DICTIONARY-BASED CONCEPT MINING: AN APPLICATION FOR TURKISH

David C. Wyld et al. (Eds) : CST, ITCS, JSE, SIP, ARIA, DMS - 2014
pp. 27–38, 2014. © CS & IT-CSCP 2014 DOI : 10.5121/csit.2014.4103
DICTIONARY-BASED CONCEPT MINING:
AN APPLICATION FOR TURKISH
Cem Rıfkı Aydın1
, Ali Erkan1
, Tunga Güngör1
and Hidayet Takçı2
1
Department of Computer Engineering, Boğaziçi University, Istanbul, Turkey
cemrifkiaydin@gmail.com, alierkan@gmail.com, gungort@boun.edu.tr
2
Department of Computer Engineering, Cumhuriyet University, Sivas, Turkey
htakci@gmail.com
ABSTRACT
In this study, a dictionary-based method is used to extract expressive concepts from documents.
So far, there have been many studies concerning concept mining in English, but this area of
study for Turkish, an agglutinative language, is still immature. We used dictionary instead of
WordNet, a lexical database grouping words into synsets that is widely used for concept
extraction. The dictionaries are rarely used in the domain of concept mining, but taking into
account that dictionary entries have synonyms, hypernyms, hyponyms and other relationships in
their meaning texts, the success rate has been high for determining concepts. This concept
extraction method is implemented on documents, that are collected from different corpora.
KEYWORDS
Concept mining, Morphological analysis, Morphological disambiguation
1. INTRODUCTION
Concept can be thought of as a general idea or something specific conceived through the mind. A
human can easily have a general opinion concerning something which he/she is reading, watching
or to which she is listening. But as for computers, since they lack the functionality of human brain
that can easily classify many patterns and have conceptual ideas on an object, they have to
process many statistical or machine learning methods.
Concept mining is a field of study where text, visual or audio materials are processed and general
concepts are extracted. Amongst its applications, the most commonly used approach for
extracting concepts is concerned with text materials. In this paper we are interested in extracting
concepts from textual documents. Concept mining is a hard and very beneficial operation. In
many areas it can be used in an automatic, efficient and computerized manner. For example it can
ease categorizing the judiciary classes, such as adult court, appellate court and many others
[1].Also in medical field, it can help both the patients and doctors in many different ways
[2, 3].Similarly, banks and other financial institutions may use concept mining to track the
profiles of some creditworthy customers.
During last decades, through the enhancement in natural language processing, many artificial
intelligence methods have been developed and used for extracting concepts, such as Latent
Dirichlet Allocation (LDA) that aims to determine topics of a textual material. Besides those

28 Computer Science & Information Technology (CS & IT)
methods, the most widely used lexical database for this field of study is WordNet. In WordNet
the relationships between the words are stored in groups called synsets. These sets include
hypernymy, hyponymy, synonymy relationships which are the most important and commonly
used ones amongst many others. Hypernymy of a word can give us a general meaning of that
word, hereby this relational property is the most widely used one for concept mining. This
database is widely used for English concept extraction, but as for languages that are not as widely
spoken as English, WordNet is not properly built and many word relationships (such as
hypernymy) lack. That is why in these languages, the use of this lexical database would not yield
as successful and meaningful results as it does in English.
In this paper, we used a novel approach for extracting concepts in Turkish through the use of a
dictionary. Dictionaries have word entries, meaning texts, lexical class (noun, adjective, etc.) and
some other properties. The meaning text of a word does, in fact, include its many relationally-
relevant words, such as hypernyms or synonyms. So it is useful to benefit from the meaning texts
in this dictionary to extract concepts. But these dictionary data are in an unstructured form, so
they are pre-processed in this work, then concept mining method is run. The dictionary that is
used is TDK (Türk Dil Kurumu - Turkish Language Association) Dictionary, which is the official
Turkish dictionary that is the most respectful, accurate and most widely used one. We used the
bag-of-words model for this study and we developed an algorithm which takes the frequencies of
words in the document into account, since concepts do generally have to do with the most
frequent words and the words related to those words present in the textual material. The success
rate we observed is high compared with the other works done in concept mining for Turkish.
The remainder of the paper is as follows. In Section 2 related work concerning this field of study
is examined. Section 3 elaborately examines the method developed in this work. Section 4 gives
the evaluation results and the success rates achieved with the proposed method. Finally, the
conclusion and future work are given in Section 5.
2. RELATED WORK
Numerous comprehensive studies are done in widely used languages in the domain of concept
mining. In studies carried out in this domain, generally machine learning algorithms are used.
Also WordNet, a lexical database, is commonly used. As for Turkish, the concept mining domain
is still immature and also in this language some machine learning methods are used to extract
concepts.
In one study, it is proposed that using WordNet's synset relations and then implementing
clustering process may help obtaining meaningful clusters to be used for concept extraction [4].
The synset relationships are used as features for clustering documents and then these are used for
succeding clustering. But in this study word disambiguation has not been performed and it is
observed that using synset words has decreased the clustering success.
There have been also developed some toolkits for concept mining, one of which is ConceptNet
[5]. In accordance with this toolkit, similar to the WordNet synset relationships, spatial, physical,
social, temporal and psychological aspects of everyday life are taken into account. In this toolkit,
concepts are extracted by processing these relationships and a data structure is built. For instance
a hierarchical relationship is named "IsA" and in a document the counterpart relational word of
the word "apple" may be "fruit" in accordance with this relationship. If there are many common
words that are extracted through relationships in a document, by taking also their frequencies into
accounts, successful results may be achieved. ConceptNet is much richer than WordNet in terms
of its relational structure.

Computer Science & Information Technology (CS & IT) 29
As for Turkish, there have been a few studies concerning this study. It was proposed that
meaningful concepts could be extracted without the use of a dictionary and with the clustering
method processed on documents in corpora [6]. Here clusters would have initially been assigned
concepts and the documents to which clusters are assigned, would have those concepts.
In an another study, a method is developed for extracting concepts from web pages [7]. The
frequencies of words in the document are taken into account and words are assigned different
scores according to their html tags. Words between specific tags such as "<b>" and "<head>" are
assigned higher scores. Then the words exceeding specified thresholds are determined as
concepts. Success rates are reported to be high.
A study is done concerning extracting concepts from constructing digital library, documents in
this library are categorized through clustering in accordance with those concepts [8]. An equation
is created that multiplies term frequency (tf), inverse document frequency (idf), diameter, length,
position of the first occurrence, and distribution deviation values of the keywords. Whichever
words give the highest scores in accordance with this equation, they are selected as probable
concepts. A higher success rate is achieved as compared with methods that take only tf and idf
into account used for extracting concepts.
There have been used mostly lexical, relational databases and clustering methods for Concept
Mining, besides those also Latent Dirichlet Allocation [9] and some other artificial intelligence
methods have been used for extracting topics and concepts. Our method uses a statistical method
that makes use of a dictionary, which has not been used in Turkish so far.
3. METHODOLOGY
In the concept mining domain, several dictionaries and lexical databases such as WordNet are
used. Structural and semantic relationships between words can give us general idea (concept)
about the words. In WordNet, especially hypernymy relation [10] is preferred in concept
extraction, since it is the most relevant relation to a generalization of a word. There are also other
relationships in WordNet such as meronymy and synonymy, those relationships give us other
semantic relevances that may be useful in determining the concept of a word. For example, the
synonym of the word "attorney" is "lawyer" and if a document from which concepts are to be
extracted has many occurrences of the word "attorney", one probable concept may be "lawyer".
So far, the use of WordNet in Concept Mining has been extremely dominant. Taking it into
account, in this work we used a novel approach, that is we used dictionary to help extract
concepts from documents.
3.1 Dictionary Structure
The dictionary we used in this work includes the words and the meaning text of those words,
where the words in the meaning carry specific relationships with each other. Figure 1 shows the
XML structure of a dictionary item.

Figure 1. Structure of a dictionary entry, "jaguar", in XML format
In Figure 1, the tag "<atasozu_deyim_bilesik>" stands for "proverb, idiom, compound" and the
tag "<birlesik_sozler>" stands for "compound phrases" in Turkish. In this work we took only
"name", "lex_class" and "meaning" tag elements into account. We processed only the words,
"lex_class" of which are nouns, also analogies are made based on the meaning text nouns. There
may be several meanings of a word, we took all of them into account and selected only one of
them after disambiguation. The main relationships that would be encountered in dictionary item
meaning texts can be summarized as follows:
• Synonymy: It is a relation that two words have equivalent meanings. It is a symmetrical
relationship. (For example the words "intelligent" and "smart" have this relationship.)
• Meronymy: It is a relation that one of the words is a constituent of the other word. It is
not a symmetrical relationship. (For example the words "hand" and "finger have this
relationship.)
• Location: It is a relation that shows the location of a word with respect to the other word.
(For example the words "kitchen" and "house" have this relationship.)
• Usability: It is a relation that one word is used for an aim. (For example "toothbrush" is
used for "brushing teeth".)
• Effect: It is a relation that one action leads to a result. (For example taking medication
leads to a healthy state.)
• Hypernymy: It is a relation that one word is a general concept of an another word. (For
example the word "animal" is the hypernym of the word "cat".)
• Hyponymy: It is a relation that one word is a more specific concept of an another word.
(For example the word "school" is hyponym of the word "building".)
• Subevent: It is a relation that one action has a sub-action. (For example waking up in the
morning would make one yawn.)
• Prerequisite relation: It is a relation that one action is a prerequisite condition for another
one. (For example waking up in the morning is a prerequisite condition for hitting the
road for job.)
• Antonymy: It is a relation that one word is the opposite concept of an another word. (For
example the word "happy" is the antonym of the word "sad".)
This dictionary model may be used in many forms and has advantages as compared with the use
of specific WordNet synset relations. For example, if we want to relate words to one another and

try making analogies between those words, the use of dictionary would be very helpful. Through
implementation of clustering, many meaningful clusters can be built, because the analogies
between the properties of words (such as the words 'finger' and 'hand' have meronymy relation
and those would be in same cluster) may connect them in a semantic relationship. But instead of
clustering process, we followed a simple statistical method, which would yield successful results.
3.2 Dictionary Preprocessing
In WordNet, relations are held in specific synsets, that is they have a structure based on word to
word or word to noun-phrase relationships. So no parsing or disambiguation is needed for
WordNet since words are in their root forms. However in basic dictionaries, there are no specific
data structures that separately hold different synset relations. For example the meaning text of the
word "cat" is as follows:
"A small carnivorous mammal (Felis catus or F. domesticus) domesticated since early times as a
catcher of rats and mice and as a pet and existing in several distinctive breeds and varieties."
In the above example, the noun "rats" must be handled as the word "rat", whereas "mice" should
be normalized as "mouse". The words should go through tokenization, stemming and
normalization processes and then be taken into account. English is a language that has no much
complexity concerning stemming, but as for Turkish, it is an agglutinative language and the
parsing and disambiguation processes would quite matter. We used the parser and disambiguator
tools, that are BoMorP and BoDis, developed for Turkish by Hasim Sak, Boğaziçi University [11,
12]. BoMorP implements the parsing operation and analyses the inflectional and derivational
morphemes of a word and proposes part-of-speech (POS) tags. BoDis disambiguates amongst
those POS tags and returns the one having the highest score according to an averaged perceptron-
based algorithm.
A concept may be concerning a general idea of an abstract or concrete object, so the most
probable concepts are generally nouns. So in this study we assumed that nouns represent concepts
more than do any other word categories, so we took only nouns as concepts in the documents into
account.
3.3 Document Analysis using Dictionary
A dictionary item may have many meanings and we have to determine which meaning is used in
the context of a document. For example the word "cat" has many meanings and if the document
contains this word we have to extract which meaning of this word is used in the document we are
handling. For this, we looked up in the dictionary meaning text nouns as well as the nouns in the
context of the word in the document and took the one yielding the highest similarity measure that
has most common words. This formula is as follows:
argmaxmSimilarity(m, cw) = CommonCount(m, cw)
In this formula, m denotes a specific meaning of a word in a dictionary, whereas Cw denotes the
context of a word in a document. In this work, we used a context size of 30 words, such that 15
words that are on the left and 15 words that are on the right of the candidate word are taken into
account. If amongst those words, many occurrences are encountered also in a specific meaning of
the word to be disambiguated, that meaning would be selected as the true meaning. We also
normalized the score (CommonCount) by dividing it by the number of nouns in the dictionary
meaning text to get more accurate results.

3.4 Mapping Concepts
3.4.1 Simple Frequency Algorithm
A concept that can be extracted from a document has generally to do with frequency of the
specific word encountered in that document. For example, if we encounter a document that
abounds with the word "football", we may be inclined to think that the concept of this document
would be concerning "sport". Taking frequency into account, we developed a statistical method,
that extracts concepts favouring the words that are more frequent.
We first take all the nouns in the document(s) and label them as pre-concepts. Here we eliminate
other types of words, such as adjectives and verbs. Then we start building a matrix. This matrix
has rows representing the nouns encountered in the document and columns representing the nouns
encountered in the meaning text sentences of those row words. But we also have to take into
account that the rows on the words are also added as column items. For example the word
'football' may be very frequent in the document, so this word should be regarded as a probable
concept as well.
The cells in the matrix are filled as follows: After we built the matrix, we fill the cells by 1 or 0
depending on whether the column word appears in the row word's meaning text or not. Then we
implement frequency operation: We multiply all the cell values in the matrix by the
corresponding row word frequency. For instance, if the word "football" is encountered 10 times in
a document and its meaning text nouns in the dictionary are "sport", and "team", then those
columns' ("sport", "team" and "football") values in the correspondent row "football" would be
updated as 10, whereas the other columns would be updated as 0.
Then we summed up the column values in the matrix and took the column word that yields the
highest summation as the probable concept. This is meaningful since the concept may or may not
be present in the document, so the use of dictionary would be beneficial. An example showing the
mapping of terms into concepts is shown in Figure 2.
Figure 2. An example showing the mapping of words in a document into probable concepts.
In the above example, the column to the left is a representative of words encountered in the
document, whereas the column to the right includes the representative nouns encountered in their
meaning texts. Due to that the words "football" and "handball" are frequent in the above example
and the word "sports" is present in their meaning texts, its score would be three and this word
would be assigned as the probable concept of the document or corpus.
As mentioned above, we benefit from dictionary to extract concepts from documents, but instead
of using just meaning text nouns for this concept extraction process, we also built a hierarchical
data structure that contains 2, 3 and 4 levels. In accordance with this structure, the main word is
atop the hierarchy, then the meaning text nouns of this word is in the lower level, whereas the

respective meaning text nouns of these meaning text nouns are in the lower levels. An example of
this data structure with 3-levels is depicted in Figure 3.
This hierarchical structure may have some specific features, for example each word in different
levels may be assigned a different coefficient and we may take this coefficient factor into account
when building up the matrix. If we construct 3-level hieararchies built through the dictionary, we
may assign high values for the top levels and low values for lower levels. This is the case because
the semantic relationship between the main word and the lower level nouns weakens while going
down through the hierarchy structure. We multiplied the top-level words in the matrix by 1, the
second-level words by 0.5 and the lowest-level words by 0.25. We used this geometric approach
since the meaning text nouns' frequencies increase geometrically from one level to the below one.
But we noticed that 3-level structure gives a very low success rate, so we preferred 2-level
structure with no coefficients yielding high accuracy values. Also 4-level structure gave even
worse results than did 3-level one, so using a 2-level structure was the best choice.
Figure 3. A hierarchical data structure with 3-levels of the word 'cat' in the dictionary.
We filled the matrix cells, as mentioned above, without taking frequency into account and saw
that the results yielded were much less successful. That shows the importance of taking frequency
into account.
We also have to take into account that some words are quite common in the dictionary, such as
"situation", "thing", "person" and so on. Here we determined the top 1% most frequent words in
the meaning texts in the dictionary as stop-words and eliminated them. Generally tf-idf is used
for elimination of words, but since we make use of the dictionary as a base, top words elimination
is sufficient.
3.4.2. Context and Frequency Algorithm
Although we noticed the algorithm we developed stated above gave meaningful concepts,
drawbacks can be clearly seen. For example, let's assume there is a document containing the noun
"football" and there is no other noun and its meaning text in the dictionary is as follows:
"A game played by two teams of 11 players each on a rectangular, 100-yard-long field with goal
lines and goal posts at either end, the object being to gain possession of the ball and advance it in
running or passing plays across the opponent's goal line or kick it through the air between the
opponent's goal posts."
According to the algorithm stated above (Section 3.4.1), we take the nouns in this meaning text
into account and build up a matrix containing those nouns, including the word 'football'. Since the
word "football" is seen 3 times, the column labeled "goal" has a value of 3 as well and at the end
the probable concept may be the word "goal", as well as the other concepts may be "game",
"team", "line" and other nouns in the meaning text. (This is the case since the matrix would be of
size 1 x CountNoun(MeaningTextOf(Football)), indicating that there is only one noun, that is

"football", in the document.) Having a concept of "goal" through this document would be a bit
nonsense, hence we modified the algorithm in the following manner:
All the dictionary meaning text nouns would not be useful in determining the general idea of the
main word, so some of those nouns have to be eliminated. In order to determine which meaning
text noun is relevant in the use of the main word, we used a corpus-based context analysis. We
had a few corpora and for each corpus, we used a 30-word window size context analysis, that is
we looked up 15 words on the left of the test words and 15 words on the right of the test words.
Hereby we eliminated the context words which are not nouns, because we think of concepts as
only nouns. Then we assumed that if a context word is also present in the meaning text of the
main word in dictionary, we take this context word into account. After scanning the whole
corpus, whichever context word is seen most, given that context word is also seen in the meaning
text of the main word, we add this word as a column word in the matrix corresponding to the row
word. Then, similar to what we have done in (Section 3.4.1), we multiply the row elements values
by the frequency of the row representative word and sum up the columns values. Whichever
column value has the maximum value, we define that column representative word as the probable
concept. In this case, we take mostly two words for each word in the document: The word itself
and the word in the meaning text of this word that is most widely seen in the contexts in the
corpus. We again, of course, firstly eliminated the stop words present in the TDK Dictionary.
This approach makes sense, since all meaning text nouns would not be useful in determining the
general idea, that is concept, of a word. Also the corpus-based approach shows that the most
relevant word in the meaning text of a test word is extracted through the context analysis.
Selecting at most two words, that are the word itself and the most frequent word in the contexts of
the word that is also present in the meaning text of the row word in the matrix rather than taking
into account all nouns in the meaning text of a word increased the success rate for three of the
corpora.
4. EVALUATION AND RESULTS
We tested our algorithms on four corpora in Turkish, which are as follows: Gazi University, Sport
News, Forensic News and Forensic Court of Appeals Decisions. The corpora we handled are in
different fields of topics as follows:
• Gazi University Corpus: This corpus includes 60 text files. There is no specific topic in
this corpus, instead the files collected represent different fields of topics, such as tutorials
on specific architectural or engineering subjects and many others.
• Sport News Corpus: This corpus includes 100 text files. Each file has a topic concerning
sport news, especially Turkish football.
• Forensic News Corpus: This corpus includes 100 text files. These files are collected
from news that are concerning forensic domain.
• Forensic Court of Appeals Decisions Corpus: This corpus has 108 text files. These
files are concerning the forensic decisions and this corpus is similar to the corpus
Forensic News. The difference is that this one is not collected from news documents
database.
Initially we parsed and disambiguated the files in those corpora through the tools BoMorp and
BoDis to extract the nouns needed for concept mining. Then we discarded the stop-words that
abound in the dictionary.

4.1 Evaluation Metric
As an evaluation metric, we made use of the accuracy metric, which can be defined as follows:
ܽܿܿ‫ݕܿܽݎݑ‬ =
‫݁ݑݎݐ‬ ‫݁ݒ݅ݐ݅ݏ݋݌‬ + ‫݁ݑݎݐ‬ ݊݁݃ܽ‫݁ݒ݅ݐ‬
‫݁ݑݎݐ‬ ‫݁ݒ݅ݐ݅ݏ݋݌‬ + ݂݈ܽ‫݁ݏ‬ ‫݁ݒ݅ݐ݅ݏ݋݌‬ +
‫݁ݑݎݐ‬ ݊݁݃ܽ‫݁ݒ݅ݐ‬ + ݂݈ܽ‫݁ݏ‬ ݊݁݃ܽ‫݁ݒ݅ݐ‬
In order to evaluate the concepts extracted through the algorithm we developed, we also manually
determined the concepts in those corpora. Then we compared the concepts with one another. We
handled many ways to compare the concepts suggested by the algorithm with the ones extracted
manually: For documents we made comparisons in 3, 5, 7, 10 and 15 window sizes. These
windows include the top concepts, for example the one with size 5, includes the top 5 concepts. A
comparison example table is shown below:
Table 1. An example showing the top 3 concepts in 2 documents.
Documents Algorithm Manual
Document 1 Sport, Game,
Match
Sport, Match,
Politics
Document 2 Court,
Attorney,
Judge
Attorney,
Accused,
Match
Table 1 shows the top three concepts for two documents, extracted both manually and
algorithmically. In the first document, it can be seen that the success rate is 2 / (2 + 1) = 0.66,
since there are two words in common, which are "sport" and "match" that are found both in
concept clusters extracted manually and algorithmically. However the word "game" is not in the
top three concept cluster yielded manually, so it decreases the success rate. In Document 2, the
success rate is 0.33, since only the word "attorney" is common amongst the three top concepts.
We also made comparison in a manner which compares the top 3, 5, 7, 10 and 15 concepts found
through the algorithm with all the concepts found manually and evaluated the success ratio. It
generally gave the highest success rates.
4.2 Results
Table 2 summarizes the accuracy results for different corpora. Window sizes 3, 5, 7 and 10 are
tested for comparison. For algorithm 1 (Simple Frequency Algorithm) the most successful results
are achieved through method, which uses 2-level structure taking frequency into account. All
values shown in Table 2 for algorithm 1 are that achieved by 2-level structure taking frequency
factor into account. For this algorithm, the sub-methods taking 3-level structure into account or
eliminating frequency factor gave unsuccessful results. It is seen, on average, comparison with
window size of three gives the highest success rate. The second algorithm gives, on average,
better evaluation results. It can clearly be seen that results for different corpora vary a lot which
shows that concept mining would be biased for documents concerning specific topics. The first
algorithm gives, on average, a success rate of 63.62%, whereas the second one gives a success
rate of 75.51%.

Table 2. Performance results for different corpora in terms of accuracy percentage.
Corpora Comparison
Window Size
Simple Frequency Algorithm Frequency and Context
Algorithm
Gazi
k = 3 58.40 69.30
k = 5 55.70 67.00
k = 7 54.70 64.30
k = 10 53.90 62.60
Sport
News
k = 3 57.40 55.40
k = 5 56.90 54.20
k = 7 56.00 53.40
k = 10 55.30 52.68
Forensic
News
k = 3 58.74 77.53
k = 5 56.81 71.72
k = 7 55.52 68.53
k = 10 55.18 67.79
Forensi
c
Decision
s
k = 3 76.81 95.74
k = 5 67.45 91.54
k = 7 63.21 84.86
k = 10 59.44 78.95
Table 3 shows, as an example, the results for Forensic Decisions corpus comparing all methods.
First Algorithm (2-levels, 1-0) is the method that fills the matrix with values 1 or 0 using 2-level
structure. If a column word is present the cell value in matrix is 1, if it is not present, cell value is
0. Frequency factor is overlooked in this sub-method. First Algorithm (2-levels, frequency) takes
frequency into account. First Algorithm (3-levels, coefficients) takes frequency into account using
3-level structure. This algorithm assigns different scores for different hierarchical levels of a
word. The second algorithm is the algorithm explained in Section 3.4.2.
5. CONCLUSION AND FUTURE WORK
In this work, we developed two novel algorithms which make use of the dictionary. So far,
generally WordNet has been preferred for its synset relations, especially hypernymy, in concept
mining, because the general idea of a word is generally related to its hypernym word. But only
taking a solitary synset relation into account may be insufficient, also other relationships may
give us an idea concerning a word, that is why we made use of a general language dictionary.
Table 3. Performance results of four methods for Forensic Decisions corpus with different
comparison windows sizes.
Comparison Window Size
Algorithms k = 3 k = 5 k = 7 k = 10
First Algorithm (2-
levels, 1-0)
76.81 67.45 63.21 59.44
First Algorithm (2-
levels, frequency)
68.84 65.43 66.72 63.92
First Algorithm (3-
levels,
coefficients)
64.81 60.73 57.87 55.72
Second Algorithm 95.74 84.77 78.58 73.11

Besides the use of a dictionary, we took the frequencies into account, because the more frequent a
word in the document, the higher contribution of this word to the general idea and concept of
documents amongst many. We also developed context-based algorithm which eliminates some of
the meaning text nouns in dictionary from probable candidate concepts set, because some of those
words would not be convenient in determining the general idea, that is concept, of a word. This
approach increased our success rate.
As a future work, we would improve the second algorithm, where we can take not only the word
and the noun that is most frequently found in its contexts in all corpus and present in its meaning
text in dictionary into account, but also the other nouns present in both meaning text and contexts.
We can give them coefficients in accordance with their frequencies instead of eliminating them.
Apart from the direction of this paper's algorithm, for concept mining we may also implement
clustering through all dictionary words, in accordance with the analogies between their meaning
texts. This would approach the whole dictionary word-set as a training data.
ACKNOWLEDGEMENT
This work was supported by the Boğaziçi University Research Fund under the grant number
5187, and the Scientific and Technological Research Council of Turkey (TÜB TAK) under the
grant number 110E162. Cem Rıfkı Aydın is supported by TÜB TAK B DEB 2210. We thank
Hasim Sak for providing us the tools for pre-processing and morphological analyses.
REFERENCES
[1] Marie-Francine, Moens & Roxana, Angheluta, (2003) "Concept Extraction from Legal Cases: The
Use of a Statistic of Coincidence", International Conference on Artificial Intelligence and Law,
ICAIL, ACM
[2] Vance, Faber & Judith G., Hochberg & Patrick M., Kelly, Timothy R. Thomas, James M. White,
(1994) "Concept Extraction – a data-mining technique", Los Alamos Science
[3] Nuala A., Bennett & Qin, He & Conrad T. K., Chang & Bruce R., Schatz, (1999) "Concept Extraction
in the Interspace Prototype", Technical Report, Dept. of Computer Science, University of Illinois at
Urbana-Champaign, Champaign, IL
[4] David M., Pennock & Kushal, Dave & Steve, Lawrence, (2003) “Mining the Peanut Gallery: Opinion
Extraction and Semantic Classification of Product Reviews”, Twelfth International World Wide Web
Conference (WWW’2003), ACM
[5] H., Liu & P., Singh, (2004) "ConceptNet - a practical commonsense reasoning toolkit", BT
Technology Journal, Vol 22, No 4
[6] Meryem, Uzun, (2011) "Developing a concept extraction system for Turkish", International
Conference on Artificial Intelligence, ICAI'11
[7] Paul M., Ramirez & Chris A., Mattmann, (2004) "ACE: Improving Search Engines via Automatic
Concept Extraction", Information Reuse and Integration
[8] Zhang, Chengzhi & Wu, Dan, (2008) "Concept Extraction and Clustering for Topic Digital Library
Construction", IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent
Technology
[9] Loulwah, AlSumait & Daniel, Barbar´a & Carlotta, Domeniconi, (2008) "On-Line LDA: Adaptive
Topic Models for Mining Text Streams with Applications to Topic Detection and Tracking", ICDM
'08 Proceedings of the 2008 Eighth IEEE International Conference on Data Mining
[10] Zakaria, Elberrichi & Abdelattif, Rahmoun & Mohamed Amine, Bentaalah, (2008) "Using WordNet
for Text Categorization", The International Arab Journal of Information Technology, Vol. 5, No. 1
[11] Haşim, Sak & Tunga, Güngör & Murat, Saraçlar, (2008) “Turkish Language Resources:
Morphological Parser, Morphological Disambiguator and Web Corpus”, GoTAL 2008, vol. LNCS
5221, pp. 417-427, Springer.
[12] Haşim, Sak & Tunga, Güngör & Murat, Saraçlar, (2007) “Morphological disambiguation of Turkish
text with perceptron algorithm”, CICLing 2007, vol. LNCS 4394, pp. 107-118.

AUTHORS
Cem Rıfkı Aydın received his B.Sc. degree from Bahçeşehir University, Istanbul,
Turkey, in Department of Computer Engineering. He is currently an M.Sc. student in
Department of Computer Engineering at Boğaziçi University, and awarded with
TÜB TAK BIDEB scholarship throughout his M.Sc. studies. His areas of interest include
natural language processing, artificial intelligence applications, pattern recognition, game
programming, information retrieval and genetic algorithms.
Ali Erkan received his B.Sc. and M.Sc. degrees from Department of Industrial
Engineering, Bilkent University, Ankara, Turkey, and he received M.Sc. degree in
Software Engineering from Department of Computer Engineering, Boğaziçi University,
Istanbul, Turkey, in 2010. He is currently studying for Ph.D. degree at Department of
Computer Engineering, Boğaziçi University, Istanbul, Turkey. His research interests
include natural language processing, machine learning, pattern recognition,
bioinformatics and statistics.
Tunga Güngör received his Ph.D. degree from Department of Computer Engineering,
Boğaziçi University, Istanbul, Turkey, in 1995. He is an associate professor in the
department of Computer Engineering, Boğaziçi University. His research interests
include natural language processing, machine translation, machine learning, pattern
recognition, and automated theorem proving. He published about 60 scientific articles,
and participated in several research projects and conference organizations.
Hidayet Takçı is an academician at Cumhuriyet University, Sivas, Turkey. He studies
on some fields such as data mining, text mining, machine learning and security.
Hitherto he has lectured some courses such as data mining and applications, text
mining, neural nets, etc. In addition, he has many papers and projects in the field of text
classification and text mining, information retrieval and web security.

DICTIONARY-BASED CONCEPT MINING: AN APPLICATION FOR TURKISH

Recommended

More Related Content

What's hot (20)

Similar to DICTIONARY-BASED CONCEPT MINING: AN APPLICATION FOR TURKISH (20)

More from cscpconf (20)

Recently uploaded (20)

DICTIONARY-BASED CONCEPT MINING: AN APPLICATION FOR TURKISH