation and Evaluation

AcSeMod web application, implementing ASM structure and accompanying cognitive

data processing algorithms, has been developed for testing purposes and dissemination.
The web application was developed using Java programming language. Apache Tomcat
v6 was used as application server. Associations and other elements of the ASM
 Towards Analogy-Based Reasoning in Semantic Network 1003

structure are stored using MySQL Community Server v5 relational database

management system. Presented association plexus upgrading procedure is one of the
cognitive data processing algorithms, and is implemented on database level through
stored procedures and views.
Performance of the association plexus upgrading procedure was assessed on desktop
computer with Windows 7 Enterprise operating system and the following hardware
specifications: Intel Core i5 quad-core processor with 3.0 GHz clock speed and 6 MB of
cache memory; 8 GB DDR3 RAM memory; hard disk with 1 TB capacity and 32 MB
of cache memory. As for the software specifications, MySQL Community Server
v5.5.27 relational database management system, Apache Tomcat v6.0.37 application
server, and Google Chrome v34.0.1847.131 m web browser were used. Semantic
network contained seven contexts with 98 associations. All three iterations were
finished in three or less seconds.

8. Conclusion

As it is shown, ASM brings original approach in realization of ABR in semantic

network. The core of the ABR process and semantic interpretation of data is performed
by ASMs association plexus upgrading procedure which is based on recognition and
determining the similarity of association plexuses. Determining the similarity of
association plexuses is performed by the recognition of topological analogy between
association plexuses. Relaying on this approach ASM responds to an unpredicted input,
which is defined through input association plexuses, by upgrading that association
plexus modeled on remainder of the context whose subset is recognized as topologically
analogous association plexus. ABR process designed in this way enables autonomous,
flexible and analytic semantic interpretation of data described in the semantic network.
Limitations: First users involved in testing or incorporating knowledge can be
discouraged by difficulty to understand ASM operating mode. The degree of
meaningfulness of ASM responses depends to a large extent on the way in which user
described the request for semantic interpretation. Request description is currently the
biggest functional problem of ASM. Of course, before describing request some
knowledge should be incorporated in ASM. Testing showed that for some successful,
meaningful conclusions ASM needed only small portions of knowledge in the network.
This can be considered as advantage (like with the case presented in this paper).
Another weakness of ASM is the imprecision (inaccuracy) of the responses. This is
also the problem of all approaches which base semantic interpretation of data on
analogies. Related to this problem is also the problem of imaginative period of work
in ASM. In situations when there is little knowledge incorporated in the network, ASM
responses can be characterized by imagination, and ASM can relate concepts that
cant be related.
Since responses of ASM are based on knowledge previously incorporated in the
network, there is a risk of indoctrination. Incorrect indoctrination of ASM, leading to
creation (generation) of incorrect conclusions, can appear when multiple users are
incorporating their knowledge about the same domain. In this case there is a risk of
making semantic content inconsistent (if their interpretations are extremely
inconsistent).
1004 Milos Stojkovic et al.

Future work on ASM could include testing and adjustment of data structure
(semantic network) and set of algorithms for processing of data. Development of
intuitive interface, which will enable other software applications from various fields to
connect with ASM could be second direction of ASM development. As for the
enhancement of ASM functionality there is a need for developing structural elements
for incorporation and semantic categorization of events, i.e. contexts which string one
after another in the time sequence of discrete time instants.
The ability to recognize analogy between semantically very distant situations is
considered as one of the essential characteristics of creativity. Creative conclusions
usually start by recognizing similarity between apparently semantically disconnected
elements and arise by creating new semantic relations between these elements or ideas.
According to another stand [25], creative conclusions arise by creating new context-
suitable semantic relations between elements or ideas which are already connected by
some old semantic relations, which are not applicable for the actual context. In ASM
topologically correspondent associations from completely semantically distant contexts
can be used for drawing conclusions. In this way knowledge from one context can be
applied in situations which belong to other completely different contexts enabling ASM
to demonstrate creativity. Associations between the same concepts, belonging to
different contexts (and having different parameters), participate in the decision making
process in a completely different way, depending on the context they belong to, which
makes ASM more flexible and productive in capturing and interpreting semantics of
data compared to existing semantic models [49].
One can see ASM as a kind of layer above the DL ontologies layer (strongly
structured knowledge, i.e. richly axiomatized discourse) which helps semantic
interpretation.

Acknowledgment. The paper represents a summary about a part of the research that is conducted
within the project Virtual human osteoarticular system and its application in preclinical and
clinical practice (project id III 41017) which is funded by the Ministry of Education, Science and
Technological Development of the Republic of Serbia for the period 2011-2014.

References

1. Daconta, M.C., Obrst, L.J., Smith, K.T.: The Semantic Web: A Guide to the Future of XML,
Web Services, and Knowledge Management. Wiley Publishing, Inc., Indianapolis, USA.
(2003)
2. Forbus, K.D., Mostek, T., Ferguson, R.: An Analogy Ontology for Integrating Analogical
Processing and First-principles Reasoning. In Proceedings of the 14th Conference on
Innovative Applications of AI, IAAI-02. AAAI Press, Menlo Park, California, USA. (2002)
3. Corcho, O., Gomez-Perez, A.: Evaluating Knowledge Representation and Reasoning
Capabilities of Ontology Specification Languages. In Proceedings of the Workshop on
Applications of Ontologies and Problem Solving Methods, 14th European Conference on
Artificial Intelligence, ECAI'00. IOS Press, Amsterdam, The Netherlands. (2000)
4. Dentler, K., Cornet, R., ten Teije, A., de Keizer, N.: Comparison of reasoners for large
ontologies in the OWL 2 EL profile. Semantic Web Interoperability, Usability,
Applicability, Vol. 2, No. 2, 71-87. (2011)
5. Pan, Z.: Benchmarking DL Reasoners Using Realistic Ontologies. In Proceedings of the 1st
OWL: Experiences and Directions Workshop. Galway, Ireland. (2005)
 Towards Analogy-Based Reasoning in Semantic Network 1005

6. W3C Recommendation RIF Production Rule Dialect (Second Edition) (2013). Available:
http://www.w3.org/TR/2013/REC-rif-prd-20130205/ (current February 2013)
7. Nikitina, N., Rudolph, S., Blohm, S.: Refining Ontologies by Pattern-Based Completion. In
Proceedings of the Workshop on Ontology Patterns, WOP 2009. Washington, DC, USA.
(2009)
8. Simperl, E., Burger, T., Hangl, H., Worgl, S., Popov, I.: ONTOCOM: A reliable cost
estimation method for ontology development projects. Web Semantics: Science, Services and
Agents on the World Wide Web, Vol. 16, 1-16. (2012)
9. Hitzler, P., van Harmelen, F.: A Reasonable Semantic Web. Semantic Web
Interoperability, Usability, Applicability, Vol. 1, No. 1-2, 39-44. (2010)
10. Horrocks, I.: Description Logic Reasoning (tutorial). 11th International Conference on Logic
for Programming Artificial Intelligence and Reasoning, LPAR 2004. Montevideo, Uruguay.
Available: http://www.cs.ox.ac.uk/people/ian.horrocks/Seminars/download/lpar05.ppt (2005)
11. Lehmann, J., Volker, J.: An Introduction to Ontology Learning. In: Lehmann, J., Volker, J.
(eds.): Perspectives on Ontology Learning. IOS Press, Amsterdam, The Netherlands. (2014)
12. Shamsfard, M., Barforoush, A.A.: The state of the art in ontology learning: a framework for
comparison. The Knowledge Engineering Review, Vol. 18, No. 4, 293-316. (2003)
13. Hazman, M., El-Beltagy, S.R., Rafea, A.: A Survey of Ontology Learning Approaches.
International Journal of Computer Applications, Vol. 22, No. 9, 36-43. (2011)
14. Arnold, P., Rahm, E.: Enriching ontology mappings with semantic relations. Data &
Knowledge Engineering, Vol. 93, 1-18. (2014)
15. Lu, Y., Panetto, H., Ni, Y., Gu, X.: Ontology alignment for networked enterprise information
system interoperability in supply chain environment. International Journal of Computer
Integrated Manufacturing, Vol. 25, No. 12, 1148-1159. (2012)
16. Petrakis, E.G.M., Varelas, G., Hliaoutakis, A., Raftopoulou, P.: Design and Evaluation of
Semantic Similarity Measures for Concepts Stemming from the Same or Different
Ontologies. In Proceedings of the 4th Workshop on Multimedia Semantics, WMS'06.
Chania, Crete, Greece. (2006)
17. Rahm, E., Arnold, P., Raunich, S.: Semantic ontology mappings: How to determine and use
them. Colloquium Talk, University of Paris-Sud. Available: http://dbs.uni-
leipzig.de/file/paris-Octob2014.pdf (2014)
18. Zdravkovic, M., Panetto, H., Trajanovic, M., Aubry, A.: An approach for formalising the
supply chain operations. Enterprise Information Systems, Vol. 5, No. 4, 401-421. (2011)
19. Borgida, A., Walsh, T.J., Hirsh, H.: Towards Measuring Similarity in Description Logics. In
Proceedings of the 2005 International Workshop on Description Logics, DL2005. Edinburgh,
Scotland, UK. (2005)
20. Kim, K.-Y., Manley, D.G., Yang, H.: Ontology-based assembly design and information
sharing for collaborative product development. Computer-Aided Design, Vol. 38, No. 12,
1233-1250. (2006)
21. dAmato, C., Fanizzi, N., Esposito, F.: Reasoning by Analogy in Description Logics through
Instance-based Learning. In Proceedings of the 3rd Italian Semantic Web Workshop Scuola
Normale Superiore. Pisa, Italy. (2006)
22. Raad, E., Evermann, J.: Is Ontology Alignment like Analogy? Knowledge Integration
with LISA. In Proceedings of the 29th Annual ACM Symposium on Applied Computing,
SAC14. Gyeongju, Korea. (2014)
23. Raad, E., Evermann, J.: The role of analogy in ontology alignment: A study on LISA.
Cognitive Systems Research, Vol. 33, 1-16. (2015)
24. Janowicz, K., Raubal, M., Kuhn, W.: The semantics of similarity in geographic information
retrieval. Journal of Spatial Information Science, No. 2, 29-57. (2011)
25. Kosti, A.: Kognitivna psihologija. Zavod za udbenike i nastavna sredstva, Beograd, Srbija.
(2006)
26. Gentner, D., Holyoak, K.J., Kokinov, B.N.: The Analogical Mind: Perspectives from
Cognitive Science. The MIT Press, Cambridge, USA. (2001)
1006 Milos Stojkovic et al.

27. Forbus, K.D.: Exploring Analogy in the Large. In: Gentner, D., Holyoak, K.J., Kokinov,
B.N. (eds.): The Analogical Mind: Perspectives from Cognitive Science. The MIT Press,
Cambridge, USA, 23-58. (2001)
28. Newel, A., Simon, H.A.: Human Problem Solving. Prentice Hall, New Jersey, USA. (1972)
29. Aamodt, A., Plaza, E.: Case-Based Reasoning: Foundational Issues, Methodological
Variations, and System Approaches. AI Communications, Vol. 7, No. 1, 39-59. (1994)
30. Davies, J., Goel, A.K., Yaner, P.W.: Proteus: Visuospatial analogy in problem solving.
Knowledge-Based Systems, Vol. 21, No. 7, 636-654. (2008)
31. Bergman, R., Kolodner, J., Plaza, E.: Representation in case-based reasoning. The
Knowledge Engineering Review, Vol. 20, No. 3, 209-213. (2006)
32. Holyoak, K.J.: Analogical thinking and human intelligence. In: Sternberg, R.J. (ed.):
Advances in the psychology of human intelligence. Erlbaum, Hillsdale, NJ, USA. (1984)
33. Chang, P.-C., Lin, J.-J., Dzan, W.-Y.: Forecasting of manufacturing cost in mobile phone
products by case-based reasoning and artificial neural network models. Journal of Intelligent
Manufacturing, Vol. 23, No. 3, 517-531. (2012)
34. Begum, S., Ahmed, M.U., Funk, P., Xiong, N., Scheele, B.V.: A case-based decision support
system for individual stress diagnosis using fuzzy similarity matching. Computational
Intelligence, Vol. 25, No. 3, 180-195. (2009)
35. Passone, S., Chung, P.W.H., Nassehi, V.: Incorporating domain-specific knowledge into a
genetic algorithm to implement case-based reasoning adaptation. Knowledge-Based Systems,
Vol. 19, No. 3, 192-201. (2006)
36. Forbus, K.D., Gentner, D., Law, K.: MAC/FAC: A Model of Similarity-based Retrieval.
Cognitive Science, Vol. 19, No. 2, 141-205. (1995)
37. Holyoak, K.J., Koh, K.: Surface and structural similarity in analogical transfer. Memory &
Cognition, Vol. 15, No. 4, 332-340. (1987)
38. Negny, S., Riesco, H., Le Lann, J.M.: Effective retrieval and new indexing method for case
based reasoning: Application in chemical process design. Engineering Applications of
Artificial Intelligence, Vol. 23, No. 6, 880-894. (2010)
39. Smyth, B., Keane, M.T.: Adaptation-guided retrieval: questioning the similarity assumption
in reasoning. Artificial Intelligence, Vol. 102, No. 2, 249-293. (1998)
40. Foxwell, H.J., Menasce, D.A.: MARVIN: A Web-Based System for Representing,
Retrieving, and Visualizing Analogies. World Wide Web: Internet and Web Information
Systems, Vol. 7, No. 4, 385-419. (2004)
41. Hall, R.P.: Computational Approaches to Analogical Reasoning: A Comparative Analysis.
Artificial Intelligence, Vol. 39, No. 1, 39-120. (1989)
42. Kedar-Cabelli, S.: Analogy - From a Unified Perspective. In Helman, D.H. (ed.): Analogical
Reasoning: Perspectives of Artificial Intelligence, Cognitive Science, and Philosophy, Vol.
197. Kluwer Academic Publishers, Dordrecht, Netherlands, 65-103. (1988)
43. Kokinov, B.: Analogy is like Cognition: Dynamic, Emergent, and Context-Sensitive. In
Holyoak, K., Gentner, D., Kokinov, B. (eds.): Advances in Analogy Research: Integration of
Theory and Data from the Cognitive, Comptational, and Neural Sciences. New Bulgarian
University Press, Sofia, Bulgaria, 96-105. (1998)
44. Kokinov, B.N., Petrov, A.A.: Integration of Memory and Reasoning in Analogy-Making:
The AMBR Model. In: Gentner, D., Holyoak, K.J., Kokinov, B. (eds.): The analogical mind:
Perspectives from Cognitive Science. MIT Press, Cambridge, MA, USA. (2000)
45. Hummel, J.E., Holyoak, K.J.: LISA: A Computational Model of Analogical Inference and
Schema Induction. In Proceedings of the 18th Annual Conference of the Cognitive Science
Society. Lawrence Erlbaum Associates, Mahwah, NJ, USA. (1996)
46. Falkenhainer, B., Forbus, K.D., Gentner, D.: The Structure-Mapping Engine: Algorithm and
Examples. Artificial Intelligence, Vol. 41, No. 1, 1-63. (1989)
47. Holyoak, K.J., Thagard, P.: Analogical Mapping by Constraint Satisfaction. Cognitive
Science, Vol. 13, No. 3, 295-355. (1989)
 Towards Analogy-Based Reasoning in Semantic Network 1007

48. Stojkovi, M.: Analysis of the Manufacturability Parameters Based on Semantic Structures
of the Digital Product Model. Ph.D. Thesis, University of Ni, Faculty of Mechanical
Engineering in Ni, Ni, Serbia. (2011)
49. Stojkovi, M., Mani, M., Trifunovi, M., Mii, D.: Semantic categorization of data by
determining the similarities of associations of the semantic network. E-Society Journal
Research and Applications, Vol. 2, No. 1, 3-13. (2011)
50. Trifunovi, M., Stojkovi, M., Trajanovi, M., Mii, D., Mani, M.: Interpreting the
meaning of geometric features based on the similarities between associations of semantic
network. FACTA UNIVERSITATIS Series: Mechanical Engineering, Vol. 11, No. 2, 181-
192. (2013)
51. Anderson, J.R., Bower, G. H.: Human Associative Memory. Winston, Washington, USA.
(1973)
52. Snow, R., Kyllonen, P., Marshalek, B.: The topography of ability and learning correlations.
In: Sternberg, R.J. (ed.): Advances in the Psychology of Human Intelligence, Vol. 2.
Lawrence Erlbaum Associates, Hillsdale, New Jersey, USA, 47-103. (1984)
53. Raven, J.C.: Mental tests used in genetic studies: The performances of related individuals in
tests mainly educative and mainly reproductive. Masters Thesis, University of London.
(1936)
54. Kunda, M., McGreggor, K., Goel, A.: Addressing the Ravens Progressive Matrices Test of
General Intelligence. In Proceedings of the AAAI Fall Symposium on Multi-
representational Architectures for Human-level Intelligence. AAAI Press, Menlo Park,
California, USA. (2009)
55. Stojkovic, M., Manic, M., Trifunovic, M., Vitkovic, N.: Semantic interpretation of the
product model features in product quality assessment. In Proceedings of the 6th International
Working Conference Total Quality Management Advanced and Intelligent Approaches.
481-484. (2011)
56. Manic, M., Stojkovic, M., Trifunovic, M.: Semantic features in computer aided
manufacturing systems. In Proceedings of the International Conference Mechanical
Engineering in XXI Century. Faculty of Mechanical Engineering in Ni, Ni, Serbia, 179-
182. (2010)
57. Misic, D., Stojkovic, M., Domazet, D., Trajanovic, M., Manic, M., Trifunovic, M.: Exception
detection in business process management systems. Journal of Scientific and Industrial
Research, Vol. 69, No. 3, 188-193. (2010)

Milos Stojkovic is assistant professor in the Department for Production, IT and

Management at the Faculty of Mechanical Engineering in Nis, University of Nis. He
graduated with a degree in mechanical engineering from the same University in 1996
and received his Ph.D. in 2011. His areas of research are knowledge-based engineering
systems (R&D in semantic networks and cognitive data processing models for
CAD/CAM systems) and bioengineering (R&D in customized tissue scaffold design
and fabrication and R&D in CAD methods for reverse modeling of the human bones).
He is the author of more than 40 international scientific publications, 15 of them in
international journals.

Milan Trifunovic is teaching and research assistant in the Department for Production,
IT and Management at the Faculty of Mechanical Engineering in Nis, University of Nis.
He graduated with a degree in mechanical engineering from the same University in
2003. His areas of research are application of information technologies in
manufacturing, artificial intelligence and knowledge-based engineering systems
1008 Milos Stojkovic et al.

(semantic networks and cognitive data processing models for CAD/CAM systems). He
is the author of approximately 14 international scientific publications, 3 of them in
international journals.

Dragan Misic is assistant professor in the Department for Production, IT and

Management at the Faculty of Mechanical Engineering in Nis, University of Nis. He
graduated with a degree in mechanical engineering from the same University in 1991
and received his Ph.D. in 2010. His areas of research are application of information
technologies in manufacturing, simultaneous design of technological processes and
integration of CAD/CAPP/CAMM systems. He is the author of approximately 21
international scientific publications, 5 of them in international journals.

Miodrag Manic is full professor in the Department for Production, IT and Management
at the Faculty of Mechanical Engineering in Nis, University of Nis. He graduated with a
degree in mechanical engineering from the same University in 1980 and received his
Ph.D. in 1995. His areas of research are application of information technologies,
including methods of artificial intelligence, computer numerically controlled machine
tools and their programming and technology and manufacturing systems. He is the
author of approximately 35 international scientific publications, 9 of them in
international journals.

Received: November 3, 2014; Accepted: April 29, 2015.