Online-Journals.

org

JET International Journal of

Emerging Technologies in Learning
iJET | eISSN: 1863-0383 | Vol. 19 No. 1 (2024) |

https://doi.org/10.3991/ijet.v19i01.46079

PAPER

Analysis of Abstractive and Extractive

Summarization Methods

Mahira Kirmani1(), ABSTRACT

Gagandeep Kaur1, This paper explains the existing approaches employed for (automatic) text summarization.
Mudasir Mohd2 The summarizing method is part of the natural language processing (NLP) field and is applied
to the source document to produce a compact version that preserves its aggregate meaning
1
University Institute of
and key concepts. On a broader scale, approaches for text-based summarization are catego-
Computing, Chandigarh
rized into two groups: abstractive and extractive. In abstractive summarization, the main
University, Punjab, India
contents of the input text are paraphrased, possibly using vocabulary that is not present in the
2
Department of Computer source document, while in extractive summarization, the output summary is a subset of the
Application South Campus, input text and is generated by using the sentence ranking technique. In this paper, the main
University of Kashmir, Jammu ideas behind the existing methods used for abstractive and extractive summarization are dis-
and Kashmir, India cussed broadly. A comparative study of these methods is also highlighted.

mahirakirmani68@
KEYWORDS
gmail.com
textual summarization, structure-based approach, extractive summary, sentence ranking
methods, abstractive summary, semantic-based approach

1 INTRODUCTION

The exponentially increasing digital data that is accessible worldwide makes the
utilization of an automatic text summarization tool inevitable, as manual text sum-
marization entails a considerable number of impartial and knowledgeable experts.
The sole objective of automatic text summarization is to express all information in
the input text in a vivid, concise, and comprehensive manner, enabling users to save
effort and time. Initially, automatic text summarization techniques were applied to
one input document, called single document text summarization. The enormous
amount of redundant data present on the web provoked the use of multi-document
text summarization [1], where a set of multiple documents served as an input to the
system. The [2] process of automatic summarization can be divided into the follow-
ing steps: (a) Preprocessing of the original text, (b) Intermediate representation, and
(c) Generating an output as a summary. The summarist text summarization system

Kirmani, M., Kaur, G., Mohd, M. (2024). Analysis of Abstractive and Extractive Summarization Methods. International Journal of Emerging Technologies in
Learning (iJET), 19(1), pp. 86–96. https://doi.org/10.3991/ijet.v19i01.46079
Article submitted 2023-10-20. Revision uploaded 2023-11-22. Final acceptance 2023-11-24.
© 2024 by the authors of this article. Published under CC-BY.

86 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)
 Analysis of Abstractive and Extractive Summarization Methods

introduced in [3] implements three phases: (a) topic identification, (b) interpreta-
tion, and (c) generation. Textual summarization tasks are generally divided into two
classes: abstractive and extractive [3]. Extractive summaries are formed by concat-
enating the main sentences or phrases of the source document. It is a difficult effort
to identify the key sentences in the input document; sentence scoring or ranking
algorithms are used to solve this problem. On the other hand, [4] [5] abstractive sum-
maries are the compressed paraphrased version of the input text and thus are not a
mere concatenation of the main sentences or phrases present in the input document.
Summaries may be divided into two categories based on the original content: an
indicative summary and an informative summary. An indicative summary refers
to the main concepts of the input document, while all of the pertinent information
reported in the input document is included in the informative summary. [6] [4].
Table 1 briefly describes the summarization types. This paper explains the differ-
ent methods used for extractive and abstractive summarization. Section 2 shows
the related work; Section 3 defines the different extractive summarization methods;
Section 4 presents the different abstractive summarization methods; Section 5 gives
a conclusion; and Section 6 contains the references.

2 RELATED WORK

The 1950s saw the start of the automatic text summarization task [10]. It is now
over half a century old and is still progressing because of the increased use of digital
data. Luhn [10] unfolded the concept of how frequently occurring words can help in
determining important sentences. Then Edmundson [6] broadened Luhn’s approach
by imparting several other features for indicating salient sentences: (a) Frequency or
count of the word in the input text; (b) Frequency of the title terms in the sentence of
the source document; (c) Position of the sentence; and (d) Count of cue-phrases such
as “significantly,” and “concluding” [6]. Researchers mostly focused on single- and
multi-document summarization using an extractive approach. At that time, Paice was
the one who focused on the techniques for language generation. He pinpointed the
main problem that sentence extraction algorithms suffered from (that was the unin-
tended inclusion of those sentences that contained references to the sentences absent
in the summary), which resulted in inconsistent summaries [11]. This was relatively
early research; the part that follows discusses more recent studies in the area of auto-
matic text summarization. Methods such as lexical aggregation used also helped in
condensing the input text by replacing two related concepts with another concept; for
example, selling and buying are related to each other, so we replaced them with busi-
ness. For redundancy removal, syntactic aggregation method was used; for example,
Sam plays and Lin plays becomes Sam and Lin plays. Summaries generated on the
basis of the keywords are called keyword summaries. We must determine the key-
words contained in the input document for keyword summarization. [12] describes
the methods for keyword identification. Query-focused summarization determines
important parts of the input document based on the user-provided query. The simi-
larity between the query and the sentences in the input document is calculated using
support vector regression (SVR). Also summarizes multiple documents on the basis of
user queries [9]. In order to generate a quality summary, researchers in recent work
have focused on employing neural networks and fuzzy logic [13] [14].
Additionally, it was shown that summarizers based on fuzzy logic and neural
networks perform better than those based on statistical methods. Neural networks
and fuzzy logic were even used for improving and addressing sentence scoring

iJET | Vol. 19 No. 1 (2024) International Journal of Emerging Technologies in Learning (iJET) 87
Kirmani et al.

techniques [15]. Recently, [16] [10] used neural networks in order to summarize the
news articles. In the training phase, neural network learned how to check import-
ant features of sentences. On the basis of these features, input text was filtered by
the neural network, and in the end, a summary of the news article was generated.
Deep learning approaches for text summarization have also shown considerable
results. In [17], a deep auto-encoder is used to generate an extractive query-focused
summary for a single document. Ensemble noisy auto-encoder, an extension of deep
auto-encoder, creates noisy input by adding random noise to the input representa-
tion in order to choose sentences from a cluster of noisy inputs. Experiments were
performed on two separate email corpora that were publicly available, and the sys-
tem was evaluated using Rouge. Lately, [18] has used attentional encoder-decoder
recurrent neural networks to frame their abstractive automatic text summarizer.
This work also attempted to address serious problems occurring in the basic model
by proposing a few novel models. This paper claimed that these models contributed
to boosting the system’s performance further.

Table 1. Shows the summarization types

Summarization Types Definition

Extractive Concatenation of important sentences or phrases of input text. [3]
Abstractive The compressed paraphrased version of the input text. [7] [8]
Single document Single document serves as an input to the system.
Multi-document Multiple documents serve as an input to the system. [1] [9]
Indicative Points to the main concepts of the input document.
Informative Includes all the relevant information that is reported in the input document.
Keyword Consists of set keywords or phrases present in the input text.
Headline Summarizes input document by a single important sentence.
Generic Does not make any assumptions regarding domain or genre of the input;
determines importance with respect to the contents of the input document
Query Focused Based on query given by user it determines important sentences from the
input document

3 EXTRACTIVE SUMMARIZATION METHODS

To understand the different methods used for extractive summarization, it is pref-

erable to understand the primary stages of the extractive summarization approach.
The predominantly extractive approach is divided into three primary stages:
1) intermediate input text representation, 2) calculating sentence ranks or scores,
and 3) generating a summary. These stages are interdependent; that is, each stage’s
output can be used as an input for the next step.

1. Intermediate input text representation: An input text document is considered raw

until it is preprocessed and then transformed into a particular format. In order
to apply scoring algorithms, raw input needs to be transformed into a scoring
algorithm-specific representation. Frequency-based algorithms consider frequent
words as keywords; therefore, text segmentation takes place at the word level.
The segmented input text is transformed into a table representation, containing

88 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)
 Analysis of Abstractive and Extractive Summarization Methods

words and their corresponding frequencies. Likewise, sentence length, sentence

position, etc., based algorithms segment input text at the sentence level, which
represents each sentence as an indicator. Graph-based algorithms may represent
the entire input document as a set of interconnected sentences.
2. Calculating sentence ranks or scores: After transforming the input text into a cer-
tain intermediate representation, sentence ranking algorithms are applied to it in
order to assign scores to the sentences. TF or IDF, sentence position, sentence length,
word co-occurrence, lexical similarity, and proper noun are some of the existing
sentence scoring algorithms. The sentence scores assigned determine the
importance of the sentences; highly scored sentences have greater chances of
being selected for the summary.
3. Generating summary: In the last phase, a linear combination of highly ranked
sentences forms a summary. The summary size is apparently less than the size
of the original text document. In this phase, similarity check algorithms can be
employed in order to remove redundancy in the summary.

The subsequent section presents the main sentence ranking methods employed
for extractive summarization approach. The various sentence ranking methods are
widely categorized as statistical methods and semantical methods [19].

A) Statistical methods
The methods most widely used in the literature for the extractive summariza-
tion approach are statistical methods. Statistical methods operate by observing
statistics (such as the number of words, probability of a particular word, term
frequency (TF)–inverse document frequency (IDF), etc.) of the text document to
identify salient sentences. Such methods do not take into consideration the mean-
ing or sense of the words, phrases, or sentences contained in the input text docu-
ment. The methods are described below.
1. Word frequency method: The concept of word frequency is quite old and
was unfolded by Luhn. According to this method, the frequency of each word
is recorded, and the sentences are sorted in accordance with the noted fre-
quencies. Sentence rank is incremented for every frequent word that appears
in the sentence. Thus, sentences containing the most frequent words are said
to be salient.
2. TF-IDF method: The trouble with the simple frequency method is that prep-
ositions, determiners, and domain-specific words always acquire the highest
frequency counts. These words do not play any role in determining the impor-
tance of the sentence; instead, they could affect the consistency of the sum-
mary. The TF-IDF method eliminates the impact of these words by comparing
each word’s frequency (f(w)) in the input document with its frequency in all
the background documents (bg(w)).

TFi * IDFi = f(w) * log(bg/(bg(w))) (1)

TFi is the term frequency, IDFi is the inverse document frequency (where i
indicates the ith word in the input) and bg is the total number of background
documents taken.
3. Sentence length method: Long sentences sometimes include information
that should be included in the summary; hence, sentence length is significant
[20] [14]. For the optimal selection of sentences this method constrains short-
ened and lengthy sentences.

iJET | Vol. 19 No. 1 (2024) International Journal of Emerging Technologies in Learning (iJET) 89
Kirmani et al.

4. Uppercase method: This method tries to identify the important words by

assigning higher scores to the words containing uppercase alphabets [20] [14].
This method accounts for the importance of acronyms, initials, and
proper names.
5. Sentence position method: The position of the sentence with respect to
the entire input document is used as a criterion in this method to indicate
sentence importance [6] [4]. In this paper, the leading sentence of the docu-
ment is considered important and, therefore, should be the candidate for the
final summary.
6. Cue-phrase method: This method identifies summary sentences on the basis
of cue-phrases (in particular, salient, the best, hardly, the most important,
according to the literature, etc.) present in the input sentences.
7. Proper noun method: Sentences which contain one or more than one proper
nouns are given higher scores.
8. Numerical data method: Important information such as bank transactions,
amount, balance, event date, time, etc., is always numerical. This method
treats those sentences as important ones that incorporate numerical data.
9. Similarity of title to sentence method: The method checks the similarity
between a sentence and the title of a document. Thus, sentences similar to the
title become summary candidates.

Table 2. Presents a comparative study on various extractive text summarization methods

Authors Year Input Methods Results

Lloret and Palomar 2009 Single Document Word Frequency The system performance
was improved by 10% over DUC 2002. [28]
Gupta et al. 2011 Single Document Word Frequency, Cue-Phrase and The deficiently connected
Sentence Position. Sentences were removed from the summary
which resulted in more coherent summary. [12]
Kulkarni and Prasad 2010 Single Document Word Frequency, Cue-Phrase, This system performed semantically better than
Numerical Data and Sentence – the MS-word Summarizer.
Title Similarity.
Abuobieda, Salim, 2012 Single Document Numerical Data and Sentence–Title Results in optimal features selection for the
Albaham, Osman, and Similarity, Sentence Length, Word summarization process [29]
Kumar Frequency, Sentence Position.
Satoshi et al. 2001 Single Document TF-IDF, Sentence Position and Having compression ratio equal to 10% the
Sentence – Title Similarity. system obtained better results compared to lead-
based and TF-based systems.
Murdock 2006 Single Document TF-IDF Shows that approaches for language modeling
that employ Statistical Translation Models are
ineffective.
Fattah and Ren 2009 Single Document Proper Noun, Sentence Position, Promising results were achieved when the
Sentence Length, Numerical Data system was tested at different compression rates.
and Sentence – Title Similarity.
Barrera and Verma 2012 Single Document TextRank, POS Tagging, wordnet System outperformed baseline and was
evaluated on DUC 2002 and set of articles from
scientific magazine. [30]

90 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)
 Analysis of Abstractive and Extractive Summarization Methods

B) Semantical methods
The summarizers that use statistical methods for the extraction of salient infor-
mation, to some extent, fail to generate coherent summaries as they do not explore
the meaning of the input text. Semantical methods such as emotion used in [19] [13]
generate rational summaries by understanding the sentiment or emotion of every
sentence in the input document.

4 ABSTRACTIVE SUMMARIZATION METHODS

Methods for abstractive text summarization are broadly categorized as structure-

based methods and semantic-based methods.

a) Structure-based methods: Structure based methods represent input document

using structures like trees, templates, cognitive schemas, etc. Important informa-
tion is then encoded in these structures. The structure-based methods include:
1. Tree based: In this method, the input text document is represented as a
dependency tree. Important information is identified by applying different
algorithms, such as the theme intersection, etc. Finally, for summary genera-
tion, language generators are used.

Fig. 1. Block diagram of tree-based method [14] [28]

2. Template based: In this method, the input text document is represented as

a template. For mapping text snippets into template slots, extraction rules or
linguistic patterns are used. Important data is indicated by text snippets.

Fig. 2. Block diagram of template-based method [9]

3. Multimodal-semantic method: This method takes input in the form of

both text and images. This multimodal input document is represented by a

iJET | Vol. 19 No. 1 (2024) International Journal of Emerging Technologies in Learning (iJET) 91
Kirmani et al.

semantic model that clearly apprehends the conception and the relationship
among them. Some measures are used to score the important concepts. Finally,
concepts chosen for summary are framed as sentences.
4. Information item-based method: This technique transforms the infor-
mation provided in the input text into an abstract representation. From this
abstract representation, the contents of the summary are selected.
5. Semantic graph-based method: The process creates a rich semantic graph
(RSG) from the supplied text document. Then reduction of the rich semantic
graph takes place. Finally, this reduced, rich semantic graph acts as the basis
for the final abstract summary generation [21].
6. Ontology-based method: The domain-related documents can be coherently
summarized by ontology-based methods because ontology can better repre-
sent a domain as each domain possesses a knowledge structure.
7. Lead and body phrase: The technique attempts to rebuild the lead sentence
by inserting or substituting phrases that have similar triggers in the body and
the lead sentences.

Fig. 3. Block diagram of body and lead phrase method [5]

8. Rule-based method: Using this approach, the original a text document is rep-
resented as list of aspects and categories. Information extraction rules are used
to generate candidates. Then the best candidates are selected by the content
selection module. Finally, a summary is generated using generation patterns.
9. Semantic-based methods: Semantic-based methods transforms the input
document into a semantic representation. This intermediate representation
is then supplied to the natural language generation system (NLGS). NLGS pro-
cesses the linguistic data to identify verb phrases and noun phrases.

Fig. 4. Block diagram of semantic graph based method [4]

92 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)
 Analysis of Abstractive and Extractive Summarization Methods

10. Distributional Semantic techniques: Distributional semantic models

(DSM), often referred to as “distributional similarity” models, are predicated
on the idea that it is possible to deduce a word’s meaning from its usage—
that is, from how frequently it appears in text—at least to a certain extent.
By statistically analyzing the situations in which words occur, these models
dynamically construct semantic representations in the form of high-dimen-
sional vector spaces. Distributional semantic models make use of the distri-
butional hypothesis, which claims that words deployed in the same context
express equivalent meanings. These models are broad and useful for any
application because they are trained on huge external datasets. Because they
are not domain-specific, these models are adaptable. These characteristics
make these models standout selections for the semantics extraction issue.

Some of semantic distributional models are discussed in detail as follows

i. Word2vec: A two-layer neural network model called Word2Vec can provide excel-
lent text semantics. A word is converted by the model into a multidimensional vec-
tor space embedding. As an output, the model creates a vector from a word, thus
the name. The vectors generated are detailed semantic expansions of the original
word. Skip-gram and continuous bag of vectors (CBOW) are the two architectures
that Word2Vec offers. In contrast to the skip-gram, which forecasts the context
from the given word, the CBOW model forecasts the word from its context [21–23].
ii. Glove: It is a method of unsupervised learning that creates word-to-vector repre-
sentations. It establishes the paradigm for converting the frequency of terms that
co-occur in the whole of the data. The inference is made using data from collected
global word-word co-occurrence statistics [24] [25].
iii. FasText: It’s an open-source, free program that teaches users how to utilize
classifiers and text representations. It is based on the approximation approach,
dimension reduction, and n-gram characteristics. The input tokens are converted
into n-gram characters. It is a tool for classifying phrases and effectively learning
token representations [26].
iv. BioBERT: The term stands for bidirectional encoder representations from trans-
formers for biomedical text mining. It is an advanced language representation model
for the biomedical sector, trained in advance, using a large biomedical corpus [27].

Table 3. Presents a comparative analysis on various abstractive text summarization methods

Authors Year Input Methods Results

Barzilay and McKeown 1999 Multiple Documents Tree based The system was able to correctly identify
74%, 69%, 74%, and 56% of predicate-argument
structures, the subjects, the main verbs, and the
other constituents in the list respectively [14].
Barzilay and McKeown 2005 Multiple Documents Tree based Summary that is grammatically strict [31].
Harabagiu 2002 Single and Template based Evaluated GISTEXTER using DUC 2002 and
and Lacatusu Multiple Documents resulted in coherent and organized summary [9].
Lee and Jian 2005 Single Document Ontology based Showed that for summarization of news
articles news agent operates effectively [20].
Tanaka and Kinoshita 2009 Single Document Body and Lead Operations such as insertion and replacements
phrase are performed on phrases [5].
Greenbacker 2011 Includes both text and images Multimodal Abstract summaries that incorporate concepts
Multimodal Document Semantic Model obtained by graphical data [4].

(Continued)

iJET | Vol. 19 No. 1 (2024) International Journal of Emerging Technologies in Learning (iJET) 93
Kirmani et al.

Table 3. Presents a comparative analysis on various abstractive text summarization methods (Continued)
Authors Year Input Methods Results
Genest and Lapalme 2011 Multiple Documents INIT based Evaluated system using TAC 2010; average
performance was satisfactory [7].
Moawad and Aref 2012 Single Document Semantic Graph Reduced input text document to almost
based 50% [8].
Genest and Lapalme 2012 Multiple Documents Rule based Results in high density information summary [1].
Yash Sharma et al. 2017 Multiple Documents Word2vec Test papers with a number between 50 and 284
were used to determine results for the ROUGE 1,
ROUGE 2, and ROUGE L tests.
The table shows the ROUGE scores’ 95%
confidence intervals [22].
Enise Karakoc et al. 2019 Single Document FasText ROUGE scores were outperformed by semantic
similarity scores in terms of performance [23].
Mohd Mudasir et al. 2020 Single Document Word2vec, 34%, 7%, and 20% were the values of precision
Clustering respectively [21].
Algorithm, NLTK, NLP
S Kulkarni et al. 2020 Single document Glove Glove is used to construct corpora using
second-order random walks and calculate graph
node embedding [25].

5 CONCLUSION

Manual text summarization entails a considerable number of impartial and knowl-

edgeable experts and a lot of time. However, digital data, which is accessible worldwide
is increasing exponentially, making the utilization of an automatic text summarization
tool inevitable in order to achieve coherent summaries in less time. Automatic text sum-
marization approaches are widely categorized as extractive approaches and abstrac-
tive approaches. This paper presents a review of both extractive as well as abstractive
approaches. Different extractive and abstractive methods are explored, and a compar-
ative analysis of the different methods implemented in the literature is presented.

6 REFERENCES

[1] P.-E. Genest and G. Lapalme, “Fully abstractive approach to guided summarization,” in
Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics
(Volume 2: Short Papers), 2012, pp. 354–358.
[2] R. Aliguliyev, “Automatic document summarization by sentence extraction,” vol. 12,
no. 5, pp. 5–15, 2007.
[3] E. Hovy and C.-Y. Lin, “Automated text summarization and the SUMMARIST system,”
in Proceedings of the TIPSTER Text Program, Association for Computational Linguistics,
Baltimore, Md, USA, 1998, pp. 197–214.
[4] C. Greenbacker, “Towards a framework for abstractive summarization of multimodal
documents,” in Proceedings of the ACL 2011 Student Session, pp. 75–80, 2011.
[5] H. Tanaka, A. Kinoshita, T. Kobayakawa, T. Kumano, and N. Kato, “Syntax-driven sen-
tence revision for broadcast news summarization,” in Proceedings of the 2009 Workshop
on Language Generation and Summarisation (UCNLG + Sum 2009), 2009, pp. 39–47. https://
doi.org/10.3115/1708155.1708163

94 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)
 Analysis of Abstractive and Extractive Summarization Methods

[6] H. P. Edmundson, “New methods in automatic extracting,” Journal of the ACM (JACM),
vol. 16, no. 2, pp. 264–285, 1969. https://doi.org/10.1145/321510.321519
[7] P.-E. Genest and G. Lapalme, “Framework for abstractive summarization using text-to-
text generation,” in Proceedings of the Workshop on Monolingual Text-to-Text Generation,
2011, pp. 64–73.
[8] I. F. Moawad and M. Aref, “Semantic graph reduction approach for abstractive text sum-
marization,” in 2012 Seventh International Conference on Computer Engineering & Systems
(ICCES). IEEE, 2012, pp. 132–138. https://doi.org/10.1109/ICCES.2012.6408498
[9] S. M. Harabagiu and F. Lacatusu, “Generating single and multi-document summaries
with gistexter,” in Document Understanding Conferences, 2002, pp. 11–12.
[10] H. P. Luhn, “The automatic creation of literature abstracts,” IBM Journal of Research and
Development, vol. 2, no. 2, pp. 159–165, 1958. https://doi.org/10.1147/rd.22.0159
[11] A. Nenkova, K. McKeown et al., “Automatic summarization,” Foundations and Trends® in
Information Retrieval, vol. 5, no. 2–3, pp. 103–233, 2011. https://doi.org/10.1561/1500000015
[12] V. Gupta and G. S. Lehal, “A survey of text summarization extractive techniques,” Journal
of Emerging Technologies in Web Intelligence, vol. 2, no. 3, pp. 258–268, 2010. https://doi.
org/10.4304/jetwi.2.3.258-268
[13] Q. Wu, X. He, Q. V. Nguyen, W. Jia, and M. Huang, “2008 IEEE 8th international confer-
ence on computer and information technology,” CIT, p. 295, 2008.
[14] R. Barzilay, K. McKeown, and M. Elhadad, “Information fusion in the context of multi-
document summarization,” in Proceedings of the 37th Annual Meeting of the Association for
Computational Linguistics, pp. 550–557, 1999. https://doi.org/10.3115/1034678.1034760
[15] L. Chengcheng, “Automatic text summarization based on rhetorical structure the-
ory,” in 2010 International Conference on Computer Application and System Modeling
(ICCASM 2010), IEEE, 2010, vol. 13, pp. V13-595–V13-598. https://doi.org/10.1109/
ICCASM.2010.5622918
[16] M. Yousefi-Azar and L. Hamey, “Text summarization using unsupervised deep learn-
ing,” Expert Systems with Applications, vol. 68, pp. 93–105, 2017. https://doi.org/10.1016/
j.eswa.2016.10.017
[17] R. Nallapati, B. Zhou, C. Gulcehre, B. Xiang et al., “Abstractive text summarization using
sequence-to-sequence rnns and beyond,” ArXiv Preprint ArXiv:1602.06023, 2016. https://
doi.org/10.18653/v1/K16-1028
[18] I. K. Bhat, M. Mohd, and R. Hashmy, “Sumitup: A hybrid single-document text summa-
rizer,” in Soft Computing: Theories and Applications: Proceedings of SoCTA 2016, Springer,
vol. 1, pp. 619–634, 2018. https://doi.org/10.1007/978-981-10-5687-1_56
[19] J. Kupiec, J. Pedersen, and F. Chen, “A trainable document summarizer,” in Proceedings
of the 18th Annual International ACM SIGIR Conference on Research and Development in
information retrieval, pp. 68–73, 1995. https://doi.org/10.1145/215206.215333
[20] C.-S. Lee, Z.-W. Jian, and L.-K. Huang, “A fuzzy ontology and its application to news sum-
marization,” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
vol. 35, no. 5, pp. 859–880, 2005. https://doi.org/10.1109/TSMCB.2005.845032
[21] M. Mohd, R. Jan, and M. Shah, “Text document summarization using word embed-
ding,” Expert Systems with Applications, vol. 143, p. 112958, 2020. https://doi.org/10.1016/
j.eswa.2019.112958
[22] Y. Sharma, G. Agrawal, P. Jain, and T. Kumar, “Vector representation of words for senti-
ment analysis using glove,” in 2017 International Conference on Intelligent Communication
and Computational Techniques (ICCT). IEEE, pp. 279–284, 2017. https://doi.org/10.1109/
INTELCCT.2017.8324059
[23] E. Karakoç and B. Yılmaz, “Deep learning based abstractive turkish news summariza-
tion,” in 2019 27th Signal Processing and Communications Applications Conference (SIU).
IEEE, pp. 1–4, 2019. https://doi.org/10.1109/SIU.2019.8806510

iJET | Vol. 19 No. 1 (2024) International Journal of Emerging Technologies in Learning (iJET) 95
Kirmani et al.

[24] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word represen-
tation,” in Proceedings of the 2014 Conference on Empirical Methods in Natural Language
Processing (EMNLP), pp. 1532–1543, 2014. https://doi.org/10.3115/v1/D14-1162
[25] S. Kulkarni, J. K. Katariya, and K. Potika, “Glovenor: Glove for node representations with
second order random walks,” in 2020 IEEE/ACM International Conference on Advances
in Social Networks Analysis and Mining (ASONAM). IEEE, 2020, pp. 536–543. https://doi.
org/10.1109/ASONAM49781.2020.9381347
[26] A. Joulin, E. Grave, P. Bojanowski, and T. Mikolov, “Bag of tricks for efficient text classifi-
cation,” ArXiv Preprint ArXiv:1607.01759, 2016. https://doi.org/10.18653/v1/E17-2068
[27] J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, and J. Kang, “Biobert: A pre-trained
biomedical language representation model for biomedical text mining,” Bioinformatics,
vol. 36, no. 4, pp. 1234–1240, 2020. https://doi.org/10.1093/bioinformatics/btz682
[28] E. Lloret and M. Palomar, “A gradual combination of features for building automatic
summarisation systems,” in International Conference on Text, Speech and Dialogue, 2009,
pp. 16–23. https://doi.org/10.1007/978-3-642-04208-9_6
[29] A. Abuobieda et al., “Text summarization features selection method using pseudo Genetic-
based model,” in 2012 International Conference on Information Retrieval & Knowledge
Management, Kuala Lumpur, Malaysia, 2012, pp. 193–197. https://doi.org/10.1109/
InfRKM.2012.6204980
[30] A. Barrera and R. Verma, “Combining syntax and semantics for automatic extractive
single-document summarization,” in Computational Linguistics and Intelligent Text
Processing, 2012, pp. 366–377. https://doi.org/10.1007/978-3-642-28601- 8_31
[31] R. Barzilay and K. McKeown, “Sentence fusion for multidocument news summari-
zation,” Computational Linguistics, vol. 31, no. 3, pp. 297–328, 2005. https://doi.org/
10.1162/089120105774321091

7 AUTHORS

Mahira Kirmani is a Research scholar at University Institute of Computing,

Chandigarh University, Punjab, India (E-mail: mahirakirmani68@gmail.com; ORCID:
0000-0001-5651-4332).
Gagandeep Kaur is an Assistant Professor at University Institute of Computing,
Chandigarh University, Punjab, India (E-mail: gagandeepkaurlogani@gmail.com;
ORCID: 0000-0002-1513-8446).
Mudasir Mohd is a Senior Assistant Professor in the Department of Computer
Application South Campus, University of Kashmir, J&K, India (E-mail: mudasir.
mohammad@kashmiruniversity.ac.in; ORCID: 0000-0003-1597-146X).

96 International Journal of Emerging Technologies in Learning (iJET) iJET | Vol. 19 No. 1 (2024)