Available via license: CC BY-SA 4.0
Content may be subject to copyright.
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Santi Purwaningrum: Synonym Recognition Influence in ... Volume 12 Number 3 August 2023
Synonym Recognition Influence in Text Similarity
Detection Using Winnowing and Cosine Similarity
Santi Purwaningrum1, Agus Susanto2, Ari Kristiningsih3
1,2 Department of Informatics and Business Engineering Politeknik Negeri Cilacap, Cilacap 53212 INDONESIA (tel.: 0282-533329; fax: 0274-4321982, email:
1santi.purwaningrum@pnc.ac.id, 2agussusanto@pnc.ac.id)
3 Department of Agroindustry Product Development, Politeknik Negeri Cilacap, Cilacap 53212 INDONESIA (tel.: 0282-533329; fax: 0274-4321982, email:
3ari.kristiningsih@pnc.ac.id)
[Received: 19 January 2023, Revised: 4 July 2023]
Corresponding Author: Santi Purwaningrum
ABSTRACT — Plagiarism is an act of imitating, quoting and even copying or acknowledging the work of others as one’s
own work. A final project is one of the mandatory requirements for students to complete learning at college. It must be
written by the students based on their own ideas. However, there is much plagiarism because it is easy to carry out just by
simply copying the text of other people’s ideas and then pasting it into a worksheet and admitting that the ideas are theirs.
In addition, replacing some words in other people’s sentences with their own language style without properly acknowledging
the original source of the quotation is also an act of plagiarism. A manual check for the final project also becomes an issue
for the final project coordinator, i.e., it needs high accuracy and a relatively long time to check the plagiarism in the final
project document. Therefore, implementing plagiarism detection mechanisms is necessary to mitigate the escalation of
plagiarism occurrences. In response to those matters, this study aims to design a system capable of identifying textual
similarities by focusing on sentences containing synonymous words. One of the used algorithms is synonym recognition,
which detects words that possess synonymous meanings by comparing each term with the entries in a dictionary. The
synonym recognition is combined with the winnowing method, functioning as a fingerprint-based text weighting. After the
weight of each document is obtained, the similarity level between documents is calculated with the cosine similarity
algorithm. The inclusion of synonym recognition in conjunction with the winnowing weighting method resulted in a notable
gain of 3.11% in the average similarity scores for title and abstract detection, compared to the absence of synonym
recognition. The results show that the used algorithms are accurate with accuracy testing and root mean squared error
(RMSE).
KEYWORDS — Synonym Recognition, Winnowing, Cosine Similarity, Plagiarism.
I. INTRODUCTION
A final project is a compulsory prerequisite for students
pursuing diploma-level courses in order to obtain an associate
expert and applied bachelor’s degree. It is taken subsequent to
the completion of an industrial internship. Therefore, the
students are expected to derive potential topics for their final
projects from the case studies encountered during their
internship experience. Case studies are not required to be
conducted at the internship industry site. Students are allowed
to find another case study site with a topic that suits the needs
of the site.
A system implementation is not a big deal for vocational
students. Nevertheless, the process of drafting the final project
report has certain challenges for students, particularly when it
comes to articulating ideas or concepts so that they are not
considered plagiarism. Plagiarism may occur either
deliberately or unintentionally. Due to a lack of reading or
difficulty in finding references, the concepts or ideas that will
be used as the topic of the final project can be the same or have
been made before [1], [2].
The process of uploading titles, abstracts, proposals,
journals, posters, and final project reports to the system is
carried out by the final project coordinator. However, the final
project coordinator only uploads without paying attention to
each title, or abstract, to the student’s final project report.
Consequently, it results in a failure to identify numerous
instances of similarity across final project reports, both across
different classes and within the same class. On the other hand,
the examination process by the final project coordinators
requires a considerable amount of time and high concentration
because they have to check the final project one by one.
Furthermore, alongside the aforementioned issue, the copy-
paste phenomenon has also developed. It is conducted by only
substituting words containing synonyms without citing their
sources. Early detection of similarities in titles and abstracts of
scientific papers is needed to reduce acts of plagiarism arising
from the students’ lack of creativity.
Many researchers are interested in conducting research on
plagiarism cases. The methods used for plagiarism detection
text weighting include the winnowing algorithm, the Boyer-
Moore algorithm, and the Rabin-Karp algorithm (or also
known as Karp-Rabin) [3], [4]. The winnowing algorithm has
been used in a study to detect the similarity level of thesis titles
submitted compared with preexisting thesis titles [5]. The
findings of this study indicate that with the score of n-gram =
3, window = 3, and prime number = 2, a similarity score of
73.86% has been achieved. This score suggests a significant
degree of similarity between the examined titles. With the input
of n-gram = 7, window = 9, and prime number = 2, a similarity
score of 19.82% has been achieved, indicating a moderate
plagiarism level [5].
A study describes an overview of the definition of
plagiarism, tools to detect plagiarism, comparison matrices,
obfuscation methods, data sets used for comparison, and the
algorithm types. It is concluded that three algorithms that are
often used in plagiarism detection are running Karp-Rabin
greedy string tiling (RKR-GST), the winnowing algorithm, and
the implementation process in the tokenization [6].
In the weighting process, another study uses the Rabin-
Karp algorithm combined with the synonym recognition
approach to identify words that have been modified into word
EN-219
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Volume 12 Number 3 August 2023 Santi Purwaningrum: Synonym Recognition Influence in ...
forms with similar meanings. The Rabin-Karp algorithm
involves a step known as the rolling hash, which necessitates
the provision of a base score for the weighting process. Not all
numbers can be used in the base score in this algorithm. The
rationale behind this phenomenon is that in certain cases, an
incorrect base score yields an identical hash score to another
hash score corresponding to a different word [7].
A study aiming at reducing bias in a description answer
scoring system using the cosine similarity method with term
frequency-inverse document frequency (TF-IDF) weighting
and word matching by adding linear regression has been
conducted [8]. Frequently, the assessment system provides
outputs with higher scores than manual assessment by lecturers,
indicating the presence of bias in the scoring. Therefore, linear
regression is used to reduce bias to prevent the score of the
description answer generated by the system from
overestimating or underestimating the manual score given by
the lecturer. In the linear regression process, the system score
refers to the answer key that has been made, which is adjusted
to the lecturer’s assessment. Therefore, lecturers are expected
to give an objective assessment [8].
Previous study has used the synonym recognition approach
to identify terms containing synonyms and have used cosine
similarity to determine similarity scores. The text weighting
process of each compared document has used term frequency
(TF), and root mean squared error (RMSE) has been used to
test the accuracy. There are several stages in RMSE testing,
namely 20, 50, 100, and 116 data. The highest RMSE score
achieved in answer number 1 corresponding to a sample size of
20 students, which was 2.07, and the lowest was associated
with a sample size of 50 students, which was 6.16. In answer
number 2, the highest RMSE score of 8.94 was obtained in the
group of 50 students, and the lowest score at 8.00 was in the
group of 20 students. The accuracy testing has a relatively high
error score because the words in the student’s answers are
included in the stopword dictionary, resulting in their loss in
the filtering process. Therefore, in further research, it is
expected that the similarity score between the two texts can be
increased [9].
State of the art of previous research has discussed text
similarity detection with many methods used for weighting a
text, including Rabin-Karp, winnowing, Smith-Waterman and
Manber, and TF [10]–[12]. In addition, algorithms used to find
similarity scores, including Jaccard, dice similarity, neural
network, and cosine similarity, are considered to have better
similarity rates [13]. Reference [14] has investigated the
collaborative use of the cosine similarity method with other
weighting algorithms, such as artificial neural network (ANN)
and support vector machine (SVM). However, the combination
of winnowing method has not been explored in the study.
The main objective of this study is to detect the similarity
of titles and abstracts in students’ final assignments earlier by
focusing on words with synonyms. The method used to
calculate the similarity level between documents is cosine
similarity. In the similarity method, the weighting process is
usually conducted using TF-IDF to determine the weight score
of each document. However, in this study, the weighting score
of each document uses the winnowing algorithm, which is a
fingerprint-based weighting. This algorithm is combined with
a synonym recognition method used to detect words in
documents compared with a dictionary of synonyms. With the
combination of synonym recognition, final assignments with
synonym-based paraphrases are expected to be detected. The
performance of each combined method is measured using the
accuracy parameter, while RMSE is used to determine the error
score of the score recommended by the system against the
actual score [15].
II. METHODOLOGY
In this stage, the impact of the synonym recognition
algorithm in conjunction with the winnowing method and
cosine similarity, is examined to detect the similarity between
titles and abstracts of final projects focusing on words
containing synonyms.
Figure 1 illustrates the stages in the research, from the initial
process, i.e., data collection, initial data processing, algorithm
implementation, i.e., synonym recognition for detection of text
containing synonyms and winnowing algorithm for weighting
a text in the document (fingerprint), to the final process that
obtains the percentage of similarity score results from the
compared documents and the accuracy score of the system
using winnowing, synonym recognition, and cosine similarity
Figure 1. Research stages.
EN-220
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Santi Purwaningrum: Synonym Recognition Influence in ... Volume 12 Number 3 August 2023
methods. Accuracy and RMSE tests were conducted to assess
the influence of synonym recognition on the system with a
combination of winnowing and cosine similarity weighting.
A. DATA COLLECTION
The initial step of the research process involves determining
the methodology and location for data acquisition.
The data acquisition involves collecting primary data, i.e., data
directly obtained from the case study site, i.e., the Department
of Informatics Engineering at Politeknik Negeri Cilacap.
B. DATA PROCESSING
The dataset is in the form of titles and abstracts of
Indonesian final assignments of the students of Politeknik
Negeri Cilacap majoring in Informatics Engineering class of
2021 and 2022. There were 182 final project reports which
were then categorized into three sections based on the
respective theme of the final project. Specifically, there were
42 reports related to augmented reality (AR), 63 reports related
to decision support systems, and 77 reports related to e-
commerce-based information systems. The average title length
is between 12 to 15 words per title, while the average abstract
length is 350 to 500 words per abstract.
In this process the data that had to be uploaded were
prepared, i.e., the title and abstract text documents of the final
project, so that they could be processed in the text similarity
measurement method. The data were obtained from the final
project information system of the Department of Informatics
Engineering of Politeknik Negeri Cilacap. The collected data
were in the form of a complete final project report, which was
subsequently organized into separate documents based on their
titles and abstracts. Examples of input data were in the form of
final project report data organized into titles and abstracts and
saved in the form of PDF files.
Table I presents examples of each document with titles and
abstracts used as datasets for model testing. Documents from
each title and abstract were compared and searched for words
containing synonyms, and then a text weighting generating a
fingerprint was made. Subsequently, the similarity score of
each document’s fingerprint was calculated.
C. PREPROCESSING
Preprocessing is the initial stage of the text mining process.
It serves to transform the unstructured text data into structured
text data [16], [17]. In preprocessing, a series of steps were
taken to remove parts of text that were not needed in a
document because they would become noise in the subsequent
process. Preprocessing is divided into three distinct stages:
tokenizing, filtering, and stemming [18], [19].
Tokenizing is the process of separating individual words in
a document, commonly known as tokens [20]. This process
also converted all uppercase letters into lowercase letters. In
addition, tokenizing also removed all punctuation marks,
numbers, and symbols, as they have no unique score because
they were not related to the string to be processed.
The filtering process serves to remove meaningless words.
The meaningless words are commonly referred to as stopwords.
Some examples of stopwords are “juga,” “dan,” “untuk,” and
“adalah” [21]. It is necessary to delete these stopwords because
if conjunctions frequently appear in a sentence, the text
similarity percentage is very high, and it interferes with the
accuracy of the text similarity method [22].
The stemming process serves to remove affixes on a word
in a text document, so the word taken is the root word. It was
conducted to facilitate the subsequent process. Some examples
of affixes are “mem-,” “-kan,” “ber-,” “-pun,” and “me-an” [23].
The obtained root words that were used as tokens within each
text content in order to enhance efficiency and accuracy in
syntactic matching. For instance, document 1 contained the
word “belajar” and document 2 contained the word “mengajar.”
After the stemming process, the words “belajar dan mengajar”
TABLE I
EXAMPLES OF INPUT DATA
Title
Abstract
Aplikasi Media Pembelajaran
Pengenalan Aksara Jawa
Menggunakan Augmented
Reality di Smartphone
Android (Studi Kasus: SDN
07 Adipala)
Teknologi mengalami kemajuan yang pesat dan mempengaruhi dunia pendidikan juga berbagai inovasi telah
dilakukan untuk menunjang kegiatan belajar mengajar. Proses belajar mengajar bahasa Jawa yang dilakukan
guru saat ini masih menggunakan cara ceramah juga membosankan inovasi teknologi yang dapat digunakan
dan dapat membantu siswa untuk belajar bahasa Jawa salah satunya adalah Augmented Reality (AR).
Penelitian ini memiliki rumusan masalah yaitu bagaimana merancang dan membuat aplikasi Augmented
Reality (AR). Pembelajaran pengenalan aksara Jawa di smartphone Android Penulis menggunakan metode
literature dan observasi lapangan sebagai metode pengumpulan data dan sebagai metode pengembangan
sistem penulis menggunakan MDLC (Multimedia Development Life Cycle). Kata kunci: Augmented
Reality, MDLC, Smartphone Android, Aksara Jawa, Teknologi, Pendidikan, Belajar Mengajar
Aplikasi Pembelajaran
Pengenalan Rangka Manusia
Menggunakan Augmented
Reality (AR). Berbasis
Smartphone Android (Studi
Kasus: SD Negeri Jepara
Kulon 01, Binangun)
Pemahaman belajar siswa dapat meningkat dengan tersedianya media belajar yang menarik. Media belajar
yang menarik dapat memudahkan pengajar dalam menyampaikan materi. Penelitian ini diambil dari studi
lapangan dengan guru dan juga siswa sekolah dasar di SDN Jepara Kulon 01, Binangun. Kurangnya alat
peraga rangka manusia dan metode pembelajaran di SDN Jepara Kulon 01, Kecamatan Binangun yang
kurang menarik karena hanya tersedianya gambar rangka manusia 2D membuat siswa merasa bosan,
sehingga dapat menjadi salah satu penyebab kurangnya pemahaman siswa terhadap materi yang diajarkan.
Penelitian ini dilakukan dengan menerapkan teknologi Augmented Reality (AR) yang diimplementasikan
dalam bentuk aplikasi berbasis android. Teknologi Augmented Reality (AR) merupakan perpaduan antara
2D, 3D, dan dunia nyata yang digabung dalam satu objek dengan satu teknologi yang dapat digunakan
sebagai media pembelajaran di bidang multimedia. Aplikasi ini dikembangkan menggunakan metode
Multimedia Development Life Cycle (MDLC). Hasil kuisioner menunjukan aplikasi ini dapat membantu
guru dan siswa dalam mengenalkan serta memahami bagian - bagian rangka manusia dan juga menarik
minat belajar dibanding menggunakan buku dengan persentase yang didapat 28% menyatakan setuju dan
72% menyatakan sangat setuju dengan aplikasi ini. Kata Kunci: android, Augmented Reality (AR), media
pembelajaran, rangka manusia
EN-221
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Volume 12 Number 3 August 2023 Santi Purwaningrum: Synonym Recognition Influence in ...
changed to “ajar” because “ajar” is the root word of “belajar”
and “mengajar.”
D. SYNONYM RECOGNITION
Synonym recognition is one of the methods of detecting
text plagiarism instances with the synonym approach [7].
Wordnet is a lexical database intended for use under program
control [24]. The database contains various types of vocabulary
organized according to their synonyms according to a lexical
concept. A word with two or more meanings of the same word
is called a synonym. In the synonym recognition process, a list
of words that have synonym meanings is required. This word
list was obtained from the Great Dictionary of Indonesian
Language (Kamus Besar Bahasa Indonesia, KBBI) online
website and kamusbesar.com [15].
Figure 2 explains the stages of the synonym recognition
method, from input to output. The preprocessed text undergoes
the synonym recognition stage in order to detect plagiarism
activity in a text document. This stage compares one document
with another by detecting words that contain synonyms. The
synonym recognition process compares the word contained in
the document with the thesaurus contained in the database.
The winnowing algorithm alone cannot detect plagiarism
of a document if the words in the document have experienced
word modification, yet the word’s meaning remains the same.
Therefore, a synonym recognition algorithm is required to
overcome the word modification problem. The synonym
recognition process changes all words diagnosed as synonyms
of a word contained in the synonym dictionary which is
considered as the main word [25].
E. WINNOWING
In the winnowing process, data are processed from string to
numeric data. In the process, the user must enter parameter
scores, namely k-gram, hash, and window. To determine the k-
gram, hash, and window scores, parameter testing must be
conducted with the sample data to obtain a good similarity
score.
The input of the winnowing algorithm is the root word text
from the data preprocessing results. Then, the output is a
collection of hash scores. A hash score is a numerical score
formed from the American Standard Code for Information
Interchange (ASCII) table calculation of each character. The
hash score can also be referred to as a fingerprint, which is used
as an indicator to compare the similarity between text
documents [26].
Before starting the winnowing process, a synonym
recognition process was performed, which was the processing
of preprocessed root words compared with the synonym
dictionary. Next, the data entered the k-gram process, which
was used to obtain a new set of strings from the old set of strings,
with a k-gram score determined by the user. The set of strings
was grouped into a new set of strings, which was a
concatenation of the initial strings, with the length of the
concatenated strings being k.
Then, a rolling hash process was carried out, which served
to generate the hash score of each gram that had been formed.
The conversion of a string of characters into a score or code
that then becomes the sign of the string of characters is called
hashing and the resulting score can be referred to as a hash
score. The obtained hash score is already numerical data. This
process could be conducted using the following formula.
(1)
where c is the ASCII score of each character, l is the length of
the string, and b is the user-defined hash base score. From this
process, the hash score of each gram will be obtained.
The next step was to form a window on the hash score
results of each previous gram. The window size was also
determined by the user. The window process was carried out
by determining the hash score of each window to be used as a
document fingerprint and if there were the same hash score, the
rightmost hash score would be selected. Then, the final process
of the winnowing algorithm was to determine the smallest hash
score from each window to be used as a fingerprint on a
document.
F. COSINE SIMILARITY
Cosine similarity is one method that can be used to calculate
the text similarity level in a sentence or document. Cosine
similarity has a high accuracy score to determine the similarity
level because it does not affect the length of a word or sentence
in a compared document [27], [28]. It was calculated with the
following formula.
(2)
where X∩Y is the number of words contained in document X
and contained in document Y, |X| is the number of words
contained in document X, and |Y| is the number of words
contained in document Y.
G. ACCURACY
Accuracy is used to provide an assessment of predictive
results which corresponds to the actual data. The higher the
accuracy score, the more accurate or better the performance of
the used method. The accuracy function is to determine the
level of accuracy of recommendations for methods used in a
system. The accuracy formula is written as follows [29]–[31].
(3)
where TN is the number of negative data detected correctly
(true negative), FP is negative data detected as positive data
(false positive), TP is positive data detected correctly (true
positive), and FN is positive data detected as negative data
(false negative).
Start
Wordnet
synonym data
list
Result text
from
preprocessing
Comparing Wordnet data
with text preprocessing
results
Is the preprocessing
result text the same as
data on Wordnet?
Change the whole
text according to
Wordnet data
End
Yes
Return text
No
Figure 2. Synonym recognition process.
EN-222
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Santi Purwaningrum: Synonym Recognition Influence in ... Volume 12 Number 3 August 2023
H. ROOT MEAN SQUARE ERROR (RMSE)
RMSE is used to determine the performance of the system
created. RMSE is the square root of the average squared
difference between predictions and actual observations [32].
The greater the similarity score between the system and manual
calculation, the more suitable the method used in the system.
. (4)
III. RESULTS AND DISCUSSION
Figure 3 describes the process of algorithms used in text
similarity detection in the title and abstract of a final project
that focuses on text containing synonyms. It is necessary to test
the scores of k-gram, rolling hash, and window parameters to
obtain maximum results in similarity detection in titles and
abstracts in the weighting process using the winnowing method.
The parameter testing was carried out using the example of two
titles and abstracts of the decision support system with
parameter scores as shown in Table II.
Table II presents the results of the winnowing parameter
test. Each parameter was tested with different scores using a
dataset of two titles and a final project abstract on the topic of
decision support systems. The best parameter results in text
weighting were obtained with the score of k-gram = 2, hash =
5, and window = 7, with an average title similarity score of
55.84% and abstract similarity of 59.42%. Therefore, these
parameters were used in this study to analyze the influence of
the synonym recognition algorithm on the title and abstract of
the final project in the text weighting of the winnowing
algorithm. Those parameters were combined with the cosine
similarity method to determine the similarity between
documents.
The initial process was carried out by comparing two text
documents, each of which was preprocessed. The example of
preprocessing results in title 1 is “Aplikasi media belajar kenal
aksara Jawa Augmented Reality smartphone android studi
kasus SDN Adipala,” and the results of preprocessing title 2 is
“Aplikasi belajar kenal rangka manusia Augmented Reality
basis smartphone android studi kasus SD Negeri Jepara Kulon
Binangun.”
After preprocessing was complete, the next step was the
process of detecting words that contain synonyms. This
synonym recognition method can add, delete, and change
synonym words contained in the synonym dictionary database.
Strings in the title or abstract were compared or matched with
the synonym dictionary. A synonym dictionary is a collection
of synonym-based words in Indonesian. The synonym
dictionary focuses on three major topics, namely on AR, e-
commerce-based information systems, and decision support
systems. If there were title and abstract words that were the
same as the synonym dictionary, the string contained in the title
or abstract was replaced with the string as in the synonym
dictionary. For instance, after going through preprocessing and
Start
Document 1
Document 2
Case folding
Filtering
Stemming
Tokenizing
Preprocessing
Preprocessing
text
Synonym
Recognition
Term weighting
with winnowing
Fingerprint
of each
document
Document 1 Ո Document 2 /
Cosine
similarity
End
Rolling hash
Window
K-gram
Figure 3. Flowchart of title and abstract similarity detection system.
TABLE II
PARAMETER TESTING
K-gram
Hash
Window
Title
Similarity
(%)
Abstract
Similarity
(%)
5
7
2
24.50
26.21
7
2
5
18.81
19.37
2
5
7
55.84
59.42
2
7
5
51.74
57.33
7
5
2
19.44
20.50
5
2
7
25.55
27,67
EN-223
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Volume 12 Number 3 August 2023 Santi Purwaningrum: Synonym Recognition Influence in ...
synonym recognition, title 1 becomes “Aplikasi media belajar
kenal aksara Jawa Augmented Reality smartphone android
studi kasus SDN Adipala” and title 2 becomes “aplikasi tatar
kenal rangka manusia Augmented Reality basis smartphone
android studi kasus SD Negeri Jepara Kulon Binangun.” There
is one string of document 1 and 2 titles identified as having
words containing synonyms found in the synonym database
dictionary, namely “belajar” to “tatar.”
After synonym recognition, the next process was text
weighting of each compared document. The weighting process
using the winnowing method was carried out by finding the
fingerprint score of each document. The initial weighting
process was to enter the parameter scores of k-gram, rolling
hash, and window.
The example of the winnowing process results by entering
k-gram score of 2 in title 1 is “ap, pl, li, ik, ka, as, si, it, ta, at,
ta, ar, rk, ke, en, na, al, lr, ra, an, ng, gk, ka, am, ma, an, nu, us,
si, ia, aa, au, ug, gm, me, en, nt, te, ed, dr, re, ea, al, li, it, ty, yb,
ba, as, si, is, ss, sm, ma, ar, rt, tp, ph , ho, on, ne, ea, an, nd, dr,
ro, oi, id, ds, st, tu, ud, di, ik, ka, as, su, us, ss, sd, dn, ne, eg, ge,
er, rj, je, ep, pa, ar, ra, ak, ku , ul, lo, on, nb, bi, in, na, an, ng,
gu, un”. The second process in winnowing is rolling hash. For
example, the entered hash score was 5, then the result of the
title 1 in the first string “ap”, the string “a” in the ASCII table
has a score of 97, and “p” has a score of 112. Therefore, the
calculation in the rolling hash formula is (97×5(2-1))+
(112×5(1-1)) = 485+112 = 597. Therefore, the overall rolling
hash result of title 1 is [ap : 597, pl : 668, li : 645, ik : 632, ka :
632, as : 600, si : 680, it : 641, ta : 677, at : 601, ta : 677, ar :
599, rk : 677, ke : 636, en : 615, na : 647, al : 593, lr : 654, ra :
667, an : 595, ng : 653, gk : 622, ka : 632, am : 594, ma : 642,
an : 595, nu : 667, us : 700, si : 680, ia : 622, aa : 582, au : 602,
ug : 688, gm : 624, me : 646, en : 615, nt : 666, te : 681, ed :
605, dr : 614, re : 671, ea : 602, al : 593, li : 645, it : 641, ty :
701, yb : 703, ba : 587, as : 600, si : 680, is : 640, ss : 690, sm :
684, ma : 642, ar : 599, rt : 686, tp : 692, ph : 664, ho : 631, on :
665, ne : 651, ea : 602, an : 595, nd : 650, dr : 614, ro : 681, oi :
660, id : 625, ds : 615, st : 691, tu : 697, ud : 685, di : 605, ik :
632, ka : 632, as : 600, su : 692, us : 700, ss : 690, sd : 675, dn :
610, ne : 651, eg : 608, ge : 616, er : 619, rj : 676, je : 631, ep :
617, pa : 657, ar : 599, ra : 667, ak : 592, ku : 652, ul : 693, lo :
651, on : 665, nb : 648, bi : 595, in : 635, na : 647, an : 595, ng :
653, gu : 632, un : 695].
The third step in winnowing is the window. The window
process is the same as k-gram. However, the processed data are
already in the form of numbers from the previous rolling hash
process. For example, the results of title 1 with the window
input score of 5 were {597, 668, 645, 632, 632, 600, 680}, {668,
645, 632, 632, 600, 680, 641}, {645, 632, 632, 600, 680, 641,
677}, {632, 632, 600, 680, 641, 677, 601}, {632, 600, 680, 641,
677, 601, 677}, {600, 680, 641, 677, 601, 677, 599}, {680, 641,
677, 601, 677, 599, 677}, {641, 677, 601, 677, 599, 677, 636},
{677, 601, 677, 599, 677, 636, 615} … {595, 635, 647, 595,
653, 632, 695}.
The last step of the winnowing process is to find the
minimum score of each window. If there is the same minimum
score from each window, only one score is taken, which is the
rightmost score. It is done to avoid redundancy, which affects
accuracy and reduces processing time. The minimum score of
each window is referred to as the fingerprint. The results of the
fingerprint score on title 1 after the winnowing process was
complete were [597, 600, 599, 593, 594, 582, 602, 605, 587,
595, 614, 615, 608, 592].
After the fingerprint score of each document was obtained,
the next process was to find the similarity score of the
compared text documents using the cosine similarity method.
This process was carried out by dividing the same number of
fingerprints from the compared documents by the sum of the
square root of the fingerprints of each compared document.
The process of using synonym recognition and without
synonym recognition combined with the weighting method
using winnowing and the similarity search method using cosine
similarity yielded the following results. In the decision support
system topic, the average score of the title similarity difference
in 2021 and 2022 is 6.8%, in the e-commerce-based
information system topic by 1.11%, and in the AR topic by
6.07%. In addition, the average scores of abstract similarity
differences on the topics of decision support systems, e-
commerce-based information systems, and AR topics are
1.41%, 2.75%, and 0.52%, respectively. The average similarity
scores of all topics in titles and abstracts compared using
synonym recognition combined with weighting method using
winnowing and similarity search method using cosine
similarity are 59.32% and 55.19%, while the average scores of
all topics in titles and abstracts without synonym recognition
combined with weighting method using winnowing and
similarity search method using cosine similarity are 54.67%
and 53.63%.
Figure 4 highlights the average score of title and abstract
similarity in 2021 and 2022 as a consequence of using and not
using synonym recognition. The data presented in the figure
shows that in 2021 and 2022 the average similarity score is
increased by 2.73%.
The title and abstract similarity detection system displays
the results of the similarity percentage of each title and abstract
compared. The accuracy and RMSE scores are sought to
determine the success rate of the algorithms in the system used.
The influence of using synonym recognition combined with a
weighting method using winnowing and a similarity search
method using cosine similarity on title and abstract similarity
detection is compared with manual calculations, yielding an
average accuracy rate of 81.05%. The accuracy score was
obtained by dividing the TP score of 147.57% by the amount
of data. In addition, the average RMSE score is 4.38%. This
score is obtained from the actual data of 63.7% minus the
forecasting result score of 2.01%, followed by calculating the
the square root of the difference and dividing it by the amount
of data.
IV. CONCLUSION
Based on the results of the conducted tests, it can be inferred
that the smaller k-gram parameter score is sufficient to affect
Figure 4. Influence of synonym recognition on the similarity score.
EN-224
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Santi Purwaningrum: Synonym Recognition Influence in ... Volume 12 Number 3 August 2023
the similarity level scores between documents. Conversely, the
input scores for hash and window do not significantly affect the
similarity level score between documents. The highest
parameter score input in the text weighting similarity score is
k-gram = 2, hash = 5, and window = 7. The influence of
synonym recognition on text weighting using the winnowing
algorithm and calculating the similarity score using cosine
similarity demonstrates an average increase in the similarity
score of titles and abstracts by 3.11% compared to synonym
recognition without the winnowing algorithm. The rise in the
similarity score of titles and abstracts is derived from the
average score of titles and abstracts using winnowing
weighting combined with synonym recognition and calculating
the similarity score using cosine similarity, amounting to
57.26%, minus the average score without synonym recognition,
amounting to 54.15%. The addition of the synonym recognition
method combined with winnowing weighting and calculation
of similarity score using cosine similarity has the potential to
increase the similarity score, especially in synonym-based text.
Testing the influence of the synonym recognition method
combined with the winnowing algorithm and cosine similarity
yielded an accuracy score of 80.97% and RMSE of 26.14%.
These results are obtained from the dataset of the final project
conducted by students of Politeknik Negeri Cilacap. This
dataset was selected for modeling purposes, as using other
datasets may lead to the possibility of getting different results.
CONFLICT OF INTEREST
The authors declare that there are no conflicts of interest in
relation to the research and development of this paper.
AUTHOR CONTRIBUTION
Conceptualization, Santi Purwaningrum and Agus Susanto;
methodology, Santi Purwaningrum and Agus Susanto;
software, Santi Purwaningrum; validation, Santi
Purwaningrum and Agus Susanto; formal analysis, Santi
Purwaningrum and Agus Susanto; formal analysis, Santi
Purwaningrum; investigation, Santi Purwaningrum; formal
analysis, Santi Purwaningrum; resources, Santi Purwaningrum;
data curation, Santi Purwaningrum; writing-original drafting,
Santi Purwaningrum; writing-reviewing and editing, Ari
Kristiningsih; visualization, Santi Purwaningrum; supervision,
Santi Purwaningrum.
ACKNOWLEDGMENT
The authors would like to express their gratitude to
Politeknik Negeri Cilacap for their support in the form of
funding grants, which made this study possible. Additionally,
the authors would like to extend their appreciation to all
individuals who contributed to the successful execution of this
study.
REFERENCES
[1] M.H.P. Swari, C.A. Putra, and I.P.S. Handika, “Plagiarsm Checker pada
Sistem Manajemen Data Tugas Akhir,” J. Sains, Inform., Vol. 7, No. 2,
pp. 192–201, Nov. 2021, doi: 10.34128/jsi.v7i2.338.
[2] M.H.P. Swari and C.A. Putra, “Sistem Manajemen Data Skripsi (Studi
Kasus: Perpustakaan Fakultas Ilmu Komputer UPN “Veteran” Jawa
Timur),” J. Pendidik. Teknol., Kejuru., Vol. 17, No. 2, pp. 198–209, Jul.
2020, doi: 10.23887/jptk-undiksha.v17i2.25436.
[3] F.E. Kurniawati and W.M. Pradnya, “Implementasi Algoritma
Winnowing pada Sistem Penilaian Otomatis Jawaban Esai pada Ujian
Online Berbasis Web,” J. Tek. Komput. AMIK BSI, Vol. 6, No. 2, pp.
169–175, Jul. 2020, doi: 10.31294/jtk.v6i2.7838.
[4] I. Ahmad, R.I. Borman, G.G. Caksana, and J. Fakhrurozi, “Implementasi
String Matching dengan Algoritma Boyer-Moore untuk Menentukan
Tingkat Kemiripan pada Pengajuan Judul Skripsi/TA Mahasiswa (Studi
Kasus: Universitas XYZ),” SINTECH (Sci., Inf. Technol. J.), Vol. 4, No.
1, pp. 53–58, Apr. 2021, doi: 10.31598/sintechjournal.v4i1.699.
[5] N. Alamsyah and M. Rasyidan, “Deteksi Plagiarisme Tingkat Kemiripan
Judul Skripsi pada Fakultas Teknologi Informasi Menggunakan
Algoritma Winnowing,” Technologia, Vol. 10, No. 4, pp. 197–201, Oct.-
Dec. 2019, doi: 10.31602/tji.v10i4.2361.
[6] M. Novak, M. Joy, and D. Kermek, “Source-Code Similarity Detection
and Detection Tools Used in Academia: A Systematic Review,” ACM
Trans. Comput. Educ., Vol. 19, No. 3, pp. 1-37, May 2019, doi:
10.1145/3313290.
[7] N.P. Putra and Sularno, “Penerapan Algoritma Rabin-Karp dengan
Pendekatan Synonym Recognition Sebagai Antisipasi Plagiarisme pada
Penulisan Skripsi,” J. Teknol., Sist. Inf. Bisnis, Vol. 1, No. 2, pp. 130–140,
Jul. 2019, doi: 10.47233/jteksis.v1i2.52.
[8] S. Fauziati et al., “Regresi Linear untuk Mengurangi Bias Sistem
Penilaian Uraian Singkat,” J. Nas. Tek. Elekt., Teknol. Inf., Vol. 10, No.
3, pp. 221–228, Aug. 2021, doi: 10.22146/jnteti.v10i3.1983.
[9] I. Mufiid, S. Lestanti, and N. Kholila, “Aplikasi Penilaian Jawaban Esai
Otomatis Menggunakan Metode Synonym Recognition dan Cosine
Similarity Berbasis Web,” J. Mnemonic: J. Tek. Inform., Vol. 4, No. 2,
pp. 31–37, Sep. 2021, doi: 10.36040/mnemonic.v4i2.4067.
[10] B. Sari and Y. Sibaroni, “Deteksi Kemiripan Dokumen Bahasa Indonesia
Menggunakan Algoritma Smith-Waterman dan Algoritma Nazief &
Andriani,” Ind. J. Comput., Vol. 4, No. 3, pp. 87–98, Dec. 2019, doi:
10.21108/indojc.2019.4.3.365.
[11] M.R. Parvez, W. Hu, and T. Chen, “Comparison of the Smith-Waterman
and Needleman-Wunsch Algorithms for Online Similarity Analysis of
Industrial Alarm Floods,” 2020 IEEE Elect. Power, Energy Conf. (EPEC),
2020, pp. 1–6, doi: 10.1109/EPEC48502.2020.9320080.
[12] V. Kumar, C. Bhatt, and V. Namdeo, “A Framework for Document
Plagiarism Detection Using Rabin Karp Method,” Int. J. Innov. Res.
Technol., Manage., Vol. 5, No. 4, pp. 17–30, Aug. 2021.
[13] T. Wahyuningsih, Henderi, and Winarno, “Text Mining an Automatic
Short Answer Grading (ASAG), Comparison of Three Methods of Cosine
Similarity, Jaccard Similarity and Dice’s Coefficient,” J. Appl. Data Sci.,
Vol. 2, No. 2, pp. 45–54, May 2021, doi: 10.47738/jads.v2i2.31.
[14] L. Meilina, I.N.S. Kumara, and N. Setiawan, “Literature Review
Klasifikasi Data Menggunakan Metode Cosine Similarity dan Artificial
Neural Network,” Maj. Ilm. Teknol. Elekt., Vol. 20, No. 2, pp. 307–314,
Jul.–Dec. 2021, doi: 10.24843/mite.2021.v20i02.p15.
[15] M.N. Cholis, E. Yudaningtyas, and M. Aswin, “Pengaruh Penggunaan
Synonym Recognition dan Spelling Correction pada Hasil Aplikasi
Penilaian Esai dengan Metode Longest Common Subsequence dan
Cosine Similarity,” InfoTekJar (J. Nas. Inform., Teknol. Jar.), Vol. 3, No.
2, pp. 242–246, Sep. 2019, doi: 10.30743/infotekjar.v3i2.1061.
[16] Sunardi, A. Yudhana, and I.A. Mukaromah, “Indonesia Words Detection
Using Fingerprint Winnowing Algorithm,” J. Inform., Vol. 13, No. 1, pp.
7–15, Jan. 2019, doi: 10.26555/jifo.v13i1.a8452.
[17] M.R. Faisal, D. Kartini, A.R. Arrahimi, and T.H. Saragih, Belajar Data
Science: Text Mining Untuk Pemula I. Banjarbaru, Indonesia: Scripta
Cendekia, 2023.
[18] H.A. Rouf, A. Wijayanto, and A. Aziz, “Deteksi Plagiarisme Skripsi
Mahasiswa dengan Metode Single-link Clustering dan Jaro-Winkler
Distance,” J. Pilar Teknol., Vol. 5, No. 1, pp. 26–31, Mar. 2020, doi:
10.33319/piltek.v5i1.50.
[19] C.D. Manning, P. Raghavan, and H. Schütze, Introduction to Information
Retrieval, Cambridge, England: Cambridge University Press, 2008.
[20] S.P. Gunawan, L. Dwika, and A.R. Chrismanto, “Analisis Fitur Stilometri
dan Strategi Segmentasi pada Sistem Deteksi Plagiasi Intrinsik Teks,” J.
RESTI (Rekayasa Sist., Teknol. Inf.), Vol. 4, No. 5, pp. 988–997, Oct.
2020, doi: 10.29207/resti.v4i5.2486.
[21] N.C. Haryanto, L.D. Krisnawati, and A.R. Chrismanto, “Temu Kembali
Dokumen Sumber Rujukan dalam Sistem Daur Ulang Teks,” J. Teknol.,
Sist. Komput., Vol. 8, No. 2. pp. 140–149, Apr. 2020. doi:
10.14710/jtsiskom.8.2.2020.140-149.
[22] I.M.S. Putra, P. Jhonarendra, and N.K.D. Rusjayanthi, “Deteksi
Kesamaan Teks Jawaban pada Sistem Test Essay Online dengan
Pendekatan Neural Network,” J. RESTI (Rekayasa Sist., Teknol. Inf.), Vol.
5, No. 6, pp. 1070–1082, Dec. 2021, doi: 10.29207/resti.v5i6.3544.
[23] N.L.W.S.R. Ginantra and N.W. Wardani, “Implementasi Metoda Naïve
Bayes dan Vector Space Model dalam Deteksi Kesamaan Artikel Jurnal
Berbahasa Indonesia,” J. Infomedia, Vol. 4, No. 2, pp. 94–100, Dec. 2019,
doi: 10.30811/jim.v4i2.1530.
[24] R.P. Nuristiqomah and Y. Anistyasari, “Pengembangan Kamus Istilah
Basis Data Berbasis Website Menggunakan Algoritma Cosine Similarity
untuk Meningkatkan Hasil Belajar Siswa,” J. IT-EDU, Vol. 5, No. 2, pp.
621–630, 2021.
[25] R. Nishiyama, “Adaptive Use of Semantic Representations and
EN-225
JURNAL NASIONAL TEKNIK ELEKTRO DAN TEKNOLOGI INFORMASI
p-ISSN 2301–4156 | e-ISSN 2460–5719
Volume 12 Number 3 August 2023 Santi Purwaningrum: Synonym Recognition Influence in ...
Phonological Representations in Verbal Memory Maintenance,” J. Mem.
Lang., Vol. 111, pp. 1–11, Apr. 2020, doi: 10.1016/j.jml.2019.104084.
[26] S. Inturi and S. Dusa, “Assessment of Descriptive Answers in Moodle-
Based E-Learning Using Winnowing Algorithm,” J. Contemp. Issues Bus.
Gov., Vol. 27, No. 3, pp. 2759–2769, 2021, doi:
10.47750/cibg.2021.27.03.331.
[27] E. Siswanto and Y.C. Giap, “Implementasi Algoritma Rabin -Karp dan
Cosine Similarity untuk Pendeteksi Plagiarisme Pada Dokumen,” J.
ALGOR, Vol. 1, No. 2, pp. 16–22, May 2020.
[28] Y. Nurdiansyah, A. Andrianto, and L. Kamshal, “New Book
Classification Based on Dewey Decimal Classification (DDC) Law Using
TF-IDF and Cosine Similarity Method,” J. Phys. Conf. Ser., Vol. 1211,
pp. 1–9, 2019, doi: 10.1088/1742-6596/1211/1/012044.
[29] R.N. Harahap and K. Muslim, “Peningkatan Akurasi pada Prediksi
Kepribadian Mbti Pengguna Twitter Menggunakan Augmentasi Data,” J.
Teknol. Inf., Ilmu Komput., Vol. 7, No. 4, pp. 815–822, Aug. 2020, doi:
10.25126/jtiik.2020743622.
[30] J. Xu, Y. Zhang, and D. Miao, “Three-Way Confusion Matrix for
Classification: A Measure Driven View,” Inf. Sci. (Ny)., Vol. 507, pp.
772–794, Jan. 2020, doi: 10.1016/j.ins.2019.06.064.
[31] Y. Zhang and J.T. Yao, “Gini Objective Functions for Three-Way
Classifications,” Int. J. Approx. Reason., Vol. 81, pp. 103–114, Feb. 2017,
doi: 10.1016/j.ijar.2016.11.005.
[32] E. Sutoyo and A. Almaarif, “Educational Data Mining untuk Prediksi
Kelulusan Mahasiswa Menggunakan Algoritme Naïve Bayes Classifier,”
J. RESTI (Rekayasa Sist., Teknol. Inf.), Vol. 4, No. 1, pp. 95–101, Feb.
2020, doi: 10.29207/RESTI.V4I1.1502.
EN-226
