SEMANTIC NETWORK BASED MECHANISMS FOR KNOWLEDGE ACQUISITION

Sundarapandian et al. (Eds) : CSE, CICS, DBDM, AIFL, SCOM - 2013
pp. 105–118, 2013. © CS & IT-CSCP 2013 DOI : 10.5121/csit.2013.3311
SEMANTIC NETWORK BASED MECHANISMS
FOR KNOWLEDGE ACQUISITION
Dariusz Ceglarek
Department of Applied Informatics, Poznan School of Banking, Poznan, Poland
dariusz.ceglarek@wsb.poznan.pl
ABSTRACT
This article summarizes research work started with the SeiPro2S (Semantically Enhanced
Intellectual Property Protection System) system designed to protect resources from the
unauthorized use of intellectual property. The system implements semantic network as a structure
of knowledge representation and a new idea of semantic compression. As the author proved that
semantic compression is viable concept for English, he decided to focus on potential applications.
An algorithm is presented that employing semantic network WiSENet for knowledge acquisition
with flexible rules that yield high precision results. Developed algorithm is implemented as a Finite
State Automaton with advanced methods for triggering desired actions. Detailed discussion is
given with description of devised algorithm, usage examples and results of experiments.
KEYWORDS
Semantic network, Semantic compression, Knowledge acquisition, Lexical relationships,
Natural Language Processing
1. INTRODUCTION
Natural language is a very complex system which needs to be represented in a way that would be
understandable for computer systems. One need to possess some structures that can represent a part of
semantic knowledge. Choosing proper knowledge representation structure is very important
determinant of the classification quality of text documents [3][11][15]. Semantic knowledge, as
identified lexical relations between concepts, should be stored in an appropriate data structure in
order to be utilized to refine Information Retrieval (IR) or Natural Language Processing (NLP)
tasks and their results. a semantic network is a structure incorporating knowledge about all possible
lexical relations between words. Lexical relations reflect the interdependences between the
concepts. Semantic networks store information about similarity relations (like a thesaurus): word
similarity, synonymy, antonymy; hierarchical relations (like a taxonomy): hypernymy, troponymy
(lexical relations existing only for verbs) or hyponymy and meronymy or holonymy relations.
Semantic network can incorporate connotations as well these are any other word associations. Using
the graph theory terminology, semantic networks can be represented as directed graphs. Direction is
crucial in case of hierarchical relations. Edges between concepts can be weighted as well in order to
reflect strength of a relation. Semantic networks are the most advanced structures representing
semantic knowledge of natural language [26]. Choosing proper knowledge representation structure is
very important determinant of the classification quality of text documents. That is why their utilization
in information retrieval systems should bring the biggest improvement in their effectiveness.

106 Computer Science & Information Technology (CS & IT)
This paper shows that improving the efficiency of NLP methods by utilizing sophisticated models
grooves on inclusion of mechanisms reflecting and using information of lexical relations between
concepts. Text documents are most often a subject of information retrieval or information
extraction. The complexity of human natural language, however, negatively affects the results of
classic algorithms implemented in NLP/IR systems. Retrieval methods are getting accommodated
to identify lexical relations between words or phrases (i.e. relations between the meanings of words
or the concepts they represent) and use this knowledge to compare and match documents with
user’s needs more accurately. Such knowledge is represented in structures like thesauri or semantic
networks. Thesauri or semantic networks can be created manually Unfortunately, it is a very
time-consuming task and needs an involvement of expert knowledge. This paper presents a method
developed to automatically extract mentioned knowledge from a corpus of documents.
The information included in semantic network can be used in order to limit the number of keywords
to describe a document, expand user queries or identify concepts if a word represents more than one
meanings. Its greatest advantage is by supplying a system with the right meaning of the concept
processed based on its contextual usage [4]. Benefits one can obtain by applying semantic nets in
classification tasks were described by [1]. Commonly used semantic network in NLP systems for
processing English is WordNet [10][20]. Its structure is organized around notion of synsets. Every
WordNet’s synset contains words which are mutually synonyms. Relationships among synsets are
hypernyms or hyponyms, when combined with previous data it is easily seen that whole WordNet acts
as a thesaurus. The details of the adoption and motivation of transferring WordNet to a new format
WiSeNet is discussed in [7]. In this paper were also enumerated various aspects and possible merits of
applying the WiSENet semantic network.
The aim of this work is to present an application of previously introduced semantic network
WiSENet (semantic network WordNet transferred into SenecaNet format introduced in [5]). Since
earlier publications, developed semantic network has grown taking in account number of concepts.
This was necessary action, as most of advanced operations that can be carried with the WiSENet
cannot function well without extensive concept vocabulary. The most important was the
recognition of named entity (proper names, geographical names, names of organizations etc.) and
further acquisition them to semantic network what is possible using e.g. shallow text processing
methods [2].
Taking into account, that some of readers may not be familiar with specifics of WiSENet a brief
summary of its origin and capabilities is given.
To begin with, the WiSENet semantic network derived its whole content from the WordNet. The
decision was based on overall number of words and potential for further development and
restructuring.
The most important fact is that, author had to dismantle a synset structure and turn it into a graph
where nodes represent concepts and vertices denote lexical relation of hypernymy/hyponymy. This
enabled devised algorithms to easily follow relations among particular concepts found in real life
textual data (generally in unstructured text files). Restructuring was carried out in a lossless manner
(the algorithm is given in [7]).
Additionally, the WiSENet proved useful in combination with frequency dictionaries developed for a
number of various domains. These frequency dictionariesallow for highly efficient disambiguation of
concepts stored in the WiSENet. To some point, frequency dictionary coupled with semantic network
resembles human cognition when confronted with decisions concerning disambiguation. New
structure aided by domain frequency dictionaries proved to work well, results of application of the
WiSENet to semantic compression for English were highly satisfactory.
Semantic compression is a process throughout which reduction of dimension space (used for indexing)
occurs. The reduction entails some information loss, but in general it aims not to degrade quality of

Computer Science & Information Technology (CS & IT) 107
results thus every possible improvement is considered in terms of overall impact on the quality.
Dimensions’ reduction is performed by introduction of descriptors for viable concepts. Descriptors
are chosen to represent a set of synonyms or hyponyms in the processed passage. Decision is made
taking into account relations among the concepts and their frequency in context domain.
2. MOTIVATING SCENARIO FOR KNOWLEDGE ACQUISITION
As mentioned earlier, it was observed that the WiSENet lacks a great number of concepts that are to
be met in various textual data. Those most impeding experiments are originating from general
culture. Vast majority of identified missing concepts are proper names of various entities. For sake of
clarification, by proper names author understand names of people, organizations, geographical names
and various objects (e.g. names of products). The WordNet in general does not miss most general
categories of entities, yet a lot of highly specialised concepts is not present. As the WordNet was not
devised for text processing tasks previous statement is offered not as a criticism but as an
observation.
Stating the above author decided to invest effort in expanding the WiSENet. What is more important,
this effort surpasses traditional methods of bulk import of all available resources and their later
refactoring to match initial structure of to be extended semantic net.
It was observed that the WiSENet is very useful in discovering concepts that represent some
specialization of other concepts by employing specially prepared rules.
The WiSENet semantic network can be applied to a set of procedures, that aim is to extract
information from some textual data. As is well known in the domain of text processing, there
should be manually prepare a set of rules that trigger when given order of elements is met. A great
disadvantage to anyone who has to prepare this set of rules is that one is in need of specifying them in a
manner that enumerates every plausible variant of a rule.
For preparing a set of rules that enabling to retrieve information from the data, one should begin with
investigation of domain. Let’s assume, that the whole process should supply its invoker with new data
on people that hold managerial positions at various companies. First of all, one should issue an
recognizance query to a search engine of his choice, probing for concepts than can denote a
managerial position in some company.
It can be easily checked, that querying with search concepts such as: chairman, CEO, chief executive
officer, managing director, manager; shall bring results similar to following ones from the Table 1.
Table 1. Examples of persons with managerial positions
William (Bill) H. Gates is chairman of Microsoft Corporation
Richard K. Matros has been Chairman and CEO of Sun Healthcare
Novartis AG chairman Daniel Vasella steps down from the company he helped
build over 25 years, he leaves behind [...]
Larry Ellison has been CEO of Oracle Corporation
Amit Singhal is Senior Vice President and a Google Fellow
TerreStar Corp. (TSTR) President and CEO Jeffrey Epstein as its new chief
financial officer
Brian McBride joined Amazon UK as Managing Director

Amazon CEO Jeff Bezos
David Drummond joined Google in 2002. Today is senior vice president
[...] said Bill McDermott, co-CEO of SAP
Mr Krishan Dhawan, Managing Director of Oracle India Pvt. Ltd.
Jonathan Ian Schwartz (born October 20, 1965) was President and CEO at Sun
Microsystems prior to its acquisition by Oracle
Source: own elaboration
It is easy to observe a vast number of possibilities when it comes to word order in researched material.
Furthermore, the given list of search concepts is far from completion.
Standard methods of local pattern matching dictate creation of rules that trigger when exact number
of tokens of right characteristics is found. Apart from great effort investment spent on rule creation,
they are prone to misfiring when slightest change of word order occurs.
Good examples of local pattern matching are regular expressions and text processing automata.
While tremendous tools they might induce considerable effort when applied to information
extraction. First of all, it was observed that regular expressions tend to fail in information retrieval
task, not because their inefficiency but due to users being overwhelmed by their syntax. To
exemplify above lets point out that, one has to be an experienced user to produce regular
Figure 1: Excerpt from the WiSENet taxonomy showing concepts related to decision maker. Elements filled
in blue constitute corporate decision makers. Source: own elaboration

expression that will match more than 99% valid emails. As with practice comes experience, more
important issue with regular expression ([19] demonstrated that regular expressions can be
converted into non-deterministic finite state automata - FSA) is its sensitivity to word order
permutations.
Considering grammars, there should be remember that they will have to face the challenge of an
alphabet that is finite but actual number of symbols cannot be counted a priori. One has to process
whole corpora to enumerate all alphabet’s symbols. When processing a language such as English
this can be troublesome, as there is no known boundaries of resources that should be processed.
Ideal solution to above mentioned issues, shall combine flexibility and ease of use. Flexibility shall
be understood as ability to adapt to natural permutations in a word order of processed text. Ease of
use shall make a user exert the least amount of effort in a formulation of his information needs.
3. APPLICATION OF WISENET
Coming back to introduced motivation scenario, there is easily to observe that given results of
recognizance query share common structure. This structure shall be treated as case analysis which
leads to introduction of method designed by the author to automate information retrieval in this
specific task.
Every result contains some information on person, its position (managerial one) and some company.
Whether there is a task to build a datastore of data on managers in some kind of industry, a method
that works with such high level query concepts as executive, person and company name will be of
tremendous help.
When there should be start with a corpus of some textual data, it can be filtered it through
envisioned method and come up with elements that become candidates to extend current knowledge
base. Found elements in textual data can be new relations among already stored data, or new more
general/specific concepts directly in relation with existing ones. The whole process of acquisition of
new concepts and relations bases on the WiSENet semantic network structure. Effects of the process are
reflected onto it, thus subsequent usage yields better result than previous ones.
The WiSENet network stores a corporate executive as a concept. This concept has other concepts in
lexical relation, such as its hypernym and various hyponyms. A list of most important is given in
Figure 1.
4. ALGORITHM OF MATCHING RULES
Before applying algorithm for matching rules there is necessary to carry out text-refinement
process (starting from unstructured text document input and resulting finally a structure containing
stacked sequentially descriptors of concepts found in the input document). An action that make up
the process of text-refinement in documents starts from extracting lexical units (tokenization), and
further text refinement operations are: elimination of the words from the so-called information
stop-list, the identification of multiword concepts, bringing concepts to the main form by
lemmatization or stemming. It is particularly difficult task for highly flexible languages, such as
Polish, Russian or French (multiple noun declination forms and verb conjugation forms).
Synonyms need to be represented with concept descriptors using a semantic network. It allows
correct similarity analysis and also increases classification algorithms efficiency without loss in
comparison quality [14].
Abstracting process faces another problem here, which is polysemy. One word/phrase can represent
multiple meanings, so the apparent similarity need to be eliminated. The problem is distinguishing
words/phrases senses: for a polysemous word/phrases, we aim to have one node per sense in the

resulting network, merging all occurrences to the correct node. This corresponds to unsupervised
Word Sense Induction and Discrimination [28]. It is done by a procedure of concept
disambiguation, which identifies word/phrase meaning depending on its context, is important to
ensure that no irrelevant documents will be returned in response to a query [16], [17], [23]. Concept
disambiguation entails indicating the appropriate meaning for ambiguous concepts, which results
in obtaining information, as inferred from the documents, which better matches information needs.
Disambiguation method based on lexical relations from the semantic network examines word context
to determine its meaning, resulted in 82 % accuracy. It seems that only linguistic analysis methods can
exceed 90 % accuracy [25], while human experts are able to recognize correct meaning of 96,8 % of
polysemous words/phrases [24].
The last operation in the text refinement procedure is a generalization of concepts using semantic
compression.
The final effect of the refinement procedure is the structure of documents containing ordered
descriptors of concepts derived from the input document. This structure can be stored as an abstract
(containing data for creating index) of the document, and then use the algorithm for discovering new
concepts or new lexical relationships between concepts already existing in the WiSENet.
Devised algorithms uses ideas already mentioned in previous publications. All operations are
performed with the WiSENet as a semantic net. The first important step in the algorithm is a
procedure that unwinds rule into all hyponyms stored inside the network. This operation can be of
considerable cost in terms of execution as it has to traverse all possible routes from chosen concept to
terminal nodes in the network. After completion a list of rules is obtained, listing every possible
permutation of concepts from the semantic network. To shorten processing time, there should be
specify a number of levels that the procedure shall descend in its course of execution.
Next phase of the algorithm is to step through textual data in order to find matches on computed
rules. Stepping through is done by employing bag of concepts approach. The bag of concepts is
implemented as a Finite State Automaton (formally the automaton is a transducer) with advanced
methods for triggering desired actions. At any state, it check whether any of the rules to be matched is
completed. Discussion covering details of implementation is beyond the scope of this article.
Nevertheless, it can be visualized as a frame passing through textual data. With every shift towards
end of text fragment, concepts inside frame are used to check whether they trigger any of the rules
obtained in the first phase. Size of the bag is chosen by researcher, yet performed experiments show
that best results are obtained for a bag of size from 8 to 12 when rules are 2 to 5 concepts long.
Bag of concepts is very important idea, as it tolerates mixins and concept order permutations. All
matchings are performed after initial text processing phase is performed. Text processing phase
consist of well known procedures such as applying stop list and words/terms normalization.
A mixin is in this case a passage of text that serves some purpose to original text author, yet it
separates two or more concepts that exist in one of the computed rules. Consider following examples
placed in the Table 2.

Table 2: Examples of matching rules.
Rule: disease (all hyponyms), therapy (all hyponyms)
Match in: chemotherapy drug finish off remaining cancer
Matched concepts therapy → chemotherapy, disease → cancer
Mixin
drug finish off remaining
Match in: gene therapy development lymphoma say woods
Matched concepts therapy → gene therapy, disease → lymphoma
Mixin development
Match in: cancer by-bid using surgery chemotherapy
Matched concepts therapy → chemotherapy, disease → cancer
Mixin by-bid using surgery
Match in: Encephalitis is an acute infection and inflammation of the brain
where therapy is supportive treatment
Match concepts disease → acute infection, therapy → supportive
treatment
Mixin
encephalitis, brain
Examples are taken from one of the experiments performed with biology and medicine corpus. It
can be observed, that bag of concepts performs well in various cases, it handles long mixins and
concept permutation. Additional observation shall be made as concepts being hyponyms to those in
the original example rule were matched (as referenced earlier).
All experiments performed took into account possibility of matching more than single rule. Thus a
mechanism for triggering a set of rules was devised and was signaled earlier along with the bag of
concepts.
A procedure matching rules holds special internal registers, that store rules that are actively valid
with given bag of concepts and actual results of filtering textual data. To give an example, please
consider a set of three following rules:
rule 1 : university, city (all hyponyms)
rule 2: university, city (all hyponyms), country (all hyponyms)
rule 3 : first name (all hyponyms), academic (all hyponyms)

Given is exemplary text fragment:
A team of chemists led by chemistry professor David Giedroc from Indiana University (in
Bloomington, USA) described a previously unknown function of a protein they now know is
responsible for protecting a major bacterial pathogen from toxic levels of copper. Co-author
with Giedroc on the paper is professor Michael J. Maroney of the University of Massachusetts.
The results were published Jan. 27 in Nature Chemical Biology.
Procedure shall match and matches previously defined rules:
The procedure shall match and matches previously defined rules:
rule number 1: with university → university, Bloomington → city, new concept: Indiana
University in Bloomington
rule number 2: with university → university, Bloomington → city, USA → country, new
concept: Indiana University in Bloomington
rule number 3: with David → first name, professor → academic, new concept: David Giedroc =
professor(Indiana University University in Bloomington)
rule number 3: with Michael → first name, professor → academic, new concept: Michael J.
Maroney = professor(University of Massachusetts)
Figure 2: Process of matching rules from Example 1. Source: own elaboration

When a complete rule or its part (one can decide whether he is interested in total matches all partial
ones) is mapped, it is presented to a user to accept match or reject it. When the bag of concepts drop
earlier concepts and is filled with new ones, rules that were not matched are dropped from register
of valid rules.
Algorithm in pseudocode is presented in listing 1
Algorithm 1: Algorithm for matching rules using WiSENet and bag of
concepts
//attach rule triggers to concepts in semantic network
mapRulesToSemNet(SN, R[])
for all Rule ∈ R do
for all Word, Relations ∈ Rule do
N = SN.getNeighbourhood(Word, Relations)
for all Word ∈N do
SN.createRuleTrigger(Word, Rule)
end for
end for
end for
// Phase 2: text processing: tokenization, phrases, stop list
T = analyzeText(Input)
foreach Word in T do
if count(Bag) = size(Bag) then
//First, deactivate rules hits for a word
//that drops out from bag of words
oldWord = pop(Bag)
end if
for all Rule ∈ SN.getTriggers(oldWord) do
Rule:unhit(Word)
push(Bag, Word)
for all Rule ∈ SN.getTriggers(Word) do
//take all relevant rules and activate word hit
Rule.hit(Word)
if Rule.hitCount = Rule.hitRequired then
//report bag of words when hits reaches required number
report(Rule, Bag)
end if
end for
end for
SN - Semantic Network
R - semantic relation pattern

5. EXPERIMENT
Devised algorithm (automaton) was used to perform an experiment on biology related data. The
main aim of the experiment was to discover new concepts and insert them in the WiSENet semantic
Netowrk structure. The test corpus of documents consisted on 2589 documents. A total number of
words in a document's corpus was over 9 million. Author decided to prepare query for searching for
specialists and their affiliations. This converges with motivating scenario, as the WiSENet semantic
network was enriched by both specialists (and their fields of interest), universities, institutes and
research centers.
Experiment used following rules:
rule 1: first name (all hyponyms), professor (all hyponyms), university (all hyponyms)
rule 2: first name (all hyponyms), professor (all hyponyms), institute (all hyponyms)
rule 3: first name (all hyponyms), professor (all hyponyms), research center (all hyponyms)
rule 4: first name (all hyponyms), professor (all hyponyms), department (all hyponyms)
rule 5: first name (all hyponyms), professor (all hyponyms), college
Size of the bag of concepts was set at 8 elements. Additionally, all rules were to match exactly all
concepts.
Out of 1326 documents where concept ”professor” was found, prepared rules matched 445 text
fragments. This gives a recall rate of 33,56%. Precision of results was 84,56%. This level is found to
be very satisfactory, especially taking into account that due to algorithm nature there can be
duplicates of matched text fragments (due to multiple triggering of rules inside current bag of
concepts).
Table 3 demonstrates sample results. Please notice, that match on its own does not discover new
concepts. Rules present potential fragments that with high likelihood contain new concepts that can be
included into semantic network.
In addition, experiment resulted in 471 concepts that were previously unknown to the WiSENet
network. The context and the type of rules that matched text fragments led to extremely efficient
updates of the network. The author developed a new version of the devised algorithm, in which the
goal is to discover new lexical relations between concepts. So, the same mechanism of creating
rules can be applied to discover and store new lexical relationships using rules containing concepts
already stored in the semantic network.
Further experiments performed with this new version showed that the same corpus and the same
rules can be used to acquisition new lexical relationships between concepts already stored in the
semantic network.

Table 3: Sample results of experiments with rules based on WiSENet on corpus of biology related
documents. Discovered concepts are written under matches.
text fragment match/discovered concept rule
explain senior author Douglas Smith Md professor
department neurosurgery director
Douglas professor department
Douglas Smith
5
Feb proceedings national academy of sciences
researcher University of Illinois entomology
professor Charles Whitfield postdoctoral
University of Illinois
professor Charles
Charles Whitfield
1
design function biological network she-bop
visiting professor Harvard University Robert
Dicke fellow visiting
professor Harvard University
Robert
Robert Dicke
1
modify bacteria Thomas Wood professor
--Artie-- --McFerrin-- department chemical
engineering have
Thomas professor department
Thomas Wood
5
Matthew --Meyerson-- professor pathology Dana
--Farber-- cancer institute senior associate
Matthew professor institute
Matthew Meyerson
2
an assistant professor medical oncology Dana
--Farber-- cancer institute researcher broad
assistant
professor Dana institute
Dana Farber
2
sun mat professor emeritus Robert --Hodson-- all
university Georgia Robert Edwards
professor Robert university
Robert Hodson
1
vacuole David Russell professor molecular
microbiology --Cornell's-- college veterinary
medicine colleague
David professor college
David Russell
4
chemistry professor David Giedroc from Indiana
University (in Bloomington, USA) described a
previously unknown function of a protein
David Professor university
Davic Giedroc
1
resistant cell professor Peter --Sadler-- chairman
chemistry department University of Warwick lead
research project
professor Peter University of
Warwick
Peter Sadler
1
said first author Quyen Nguyen doctorate assistant
professor surgery si tan University of California
San Diego school of medicine
Nguyen assistant professor
University of California
Quyen Nguyen
1
scientist --Sirtris-- co-author founder prof David
--Sinclair-- Harvard Medical School published
consecutive
professor David Harvard
Medical School
David Sinclair
1
6. CONCLUSIONS
In this article have been presented an approach for building or expanding large-scale semantic
networks automatically from text, employing deep semantic processing with appropriate
mechanisms (finite state automaton).
The work presented in this article continues research efforts started with presentation of
Semantically Enhanced Property Protection System SeiPro2S [5]. The SeiPro2S system has proved

to be a efficient tool in checking whether submitted content is not an unauthorized copy. SeiPro2S
makes it possible to not only find direct copying, but also to find passages that rephrase the copied
content with another set of words, thus reproducing the original thought. Designed SHAPD2
algorithm is highly efficient in plagiarism detection task and employing semantic compression is
strong resilient to false-positives examples of plagiarism (see [9]), which is may an issue in case of
using competitive algorithms [27].
The SHAPD2 algorithm has extremely low computational complexity estimated as linearithmic and
uses technique of hashing whole sentences. The final architecture of the SeiPro2S system and its
functionality has been obtained by introducing new mechanisms which effectiveness was established
thanks to performed experiments and was described in [9].
After realizing vision of semantic compression for English and presenting results, author decided to
focus on applications enabling network expansion with new concepts and new lexical relationships
using specially constructed automata (which is functionally a transducer) what is necessary to
increase performance quality as the WordNet realizing NLP or computational linguistics tasks.
Rules created with the WiSENet are interesting application, that has great potential for future
development, as it helps to expand body of knowledge represented by WiSENet. Experiments
performed with devised algorithm for rule matching showed that envisioned flexibility and
precision are available. Further experiments performed with devised algorithm showed that the
same corpus and the same rules can be used to acquisition new lexical relationships between
concepts already stored in the semantic network.
As reported in the experiment section, due to reasonably high precision on achieved results, unknown
concepts can be easily added, thus realizing a vision of knowledge acquisition with the WiSENet.
Future work will focus on further addition of previously unknown concepts to the WiSENet along
with restructuring of relations among them. Author believes that there are even more useful
applications of semantic compression and plan to experiment with them and share experiments’
results.
REFERENCES
[1] Baeza-Yates, R., Ribeiro-Neto, B. (1999), "Modern Information Retrieval", ACM Press,
Addison-Wesley Longman Publishing Co., New York
[2] Becker, M., Drozdzynski, W., Krieger, H. U., Piskorski, J., Shaafer, U., Xu, F. (2002), SProUT - shallow
processing with unification and typed feature structures, Proceedings of the International Conference on
Natural Language Processing, ICON-2002
[3] Blackburn, P., Bos, J. (2005), Representation and Inference for Natural Language, A First Course in
Computational Semantics, CSLI Publications
[4] Brachman, R. J., Levesque, H. J. (2004), Knowledge Representation and Reasoning, Morgan Kaufmann,
p. 381
[5] Ceglarek, D., Haniewicz, K., Rutkowski, W. (2009), Semantically Enchanced Intellectual Property
Protection System - SEIPro2S, 1st International Conference on Computational Collective Intelligence,.
Nguen N.T. (ed.) In: Lecture Notes in Artificial Intelligence - vol. 5796: Computational Collective
Intelligence - Technologies and Applications, pp. 449-59, Springer Verlag Berlin Heidelberg
[6] Ceglarek, D., Haniewicz, K., Rutkowski, W. (2010), Semantic compression for specialized Information
Retrieval systems, Studies in Computational Intelligence, vol. 283 of Lecture Notes in Artificial
Intelligence, pp. 111–121, Springer Verlag, Berlin Heidelberg

[7] Ceglarek, D., Haniewicz, K., Rutkowski, W. (2010), Quality of semantic compression in classification,
Lecture Notes in Artificial Intelligence - vol. 6421: Computational Collective Intelligence -
Technologies and Applications, pp. 162–171, Springer Verlag, Berlin Heidelberg
[8] Ceglarek, D., Haniewicz, K., Rutkowski, W. (2012), Robust Plagiary Detection Using Semantic
Compression Augmented SHAPD, 4th 1st International Conference on Computational Collective
Intelligence - ICCCI 2012, pp. 308–317, Lectures Notes in Computer Science, Springer Verlag, Berlin
Heidelberg
[9] Ceglarek, D. (2013), Architecture of the Semantically Enhanced Intellectual Property Protection System,
Lecture Notes in Artificial Intelligence - Computer Recognition System 5, pp. 12, Springer Verlag,
Berlin Heidelberg
[10] Fellbaum, C. (1998), WordNet - An Electronic Lexical Database, The MIT Press
[11] Goddard, C., Schalley, A.C. (2010), Semantic Analysis, edit. Indurkhya N. & Damerau F. In:
Handbook of Natural Language Processing, , pp. 93-12, Chapman Hall/CRC
[12] Gonzalo, J., Vardejo, F., Chugur I., Cigarrán, J.M. (1998), Indexing with WordNet Synsets can improve
Text Retrieval, Proceedings of the COLING/ACL '98 Workshop on Usage of WordNet for NLP, pp.
38-44
[13] Hotho, A., Staab, S., Stumme, S. (2003), Explaining Text Clustering Results using Semantic Structures,
Principles of Data Mining and Knowledge Discovery, 7th European Conference PKDD 2003
[14] Hotho, A., Maedche, A. , Staab, S. (2003), Ontology-based Text Document Clustering, Proceedings of
the Conference on Intelligent Information Systems, Zakopane, Physica/Springer
[15] Keikha, M., Razavian, N. S., Oroumchian, F., Razi, H. S. (2008), Document Representation and Quality
of Text: An Analysis, ed.. M. W. Berry M. Castellanos, Survey of Text Mining II: Clustering,
Classification, and Retrieval, pp. 219-232, Springer Verlag, Berlin Heidelberg
[16] Khan, L., McLeod, D., Hovy, E. (2004), Retrieval effectiveness of an ontology-based model for
information selection, The VLDB Journal, volume 13 (1), pp. 71-85, Springer Verlag, Berlin
Heidelberg
[17] Krovetz, R., Croft, W. B. (1992), Lexical Ambiguity and Information Retrieval, ACM Transactions on
Information Systems, vol. 10(2), pp. 115-141
[18] Frakes, W. B., Baeza-Yates, R. (1992), "Information Retrieval: Data Structures and Algorithms",
Prentice Hall
[19] McNaughton, R., Yamada, H. (1960), Regular expressions and state graphs for automata, IRE
Transactions on Electronic Computers EC-9(1), pp. 39–47
[20] Miller, G. A. (1995), Wordnet: a lexical database for English, Communications of the ACM, vol. 38 (11),
pp. 39–41
[21] Califf, M. E., Mooney, R. J. (2003), Bottom-up relational learning of pattern matching rules for
information extraction, Journal of Machine Learning Resources, vol. 4, pp. 177-210
[22] Sinha, R., Mihalcea, R. (2007), Unsupervised graph-basedword sense disambiguation using measures of
word semantic similarity, pp. 363–369, In: ICSC
[23] Zhu, J., Hovy, E.H. (2007), Active Learning for WordSense Disambiguation with Methods for
Addressingthe Class Imbalance Problem. In: Proceedings of the EMNLP Conference, Prague
[24] Gale, W., Church, K., Yarowsky, D. (1992), A Method for Disambiguating Word Senses in a Large Corpus,
pp. 415--439, Computers and the Humanities. vol. 26
[25] Sanderson, M. (2000), Retrieving with Good Sense, Information Retrieval, pp. 45--65
[26] Sowa, J. (1991), "Principles of Semantic Networks", Morgan Kaufmann

[27] Zhang, Q., Zhang, Y., Yu, H., Huang, X. (2010), Efficient partial-duplicate detection based on sequence
matching, In: SIGIR '10: Proceeding of the 33rd international ACM SIGIR Conference on Research and
development in information retrieval, pp. 675--682, New York, NY, USA, ACM
[28] Navigli, R. (2009), Word sense disambiguation: A survey. ACM Computational Survey 41(2), pp. 1–69
Author
Dr. D. Ceglarek is a Assistant Professor of Information Technology, Poznan
School of Banking, Poznan, Poland. He earned a Ph.D in Clustering of information
from Department of Information Systems, Poznan University of Economics,
Poznan, Poland in 1997. He has published more than thirty of research papers in
the area of classification of information, plagiarism detection, natural language
processing and semantic technology. He has served as Assistant Professor in
Department of Information Systems, Poznan University of Economics during 1997
to 2007.

SEMANTIC NETWORK BASED MECHANISMS FOR KNOWLEDGE ACQUISITION

More Related Content

What's hot (17)

Similar to SEMANTIC NETWORK BASED MECHANISMS FOR KNOWLEDGE ACQUISITION (20)

More from cscpconf (20)

Recently uploaded (20)

SEMANTIC NETWORK BASED MECHANISMS FOR KNOWLEDGE ACQUISITION