Vol. 9(33), Jul. 2019, PP.

4258-4270

An Overview on Extractive Text Summarization

Mohamad Abdolahi1 , Shohreh Rad Rahimi2 , Ali Toofanzadeh Mozhdehi3
1
Iranian Academic Center for Education, Culture and Research (ACECR)Mashhad- Iran
2
Department of Computer Engineering Marlik Higher Education Institute, Nowshahre, Iran
3
Faculty of Computer Engineering, Qazvin Branch, Islamic Azad University , Qazvin, Iran
Phone Number: +9153140551
*Corresponding Author's E-mail: mabdolahi512@yahoo.com

Abstract

W
ith the increasing of online information and recourse texts, text summarization has become
an essential and more favorite domain to preserve and show the main purpose of textual
information. It is very difficult for human beings to summarize manually large documents
of text. Text summarization is the process of automatically creating and condensing form of a given
document and preserving its information content source into a shorter version with overall meaning.
Nowadays text summarization is one of the most favorite research areas in natural language processing
and could attracted more attention of NLP researchers. There are also much more close relationships
between text mining and text summarization. According to difference requirements summary with
respect to input text, established summarization systems should be created and classified based on the
type of input text. In this study, at first, the topic of text mining and its relationship with text
summarization are considered. Then a review has been done on some of the summarization approaches
and their important parameters for extracting predominant sentences, identified the main stages of the
summarizing process, and the most significant extraction criteria are presented. Finally, the most
fundamental proposed evaluation methods are considered.

Keywords: Text summarization,Text mining, Fuzzy text summarization, Statistical text

summarization, Text clustering.

1. Introduction
With an increasing amount of data available on the web, rising of news websites, publication of
various electronic books, and a significant growth in the number of published articles in different
fields of study, one of the main challenges for researchers of 21st century has been that of accessing
accurate and reliable data. The widespread volume of information available on one hand and time
limitation on the other has directed the researchers to the interesting area of summarizing texts and
powerful system to summarize documents. Given that much researches paper has been done on the
summarization subject and many articles have been published about it. But there is still a lot of
weakness in summarization field and there is a large gap to achieve an efficie nt system that can acts
like a human agent. Problems outlined in the Persian language are far more than other languages.
The complexity of language and lack of precision tools are current problems facing the Persian
language processing. Therefore, the review of operations and procedures performed on other

4258
Article History:
Received Date: Dec. 12, 2018
Accepted Date: Apr. 12, 2019
Available Online: Jul. 01, 2019
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

languages, considering the semantic field and use the semantic relationship, using tools such as
graph theory; statistical methods, fuzzy logic and data mining techniques can make a significant
contribution to Persian language processing.
A text made up of components such as words, phrases and sentences connected completely and
meaningfully together. One of the main areas of natural language processing is text mining, which
means discover and extract new information from the documents. Text mining is analysis the
documents to extract valuable hidden patterns from the text. It is also involves the detection of the
connection between words and sentences, classify and summarize texts. The main propose in this
research is an overview on text summarization .
More than half a century passed since the first research on the automatic text summarization [ 8].
Studying on text summarization systems emerged in 1950‟s which focused on the some basic
features of the text such as the position of the sentences in the text due to lack of powerful
computers and other problems in natural language processing. Since then, a lot of methods with
powerful tools were presented to simulate text processing, such as human brain.

2. A review of text mining and its relationship to text summarization

A significant portion of available information is stored in text data. Text databases are growing
rapidly due to increasing the amount of information in electronic form. Today, most of the da ta
available in the industry, business and other organizations are stored in the form of text database [2].
Text mining is a research field that has emerged from the combination of several other research
areas such as data mining, natural language processing and information retrieval. But there are some
differences between data mining and text mining.
Data mining often deals with structured data, but text mining involved with unstructured or semi -
structured data. On the other hand, even if there is still some construction in the text, there is
another reason that text mining is much more difficult than data mining. Unlike other data, such
asnumerical data, there are some semantic concepts in text and it can be difficult to model them
with traditional knowledge structures. For example, there are some difficulties such as words with
same meaning and different dictation and different meaning with same dictation.

2.1. Text mining applications

There is a wide definition of text mining; As a result, there are different theories about its
applications. Text mining is a new field in NLP but its software analyzes have been available since late
1990. Among the most common applications of text mining can be noted to the search engines. They
can find the most relevant documents whether the user types a misspelled word or phrase. Some
other applications of text mining that can be mentioned are:
Spam identification: Analyzing the title and content of email to determine whether it is spam.
Supervision: Monitoring the behavior of a person or group of people. There are some software
that can detect and control people behavior from telephone, internet and other communication
devices.
Concepts relationship: Occurrence of some words dependent on some other words.
Search and Retrieval
Classification and clustering data
Text summarization
Search and Retrieval
Classification and clustering data
Text summarization
There are some big problems in text mining. A biggest problem is that any document is a very
large set of words. If any word is assumed an element of a vector and considers different presence
scenarios in the text, then we will find that we are dealing with a NP-Hard problem with high
dimensional area. Word reduction is one of the preprocessing operations that lead to reduce the
dimension of the problem. Preprocessing and word reduction should be so efficient and powerful,

4259

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

because eliminating words may be lead to some text noises such as grammatical error. Apart from
reducing the size of the text, processing consists of some other steps as follows:
Case Folding: At this stage all words with big and small letters came in a uniform. Also, always try
to converted uppercase letter to lowercase.
Stemming: This stage it root extraction of words without considering different scenarios such as
singular or plural, the time and all prefix and suffix. This requires knowledge of a particular language
and many algorithms have been proposed for each language. Examples of these types of words in the
English language can be compressed and compression that convert to compress.
Stop Words: Some frequent words without concept. Examples of these types of words in the
English language such as do, does, will, and so on.
N-grams: N-grams are a sub-set of N words come together. They should be protected, but it is
better to come in the form of uniform.
Tokenization: Tokenization is a dividing the text to smaller units, which are often words. But in
some cases it is difficult to define the word and its scope. A word as a unit of text has the following
properties:
 Each set of continuous characters is always a word. These characters can even figure as well.
 Range of a word can be a whitespace character, symbols or end of the line.
 Sometimes whitespace characters and punctuation is not range word or token.
In languages such as English and Persian that words separated with whitespace, defining the
boundary words is not always an easy task. For example acronyms that has point, words in hypertext
links, and the words that are separated with symbolic characters and words with space character like
New York [4].

One of the important functions of the text mining is text clustering. The main aim of text
clustering is to place the similar sentences in same clusters. The number of clusters can be
determined by the user or automatically by the program. The first step of text clustering is dividing
the text to its component and separated the sentences.

2.2. Text mining and text summarization

It is obviously that it cannot be separated text mining and text summarization. Several studies
have been conducted in the field of text mining and text summarization. All of them show that the
text summarization is also kind of mining and exploring to find the important parts of a document.
The main problems facing to text processing and summarizing are lack of qualified linguist expert
work with programmers, grammar based of linguistic theories, and the lack of reasoning and thinking
word processors machines.
The first problem can be overcome with employing linguistic expert. The second problem
according to the semantics can be resolved by revising the language theories. But it is difficult to
create machines with reasoning and thinking power.
On the other hand, dealing with high volume data on internet, especially when the goal
understands the main concept of documents, using machines instead of humans can be easier, faster
and more reasonable.
At the beginning consider the different views on the automatic text summarization and the
summaries created by the man. There are three main views that are: no differences between human
and machine summary, preference of human summary on machine summary and preference of
machine summary on human summary.
The theory of the superiority of processing and summarizing man believes that human mind more
powerful than machine processor. Decision-making power and choice of current machines are less
than human mind. The theory of the superiority of machine processing and summarizing believes
that in the near future computers can process the linguistic information with more speed and
accuracy than human. The reason for this claim is the more quickly operation of math and machinery
computing than a human.

4260

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

There are three major advantages of automatic generation of summary by the machines. The
advantages are: summary size is controllable, its content is predictable and it can be determined that
any part of the summary related to which part of the original text.

3. Different criteria in the text summarization sestem desining

There are various approaches to text summarization, some of which have been extant for about
50 years. Text summarization approaches can be categorized in different way s according to the
various measures and features. For example, according to Hovy and Lin [ 2], the measures and
features are related to input, output, purpose and result in different types of summary. In the
following various methods have been considered in general classification [22].

3.1. Summarization based on output summary

Text Summarization methods can be classified into extractive and abstractive summarization. An
extractive summarization method consists of selecting important sentences or paragraphs from the
original text and gathering them into shorter text. The importance of sentences is decided based on
some statistical and linguistic features of sentences, extract and placed in the output text. In this
paper, the more emphasis is on extraction techniques.

An abstractive summarization attempts to extract the main concept of the text in clear natural
language without necessity to use text phrases. Each abstractive summarization consists of
comprehension part to interpret the text and find the new concepts and production part to generate
new shorter text with most important information from the original document. In this method,
sentences could be omitted or changed or even new sentences could be generated. It should be
noted that this method is very complicated and even more complicated than machine translation.

3.2. Summarization based on details

Text Summarization methods also can be classified into another classes named Indicative and
Informative summarization. In indicative summary the main idea of the text is presented and it is
usually about 5 to 10 percent of the original text. This kind of abstraction is used to encourage the
reader to read the original text. For example a brief summary of a movie or a story in the context of
its advertising which only leads to further questions and encourage the readers to watch the film and
read the story. Informative summary consist of the main abstraction and the important issues of the
text. This kind of summary is between 20 and 30 percent of the original text and contains all the main
points of the text.

3.3. Summarization based on contents

Another classification in text summarization is content base classes. This type of classification can
be divided into two categories: Generic and Query-Based. Generic summarization is not depended on
topic of text and it is assumed that the reader does not have any basic knowledge about the text.
Generic summary contains all aspects and important issues of main text and readers are able to
achieve a thorough understanding of the subject without prior knowledge of the text.

But in Query-Based summary, it is assumed that the reader has a general knowledge about the
topic and just looking for specific information in the text. In this case based on the user's question, a
related summary is created. Most of these summarization systems are extractive.

3.4. Summarization based on limitation

There is another category based on the limitations of input text. The mentioned group has three
categories that which are: Independent, Domain dependent and Genre specific.

4261

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

Independent summarization is something like generic summary, accepted every text of each field
and generate a general summary regardless of the text scope or type.

Domain dependent summarization accepts texts with specific field of literature and type. There
are many specific text patterns such as News, science text, fiction, sports. Ge nerated summary of
domain dependent systems are according to the input text type. Genre specific summarization is
trying to summary much more specialized field. This group can be abstract specific literature such as
sports news, texts related to the field of geography, political News and etc.

3.5. Summarization based on the number of input texts

Automatic summarization systems can be divided other two main categories based on their input
text. The mentioned categories are Single document and Multi document summarization. The input
of single-document summarization systems is only one text but multi -document summarization,
which is very popular these days, is the improvement of single document summarization to
collections of related documents. The main propose of multi -document summarization is
summarizing texts and removing redundancy and considering the similarities and differences in the
information content of different documents. Multi document text summarization accepts multiple
documents with common scope in a different perspective and closely tied to answering systems and
search-based summarization [23]. There are two major approaches to summarize multiple
documents. The first approach uses the usual methods of single document summarization and
summarizes each document separately. Then combining all of the summaries together and tried to
remove redundancy by overlap similar sentences to produce the final summary. There are some
proposed methods too which behave created single summarize as inputs and then merge them to
create main summary [11].
The second approach is specifically designed for multiple documents. In the mentioned
approaches, all documents assumes as one document and all the important sentences are extracted
from all of the documents together using methods like graphs or clustering. This approach is more
challenging, intelligent and complicated. An example of the second approach is the SUMMONS
system which extracts and combines information from multiple sources and passes them to a
language generation component and produces the final summary. In general, the complexity of a
single document model is far less than multi document models [5].

3.6. Summarization based on language acceptance

A text summarization system also can be divided to mono-language and multi-language. Mono
language summarization only accepts documents in a specific language such as English, Persian. But a
multi-language summarization adopted by different languages and can be able to summarize them.

4. Similarity measures of extractive summarization

As mentioned earlier, extractive summary is using main text sentences to crate summary. This
means that more important sentences are found and transferred to the summary. To determine the
important sentences some parameters should be considered. After determining the parameters
value, sum of them are intended as the final weight of each sentence. Finally sentences with the
maximum weight based on compression rate are extracted and by ordering presence in main text
transferred to the summary. Given that some mentioned parameters reduce the importance of a
sentence their value is calculated as a negative [24]. Some of The parameters and features are
described below.

4.1. Content keyword feature

Keywords are often noun and determined by tf * idf criteria. Sentences containing keywords have
more chances to exposure in summarized output. Other methods have been proposed for extracting
keywords. Some of these methods are word analysis of morphological, statements extraction and

4262

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

scoring them, and noun phrases extraction, clustering and ranking them. Word morphological
analysis plays an important role in natural language processing and helps to resolve the ambiguity in
the words. The mentioned criteria study of root words, prefixes and suffixes attache d to a word.

4.2. Similarity of sentence and title of document

Similarity of sentence and title of document is the number of common words between title and
sentence. Sentences including title words have high importance and more chance to place in output
summary. In most of the proposed methods the criteria is calculated as Eq. (1).

(1)
√

In Eq. (1), the numerator is the number of similar sentence and title words and the denominator is
the square root of the product of the title length and sentence length.

4.3. Sentence location in the document

This measure is based on the assumption that sentences occurring in initial and end position of
both text and individual paragraphs have a higher probability of being relevant to main topic of the
document. The experimental results showed that the best correlation between the automatic and
human-made extracts was achieved using first and end sentences in summarized output. But the first
sentence is more important than last one. By reading the first sentence of every paragraph can be
understand its main topic. The Eq. (2) is used to calculate the importation rate of sentence location:

) ))) )) )))

(2)

PSi is effective value of the relative position of the ith sentence that calculated based on entropy.
The C is a constant value between zero and one and Xi is a sentence location value.

4.4. Important words and phrases

There are some words and phrases that increase the sentence importance. Numbers (including
date, time, price, percentage and weight) and proper names (including person, place, and time) have
specific important information and determining their importance depends to texts type. The score
rate of a sentence regard to the important words and phrases are obtained according to the Eq. (3)
by dividing the number of these words on the total words in sentence.

(3)

4.5. Capitalized words and acronyms

Capitalized words or acronyms are having special concepts. The words with capital letters are in
the Latin alphabet languages and there are no such words in Persian. But there are acronyms in all
languages and in Persian language can be used them to determine the significance of a sentence.

4.6. Cue-Phrase
Cue-Phrases are some phrases like “finally”, “as result”, “in this paper”. Statements containing
these words certainly including important topic and must be preserved. There are some words that
lead to increase the importance of a noun and its sentence. For example the words like Dr., Mr., and
Miss.
There are some other words placed at the beginning of a sentence. These words lead to decrease
the importance of the sentence. The mentioned sentences are come to complete the meani ng of
previous sentence and can be ignore it. Some example of them are “because”, “for example”,

4263

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

“therefore” and “thus”. The weight of statements containing these words in both positive and
negative form calculated as Eq. (4).
| | )
(4)

In the Eq. (4), Si is the number of words in sentences i, Spi is the number of positive Cue -Phrases
and Sni is the number of negative Cue-Phrases.

4.7. Specific words

Specific words are certain words that have special meaning and may vary by keyword in the text.
These words can be scientific issues such as "mathematics" psychology ". They increase the
importance of sentence.

4.8. Sentences containing words with different fonts and font effects
Some of the words that are important or specific issues charged with capital letters started or
bold, italic or underlined. They have a higher chance of exposure sentences in the output.

4.9. The continuity and similarity of a sentence with other sentences

For each sentence in the text resemblance with other sentences are considered and sentence
with more resemblance has high score. The process is repeated for all sentences of text and
sentences with higher scores are more likely to have a presence in summary. The similarity of two
sentences is similarity between contained words, sentence length and other criteria mentioned
before.

4.10. The similarity of a sentence with a paragraph topic sentence

In this section, the similarity of each sentence with paragraph topic sentence is considered.
Usually the topic sentence in each paragraph is a first sentence in paragraph. Similarity measures in
this section are the same criteria like in the previous sections.

4.11. Specific terms

Specific terms and phrases in text are marked by symptoms such as " ", << >>, etc. Review of the
literature summarized by the human shows that the sentences containing these phrases are of more
important. Sentences with specific terms scores as Eq. (5):

(5)

Sip is the number of specific terms in the sentence, Si length of sentence and Xi scoring rate of the
sentence based on mentioned criteria.

4.12. The length of sentences

Sentence length is an important criterion in text summarization. The best sentences for summary
output are the ones with average length. Sentences containing less than a pre-defined number of
words are not included in the abstract. But sometimes the effects of previous measures increases the
short sentence importance and cannot be absolutely eliminated it. The Eq. (6) is determine d and
rated the effect of sentence length. The sentences out-of-threshold with negative rate is also
considered. In the Eq. (6) Si is sentence length and Max j=1:n(|Si|) is maximum sentence length in
the document.
) ) (6)

4264

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

4.13. Pronouns
Sentences with pronouns have less importance than the one with nouns. Pronouns refer to nouns
that described in the previous sentences. The position of the pronoun in the sentence is also
important. For example, a sentence with pronoun at the beginning is less important sentence. But if
the pronoun is placed at the middle and end of a sentence, it is a little more important sentence. The
score of sentence including pronoun calculated as Eq. (7):

)
(7)
)

In Eq. (7), the | Spi | is the number of pronouns in sentences, Si is the sentence length and BPi is a
constant parameter. If the pronoun location is in the first three word in the sentence, BPi = 1,
otherwise BPi = 0.

5. Different approaches to text summarization

Although summaries created by human are often abstractive, but the most successful machine
summary are extractive. Because of the semantic ambiguity in natural language processing, a good
extractive summary is much better result than abstractive summarization. In fact it is far away to
create an ideal automatic abstractive summarization method. The first automatic text summarization
and the most important new proposed methods are extractive systems that extract im portant
sentences based on heuristic features such as sentence positions in the text, the frequency of words
in it and some important keywords related to text [8] [13]. To create a system with high-quality due
to a different type of language and text input, various methods and algorithms in machine learning
and natural language processing is proposed. Most of extractive summary methods are emphasis on
key sentences. The differences between methods are the different algorithm used to detect, identify,
extract and the way to put sentences in the summary. In the following the most important
approaches will be discussed.

5.1. Statistical approaches

Using statistical methods to summarize the text is an effective approach that has been used in
many articles. In statistical methods, the important sentences are selected based on word frequency,
indicator phrases and other features regardless of the meaning of the words that mentioned in
previous sections.
There are several methods to determine the key sentences such as, The Title Method [13], The
Location Method [16], The Aggregation Similarity Method [26], The Frequency Method [6], TF- Based
Query Method [1], and Latent Semantic Analysis. But the most common methods are Classifier
Bayesian (Bayesian Classifier) and communication concept relations approaches. One of the most
famous statistical methods to summarize the text is LSA method [8]. LSA is an algebraic-statistical
method that extracts hidden structures meaning of words and sentences [5]. This approach is an
unsupervised method that extracts text structures just by information of the words in sentence
without the need of any other knowledge. The idea in LSV is words that they are shared between
different sentences are the reason of meaning dependence. LSA also be able to show the meaning of
words and vocabulary simultaneously. It uses algebraic singular value decomposition (SVD) metho d
to determine the relationship between sentences and words. Addition to being able to model the
relationship between words and sentence, SVD also can reduce the noises and lead to improve the
accuracy.

4265

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

Summarization algorithm based on LSA method consists of three steps: create the input matrix,
applied the SVD method on the created matrix and sentence extraction. LSA also has some
limitations. The most important of them are as follows:
 The algorithm does not use the information about the words arrangement in sentences,
grammar and morphology relationship. However, this information can be useful to better
understand words and sentences.
 The algorithm does not use any word knowledge and word database.
 By increasing the number of different words and heterogeneous data, the performance of
the algorithm greatly reduced. Performance reduction is due to the time and memory
complexity of the SVD method.

5.2. Lexical Chain based approaches

Lexical chain produces a presentation of text contiguous structures. Basical ly, it uses the word net
database to determine the connection between the terms and then creates a continuum between
these terms. The first computational model of lexical chains was presented by Moris and Hirst in
1991 and the first use of lexical chains in text summarization was proposed by Barzilay and Elhadad
[17] [19].
The scores that considered for terms are according to the type and number of the relationship of
chains set. The final summary is included sentences with strongest chain connection. Lexi cal chains
can be computed the semantically relation of words. It cans also identify the synonyms and
hyponyms to place them in a group into the same lexical chain. Lexical chain also used for
information retrieval and grammatical error corrections.
There are two drawbacks in lexical chain approaches. The first one is ambiguity in the word chain.
If some of words have semantic ambiguity, the created chain also will have semantic ambiguity. The
second drawback is the relation of created chain to main topic. All chains are not related to the main
topic.

5.3. Graph based approach

Graph based approach provides text summarization methods using graph theories. After common
preprocessing steps such as stemming and stop word removal, sentences in the documents are
represented as nodes in a directed graph. Sentences are connected to each other by edges according
to sentence. The basic idea of graph-based approaches is something like voting. As result when an
edge connected a node to other node, it means that votes to it and whatever the input degree of a
node is high, it has a higher priority. In the mentioned method there is also voting score. If a node
has greater output degree, its importance increases. The importance degree of each node is
calculated by S(Va) as Eq.(8) for all nodes in a graph recursively. The process continues until coverage
and no change in S(Va).

)
S (Va) = (1 – d) + d * ∑ ) (8)
)

In the Eq. (8), V is nodes, E is edges, in(Va) is the number of Va input edges, out(Va) is the number
of Va output edges and d is an input parameter between zero and one. S(Va) is Va score and S(Vb) is
Vb score. The algorithm can also be applied on an undirected graph, but the output summary is more
different with complexity of time.
An effective graph based summary proposed algorithm is Text Rank algorithm [18]. It uses an
unsupervised method to extract key words and sentences with scoring to nodes based on previous
mentioned similarity measures [17]. The algorithm is started with an optional node values and
recursively repeated until coverage to predefined threshold. The optional nodes value has no effect
on the final scores.

4266

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

5.4. Cluster based approaches

Some automatic summarization systems are used clusters to produce significant summaries. In
this approach, different clustering algorithms are applied for dividing the text into tokens such as
words, phrases, sentences and even paragraphs. Cluster-based approach is an extractive
summarization and the most similar sentences are based on mentioned similarity measures are
placed in the clusters. The biggest cluster is selected as main topic and its sentences are extracted
and placed in summary output. Euclidean distance, Cartesian similarity, cosine and some other
similarity measures are used for defining the similarity and dissimilarity of clusters [17].
A. Agrawal & U. Gupta, 2014 proposed K-means clustering algorithm to place important sentences
within clusters and chose the biggest cluster as main topic [1]. The documents are represented using
term frequency-inverse document frequency (TF-IDF). In the context, term frequency (TF) is the
average number of occurrence (by document) in the cluster. The topic is represented by words of
which the value TF-IDF is higher in the cluster. The selection of the important se ntences is based on
the similarity measures of the sentences with the topic of cluster.
Two optimization ideas are proposed in their algorithm. The first one is a method to calculate
sentence score and the second one is a way to get the optimal number of clusters. To calculating
sentence score the algorithm uses TF-IDF and the sentence length (X).

∑
) (9)

One of the most important issues in k-means clustering is determining the optimal number of
clusters. Given that in this method largest cluster is considered as main topic, input text size has a
direct impact on determining the number of clusters. For example a big K lead to small clusters and
as result small and disperse summary with very low correlation and small K lead to big and dense
clusters and as result low compression text.
N.I.Meghana & M.S.Bewoor proposed another technique to create a query based summarization
system using clustering methods [18]. They used Expectation Maximization Clustering (EM) algorithm
and implementation methodology is divided into two sections. The EM algorithm has been
implemented in the first part and query-based summarization is done in second part. The proposed
method has been used the Word net.

5.5. Fuzzy logic based approaches

Fuzzy logic based approaches consider each characteristic of a text as the input of fuzzy system.
These methods only used fuzzy logic to detect and extract the important sentences. Fuzzy text
summarization methods are differ based on differences in extracting features, fuzzy rules, linguistic
variables, membership functions, fuzzification and defuzzification methods. Some changes in
numerical values lead to produce better linguistic variables [14]. F. Kyoomarsi, H. Khosravi proposed
a method thatsummarizes the text in two stages [3]. The first stage is pre-processing and the second
stage is performing fuzzy analysis on text. The features that they extracted are: the number of
common words including the sentence and title, the number of words in sentence, th e similarity of a
sentence with a paragraph topic sentence and similarity of sentence and first sentence in paragraph .

6. Integration similar statements in extractive summarization

Redundancy in a series of sentences in texts, especially on the web, is one of the problems in
natural language processing and on the other hand, creates a new research opportunities in text
summarization area. There are many sentences with similar meaning and conce pt in every
document. Because of determining the number of sentences in the summary by the user, statements
with a similar concept are placed in summary. There are also some other important concepts in the
main text and due to the limitation number of sentences in summary, they do not select and extract.
So many sentences with one concept placed in summary and many other important parts are

4267

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

ignored. The problem is more common in multi document summarization [20]. Similar concept
sentences are also created the ambiguity in question answering systems.
There are some proposed methods to combine similar sentences to extract much more important
sentences from the main text. One of the most important sentence fusions is dependency trees.
Katja Filippova is proposed a directed acyclic graph (DAG) to sentence fusion [12]. Clustering
approaches are also can be used for combining similar sentences. In these approaches, small clusters
are considered to place very similar sentences and then each cluster can be combined t o one
sentence.

7. Assessment and evaluation of summarizing methods

Due to extensive automatic summarization methods, assess the accuracy of created methods is
important. It can be said that the process of evaluation is much more difficult than summarizin g the
text. In most cases there is not an ideal summarize of a document. The main problem in summarizing
assessment is using extensive criteria and absence of a standard method for evaluation. Such
problems are rare in other natural language processing cases.
Most of the developed methods are used two approaches for evaluation. The first approach is to
judge by human and the second approach is compared to reference summary.
In the first approach summary generated by machine can be compare with the same summary
generated by a man. The problem with this approach is the different personal tastes and opinions. To
reduce the impact of the effects, automate summary compares with more than one summary
generated by humans and it is time consuming.
In second approach, there is more than one original text with their summary in a dataset. The
original document and its optimized summary is a good base for evaluate the proposed
summarization algorithm.
One of the most famous evaluation methods is ROUGE-N. The mention evaluation method
compares automate generated summary with summaries generated by the human [15] [9]. ROUGE-N
uses three evaluation criteria: Precision (p), Recall (R) and F1-measure and they are calculated as Eq.
(10):

(10)

In the Eq. (10) sum_ref determines summary extracted by experts and sum_cand determines
summary extracted by system.
Another evaluation criterion mentioned in ROUGE-N is calculated by the Eq.(11). The measure
compares number of N-grams in machine summary and human summary.

∑ ∑ ))
(11)
∑ ∑ ))

In the Eq. (11) N is the number of N-grams, count_match (N-gram) is maximum number of N-
grams which are in machine summary and human summary simultaneously. count(N -gram) is also
the number of N-grams in human summary.

4268

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

CONCLUSION
Text summarization is one of the most exciting research areas in natural language processing. It is
an open research areas and a lot of researches are having been done about it. Text summarization
can be classified into different groups and approaches and the most notably of them studied in this
paper. In this study, at first, the topic of text mining and its relationship with text summarization are
considered. Important designing criteria in text summarization systems are presented. Different
approaches for summarization and important parameters needed to rate important sentences are
introduced. Finally important evaluation approaches are introduced.

References
[1] A. Agrawal, U. Gupta, “Extraction based approach for text summarization using k -means
clustering”, International Journal of Scientific and Research Publications, Vol 4, Issue 11, 2014.

[2] E. Hovy, C. Lin, ”Automated Text Summarization and the SUMMARIST System, In Advances in
Automatic Text Summarization”, MIT Press, pp. 81-94, 1999.

[3] F. Kyoomarsi, H. khosravi, E. Eslami, M. Davoudi, “Extraction-based Text Summarization using

Fuzzy Analysis”, Iranian Journal of Fuzzy Systems, Vol. 7, No. 3, pp. 15-32, 2010.

[4] G.O Makbule, I. Cicekli, F. Nur Alpaslan, “Text Summarization of Turkish Texts using Latent
Semantic Analysis”, 23rd International Conference on Computational Linguistics, pp. 869–876,
2010.

[5] G.O. Makbule, “Text Summarization using Latent Semantic Analysis”, M.S thesis, Middle East
Technical University, 2011.

[6] G. Salton, “Automatic Text Processing: The Transformation, Analysis, and Retrieval of
Information by Computer”, Addison-Wesley Publishing Company, 1989.

[7] H. Edmundson, “New Methods in Automatic Extracting”, Journal of the Association for
Computing Machinery, Vol16 (2), PP. 264-285, 1969.

[8] H. Luhn, “The automatic creation of literature abstraction”, IBM Journal of research and
development, Vol 2, pp.159-165, 1958.

[9] H. Shakeri, S. Gholamrezazadeh, M. Amini Salehi, F. Ghadamyari, “A New Graph -Based

Algorithm for Persian Text Summarization”, Computer Science and Convergence, Lecture
Notes in Electrical Engineering, 2012.

[10] J. Bellegarda, “Exploiting latent semantic information in statistical language modeling,” in

Proc. IEEE, Vol. 88, No. 8, pp. 1279-1296, 2000.

[11] J.V Goldstein, J. Mittal, J. Carbonell , M. Kantrowitzt, “Multi-document summarization by

sentence extraction”. In Proc. of the ANLP/NAACL Workshop on Automatic Summarization,
2000.

[12] K. Filippova, M, Strube, ” Sentence Fusion via Dependency Graph Compression”, in

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pp
177–185, 2008.

4269

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543
M. Abdolahi et al. / Vol. 9(33), Jul. 2019, PP. 4258-4270

[13] K. Youngkoong, S. Jungyun, “An Effective Sentence Extraction Technique Using Contextual
Information and Statistical Approaches for Text Summarization”, Pattern Recognition Letters,
2008.

[14] L. Suanmali, M. Salem, B. Salim, N. Salim, “Sentence Features Fusion for Text summarization
using Fuzzy Logic, IEEE, pp.142-145, 2009.

[15] M. Pourvali, A. Abadeh Mohammad, “Automated text summarization base on lexical chain and
graph using of word net and Wikipedia knowledge base”, IJCSI International Journal of
Computer Science Issues, No. 3, vol. 9, 2012.

[16] M. Wasson, “Using leading text for news summaries: Evaluation results and implications for
commercial summarization applications”, in Proc. 17th International Conference on
Computational Linguistics and 36th Annual Meeting of the ACL, pp.1364-1368, 1998.

[17] N. Alami, M. Meknassi, N. Rais, “Automatic Texts Summarization: Current State of the Art”,
Journal of Asian Scientific Research, Vol 5(1), pp 1-15, 2015.

[18] N.I. Meghana, M.S. Bewoor, M.S, S.H. Patil, “Text Summarization using Expectation
Maximization Clustering Algorithm”, International Journal of Engineering Research and
Applications (IJERA), Vol. 2, Issue 4, pp.168-171, 2012.

[19] R. Barzilay, M. Elhadad, “Using Lexical Chains for Text Summarization”, the MIT Press, pp. 111-
121, 1999.

[20] R. Dragomir, J. Budzikowska, “Centroid-based summarization of multiple documents: Sentence

extraction, utility-based evaluation, and user studies”, In Proceedings of the ANLP/NAACL
2000 Workshop on Automatic Summarization, pp 165–172, 2000.

[21] R. Mihalcea, “Graph-based ranking algorithms for sentence extraction, applied to text
summarization”, in Proceedings of the ACL 2004 on Interactive poster and demonstration
sessions, 2004.

[22] S. Gholamrezazadeh, M. Amini, B. Gholamzadeh, “A Comprehensive Survey on Text

Summarization Systems”, IEEE, 2009.

[23] T. Hirao, Y. Sasaki, H. Isozaki, “An extrinsic evaluation for question-biased text summarization
on qa tasks”. In Proceedings of NAACL workshop on Automatic summarization, 2001.

[24] V. Gupta, G. Lehal, “A Survey of Text Summarization Extractive Techniques”, Journal of

emerging technologies in web intelligence, VOL. 2, NO. 3, 2010.

[25] V. Gupta, G.S Lehal, “A Survey of Text Mining Techniques and Applications”, Journal of
Emerging Technologies in Web Intelligence, Vol. 1, no. 1, 2009.

[26] W. al-sanie, “Towards an infrastructure for Arabic text summarization using rhetorical
structure theory”, M.S Thesis”, Department of computer science. King Saud university, Riyadh,
Kingdom of Saudi Arabia, 2005.

4270

International Journal of Mechatronics, Electrical and Computer Technology (IJMEC)

Universal Scientific Organization, www.aeuso.org
PISSN: 2411-6173, EISSN: 2305-0543