2020 International Conference on Computational Science and Computational Intelligence (CSCI)

Recent Progress on Text Summarization

Suad Alhojely Jugal Kalita
College of Engineering and Applied Science College of Engineering and Applied Science
The University of Colorado at Colorado Springs The University of Colorado at Colorado Springs
Colorado Springs, USA Colorado Springs, USA
[email protected] [email protected]

Abstract—Text summarization produces a summary of a first converts the important sentences extracted from a doc-
document by highlighting its most important content ideas. ument into an understandable and coherent semantic form,
Researchers have been developing text summarization techniques
since the 1950s. Most summarization deals with summaries of and then generates the summary from this internal form,
single documents, but recent summarization efforts have also pro- thus potentially changing the original sentences. Hybrid text
duced summaries from clusters of documents. This paper reviews summarization combines both extractive and abstractive sum-
recent approaches in three categories: Extractive, Abstractive, marization.
and Hybrid text summarization along with specific methods Generally, the processing architecture of all automatic text
within each category. Most authors focus on the Extractive
approach, but the generated summaries are very different from summarization systems contains three steps. The first is pre-
human summaries regardless of techniques used. processing to usually identify words, sentences and other
Index Terms—text summarization, extractive summarization, structural components of the text. The second is processing,
abstractive summarization, hybird summarization, approaches which converts the input text to a summary by using a text
summarization method. The third is post-processing, which
I. I NTRODUCTION fixes problems in the created draft summary [5].
Several recent surveys have been published on automatic
The amount of data and information on the Internet con- text summarization, and most focus on extractive summariza-
tinues to increase every day in the form of web pages, tion techniques [1] because abstractive summarization is diffi-
articles, academic papers, and news items. In spite of the cult and requires comprehensive Natural Language Processing
abundance, it is difficult to find information needed efficiently (NLP).
because most information is irrelevant to a particular user’s Most state-of-the-art papers focus on a part of automatic
needs at a particular time. Therefore, the need for auto- text summarization such as focusing on one approach, or
matic summarization and extraction of relevant information on one specific domain in automatic text summarization.
continues to be a productive research area within natural Mahajani et al. [6] recommended using a hybrid system that
language processing. Automatic summarization helps extract combines extractive and abstractive summarization approaches
useful information while discarding the irrelevant. It can also to leverage their respective advantages. Therefore, the goal of
improve the readability of texts, and decrease the time that this survey is to present various methods in text summarization
users spend in searching. Researchers have been trying to to help readers understand how a good summary can be
perform suitable automatic text summarization since the late generated by combining more than one approach or method.
1950s. The goal is to generate summaries, combining the main The present review is organized into three sections: a
points in a readable and cohesive way, without having unuseful brief introduction to text summarization, text summarization
or repeated information [1]. approaches, and the conclusion of the paper. The architectures,
Text summarization methods usually extract important advantages, and disadvantages of the approaches are included
words, phrases or sentences from a document and use these in detail in the second part.
words, phrases, or sentences to create a summary. Text
summarization can be classified into single document and II. T EXT S UMMARIZATION A PPROACHES
multi-document summarization, depending on the number of Conceptually, there are three approaches for text summa-
input documents. Single document text summarization only rization, which are extractive, abstractive, and hybrid sum-
accepts one document as input [2], whereas multi-document marization. Within each approach, there are many methods
summarization accepts more than one document, where each and techniques. Every approach has some advantages and
document is related to the main topic. Meaningful information disadvantages. A brief overview of the approaches along with
is extracted from each document and then gathered together some specific methods are shown in Figure 1.
and organized to generate a summary [3] [4].
Extractive summarization chooses important sentences from A. Extractive Summarization
a document and combines them to create a summary without The architecture for extractive summarization includes three
changing the original sentences. Abstractive summarization steps: Pre-processing, Processing, and Post-processing, as

978-1-7281-7624-6/20/$31.00 ©2020 IEEE 1503

DOI 10.1109/CSCI51800.2020.00278
&"# (" !

&" "$# (" !" "$# (" ' # ("

"! "! "!

"# " #"# !

&" "$"
#!"
!" "$"!
"!
""!" #!
! &" "$"%
! !" "$"!
"'!
#(('

"!!

""!
" !
" #"

!

Fig. 1. Text Summarization Approaches along with their Methods

Advantages and Disadvantages for Extractive Summa-

rization
Since extractive summarization depends on directly gener-
ating the summary from the text without changing the content
sentences in any way, it is faster and simpler [10].
The disadvantage of this approach is that it is not the same
as how humans write the summary. The approach usually re-
sults in the reduction of semantic quality and cohesion because
of wrong connections between sentences in the generated
summary, making the flow stilted and unnatural [11]. The
generated summary may not be accurate enough, and not cover
all important content sentences in the input document [12].
However, if the output summary is long enough, the issue of
Fig. 2. Extractive Text Summarization Architecture, Adopted from [5] missing significant sentences may not arise. But it may contain
unnecessary parts that may not be needed in the summary,
making it longer than necessary [9].

shown in Figure 2. Pre-processing performs tasks such as B. Extractive Summarization Methods

tokenization and extraction of sentences and paragraphs. The There are various extractive summarization methods for se-
processing step creates appropriate representation of the input lecting and scoring sentences. These include Conceptual, Lin-
text using techniques such as N-grams and graphs, or performs guistic, Statistical, Machine Learning methods, Fuzzy logic,
neural network based feature extraction and encoding [2] and Deep learning as presented in Figure 1.
followed by scoring each sentence depending on input text 1) Concept Methods: Such a method produces a summary
representation [7]. After that, the approach chooses highly of the concepts present in a document that can be found
ranked sentences and links them together as a summary [7] in external information repositories like WordNet [13] and
[8]. Post-processing involves steps such as changing pronouns Wikipedia. Depending on the concepts extracted, the important
with their antecedents, and rearranging the extracted sentences sentences are identified based on connection to external infor-
[9]. mation bases instead of words. From the external information

1504
base’s scores, a graph model or vector is built to produce the manual work to generate extractive summaries for training
connection between the sentences and the concepts. [12].
The concept methods of summarization can cover a very 5) Fuzzy Logic Based Methods: Such text summarization
large number of concepts because WordNet and Wikipedia are methods use a multiple-valued system known as fuzzy logic.
large repositories. However, such a method depends on high Fuzzy logic produces an efficient way to provide feature values
quality similarity measurements to decrease redundancies in for sentences that are between the two logical values “one”
calculating concept-sentence correlations [12]. and “zero”, because these two values often do not represent
2) Linguistic Methods: A linguistic method focuses on the “real world” [20]. For ranking sentences, the first step is
the relationships between words and concepts to get to the to choose a group of features for every sentence. The second
meaning to generate the summary. Abstractive summarization step is to apply the fuzzy logic concept to get a score for every
includes some level of semantic processing, so that, it can be sentence based on the importance of the sentence. This means
thought also of as a linguistic method. every sentence has a score value from 0 to 1, depending on
Linguistic methods are useful because they try to understand the features [1].
the meaning of every sentence in a document. However, this Fuzzy logic represents uncertainties in selecting a sentence
method is time-consuming requiring high effort. A linguistic as a ‘fuzzy’concept [20]. However, one negative factor is re-
method also needs a large amount of memory for saving dundancy in the selected sentences for the summary, impacting
additional linguistic repositories such as WordNet. It needs the quality of the generated summary. Therefore, a redundancy
powerful processors for complicated linguistic processing [14]. removal technique is required to enhance the quality of the
3) Statistical Methods: Such methods use statistical fea- generated summary [21].
tures of the document to identify the important pieces of 6) Deep Learning Methods: Kobayashi et al. [22] suggest a
the text. In a statistical method, a sentence is selected based system for text summarization using document level similarity
on features like word frequency, position of the sentence, depending on embeddings. They assume that an embedding
indicator phrases, title, location, and other features regardless of a word represents its meaning, a sentence considered as a
of the meaning of the sentence. The method calculates the bag-of-words, and a document as a bag-of-sentences. They
scores of the selected sentences and chooses a few highest formalize their task as the problem of maximizing a sub-
scoring sentences to create the summary [15] [9]. modular function which is identified by a negative summation
Baxendale [16] focused on the position of sentences in his of closest neighbors’ distance on embedding distributions.
summarization research. He found that the best locations for They found that the document level similarity is more complex
the most important parts of the paragraph are the first and in meaning compared with sentence-level similarity. In Chen
last sentences. He examined 200 paragraphs, and concluded et al. [23], they suggest automatic text summarization that
that the topic sentences are included the first sentence of the used a reinforcement learning algorithm and Recurrent Neural
paragraph in around 85% of the cases while 7% it was in the Network (RNN) model with a single document. By using a
last sentence of the paragraph. sentence level selective encoding technique, they select the
Statistical methods do not take into account the meaning of significant features, generating the summary sentences.
sentences, and as a result, they may produce low-quality sum- In deep learning methods, the network could be trained
maries. Statistical methods require low memory and processor depending on the reader’s style, and the features can be
capacity [15]. changed depending on the user’s requirement. However, it is
4) Machine Learning Methods: The idea behind machine difficult to identify how the network generates a decision [12].
learning is to use a training set of data to train the summa- Recent research shows that using a combination of various
rization system, which is modeled as a classification problem. methods helps produce a better summary by taking the ad-
Sentences are classified into two groups: summary sentences vantage of the strengths of the individual methods [12], [24],
and non-summary sentences [17]. The probability of choosing [25], [26]. For instance, Moratanch and Chitrakala [12] used
a sentence for a summary is estimated according to the training a combination of both graphs and concept based methods to
documents and corresponding extractive summaries [18]. The generate summaries. Mao et al. [26] combine three different
steps for ranking sentences in Machine Learning methods methods of supervised learning with unsupervised learning to
are extracting features from a document, and feeding those create a summary for a single document.
features to a machine learning algorithm that gives an output Combining different features together may also help pro-
score as a value [12]. Some of the common machine learning duce better outcomes during the calculation of the weights of
methods used for text summarization are linear regression, sentences [1].
naı̈ve Bayes, support vector machine, artificial neural net-
works, and fuzzy logic [19] [15]. C. Abstractive Summarization
A large training data set is necessary to improve the choices Abstractive text summarization creates a summary of a doc-
of sentences for the summary [12]. A simple regression model ument by extracting and understanding the concepts present in
may be able to produce better output when compared with the the text during processing [27] [28]. It paraphrases the text,
other classifiers [15]. Every sentence in the basic text must be but does not directly copy from the content of the original text
labeled as a summary or non-summary, demanding extensive [29]; instead it creates new sentences that better reflect the

1505
as trees, graphs, templates, rules, and ontologies. Therefore,
it recognizes in the input document, the most important
information, and then using the previously mentioned struc-
tures, it generates the abstractive summary. The semantics-
based construction of the input document generates a semantic
representation by using information items, semantic graphs,
and predicate-argument structures. Then, using approaches
in natural language generation, it generates the abstractive
summary [35].
1) Structure-Based Methods:
Templates-Based Methods:
Human summaries tend to use certain characteristic sen-
tence structures in some domains. These can be identified as
Fig. 3. Abstractive Text Summarization Architecture, Adapted from [5] templates. To perform abstractive summarization, the informa-
tion in the input document is used to fill slots in appropriate
pre-defined templates based on the input document’s style [36].
human way of constructing summaries. As a result, the input Text snippets can be extracted using rules and linguistic cues,
content needs more analysis for abstractive summarization to fill template slots [35].
[30]. Rule-based Methods:
The processing architecture for abstractive summarization To find the important concepts in the input document and
is presented in Figure 3. It is composed of Pre-processing, use them in the generated summary, one needs to define rules
Processing that contains two sub-steps, and Post-processing. and categories. To use these methods, one needs to classify the
For example, Moratanch and Chitrakala create an internal input document based on the concepts and terms present in it,
semantic representation and then use various techniques to create relevant questions depending on the domain of the input
create summaries [31]. document, answer the questions by detecting the concepts and
Advantages and Disadvantages of Abstractive Summa- terms in the document, and feed the answers into patterns to
rization: generate the summary [35].
Some of the advantages of abstractive summarization are Tree-based Methods:
that the generated summary is created to be different from To perform abstractive summarization in tree-based meth-
the original text by using more resilient expressions based ods, one needs to cluster similar sentences in the input that
on paraphrasing [11]. So, the generated summary is likely to have related information, and then work with these sentence
be closer to a human summary [32]. Compared to extractive clusters for the summary [35]. Similar sentences are formu-
summarization, abstractive summarization can decrease the lated into trees, parsers are applied to build the dependency
amount of generated text and produce a summary that removes trees, a popular tree based representation. Then, a process
any redundancy, obtaining a concise and expressive summary such as pruning linearization is used to produce trees in order
[33]. to generate summary sentences from some of the sentence
Some of the disadvantages of abstractive summarization clusters [35].
are that it is difficult to perform high-quality abstractive Graph-Based Methods: The authors in [37] used a graph
summarization [11]. It is difficult to create a good abstractive model which contains nodes, with each node expressing a
summary because it needs to use natural language generation word and positional information, that is connected to other
technology, which still needs a lot of progress [34]. Current ab- nodes. The structure of sentences is represented by directed
stractive summarization approaches seem to create repetitions edges. The steps for the graph method contain constructing a
in word choice. In addition, good abstractive summarization textual graph representing the source document and generating
should be able to explain why it creates new sentences abstractive summary. Such a method explores and scores
in the summary, which is difficult to do. The approach is many sub-paths in the graph in order to create the abstractive
also unable to handle out-of-vocabulary words properly [11]. summary [37].
Furthermore, the approach’s ability is constrained by what Ontology-Based Methods:
underlying semantic representation it uses, because a system Ontology-Based methods generate abstractive summariza-
cannot generate a summary if its representation scheme cannot tion from an input document by utilizing an ontology [38].
capture necessary nuances and details [9]. Many documents in specialized domains are connected to a
domain specific ontology, and can be mapped to such an
D. Abstractive Summarization Methods ontology. The mapping is traversed to generate a summary
Abstractive summarization methods can be classified into [39].
three categories, which are structure-based, semantics-based, 2) Semantics-Based Methods : These methods process the
and deep learning-based methods [35] as shown in Figure 1. input text to obtain semantic representations such as informa-
A structure-based approach uses pre-defined structures such tion items, semantic graphs, and predicate-argument structures.

1506
The representation is processed to provide the abstractive
summarization by performing word choices, and stringing the
words together using verb and noun phrases [35]. The authors
in [40] perform multi-document abstractive summarization by
extracting predicate-argument structures from the input text
by performing semantic role labeling. By using a semantic
similarity measurement, they cluster the semantically similar
predicate-argument structures in the text, and then score the
predicate-argument structures using feature weighting. Finally,
they use language generation approaches to create sentences
from predicate-argument structures. Fig. 4. Hybrid Text Summarization Architecture, Adopted from [5]
3) Deep Learning-Based Methods: Recent research in
generating abstractive summarization has used deep sequence-
to-sequence learning [11]. In many different NLP tasks such The advantages of hybrid summarization accrue from the
as machine translation, sequence-to-sequence learning has advantages of both approaches, and the two approaches are
led to good results [34]. RNNs with attention models have considered complementary [34]. On the other hand, the dis-
accomplished promising results in text summarization. Deep advantages of it are that the generated summary is based on
learning-based methods are being actively explored, and re- extracted sentences rather than the original text, which leads to
searchers are trying to solve many deep learning issues. Some generating low quality abstractive summarization. Researchers
of the issues are the inability to handle out-of-vocabulary who use the extractive approach are usually able to obtain a
words, and generation repeated phrases or words [11]. coherent and meaningful summary [15] because the abstractive
Abstractive summarization has recently concentrated on approach is complicated and requires comprehensive process-
utilizing deep learning methods, particularly for short text ing of natural language, which is not yet possible.
summarization [41]. It is a recommendation by some to use
more than one method to produce a better abstractive summary
by taking advantage of each method. Using different text F. Hybrid Summarization Methods
summarization algorithms on the same input document will We discuss two published methods for hybrid summariza-
produce different summaries. To generate a better summary, it tion (as shown in Figure 1): Extractive to Abstractive, and
is necessary to combine outputs of various text summarization Extractive to Shallow Abstractive.
algorithms rather than using single algorithms [42].
1) Extractive to Abstractive Methods: The approach in
Usually, structure-based methods are used as extractive
these methods is that one starts by using any one of the
techniques for generating hybrid summaries while semantics-
extractive text summarization methods. Then, one applies any
based or deep learning-based methods are used to generate
one of the abstractive text summarization methods on the
abstractive summaries [35]. For instance, one of these methods
extracted sentences.
can be used in the pre-processing step to select the important
phrases, and the other method to create the abstractive sum- Researchers in [34] suggest a hybrid text summarization
marization [35]. The authors in [41] suggest a combination of method for long text called EA-LTS, containing two phases.
semantics-based data transformation, followed by a encoder- The first phase is the extraction phase, which extracts key
decoder deep learning models for abstractive summarization. sentences by utilizing a graph model. The second phase is the
abstraction phase which builds an RNN based encoder-decoder
E. Hybrid Summarization with attention mechanisms and pointers, in order to create a
The hybrid text summarization approach combines both summary.
extractive and abstractive text summarization. The architecture 2) Extractive to Shallow Abstractive Methods: In the be-
for hybrid text summarization contains processes as shown in ginning, these methods use any one of the extractive text
Figure 4 [29] [43]. The processes are pre-processing, which is summarization methods. Then, on the extracted sentences they
usually extractive summarization to select and extract key sen- apply a shallow abstractive text summarization method, which
tences [34]; a summary generation process which is abstractive uses one or more techniques such as information fusion,
summarization to create the final abstractive summary; and information compression [44], and synonym replacement [45].
post processing, which makes sure the created sentences are The authors in [29] present a hybrid text summarization
valid. Post-processing often uses rules. These rules enforce method for a single document called SumItUp. This hybrid
heuristics such as the length has to be at least 3 words in text summarization method contains two phases. The first
a sentence, each sentence has to include a verb, and the phase is extractive sentence selection which uses semantic and
sentences must not end with a preposition, an interrogative statistical features to create a summary. The second phase is
word, an article, or a conjunction [44]. the abstractive summary generation that converts the extractive
Advantages and Disadvantages of Hybrid Summariza- summary to the abstractive summary by feeding the extracted
tion sentences to a language generator.

1507
At this time, the hybrid text summarization approach is [16] P. B. Baxendale, “Machine-made index for technical literature—an
experiment,” IBM Journal of Research and Development, vol. 2, no. 4,
pursued by many researches because it helps leverage the pp. 354–361, 1958.
advantages of both extractive and abstractive approaches [6]. [17] R. Oak, “Extractive techniques for automatic document summarization:
a survey,” International Journal of Innovative Research in Computer and
III. C ONCLUSION Communication Engineering, vol. 4, no. 3, pp. 4158–4164, 2016.
[18] B. Othman, M. Haggag, and M. Belal, “A taxonomy for text summariza-
Automatic text summarization is the mechanism that pro- tion,” Information Science and Technology, vol. 3, no. 1, p. 43, 2014.
vides a summary by reducing the size of the original docu- [19] S. K. Bharti, K. S. Babu, A. Pradhan, S. Devi, T. Priya, E. Orhorhoro,
ment, but keeping important information in the original doc- O. Orhorhoro, V. Atumah, E. Baruah, P. Konwar et al., “Automatic key-
word extraction for text summarization in multi-document e-newspapers
ument. However, automatic text summarization is still a chal- articles,” European Journal of Advances in Engineering and Technology,
lenging task, and the results are still far unlike quality human vol. 4, no. 6, pp. 410–427, 2017.
summaries even though many techniques have been proposed. [20] N. V. Kumar and M. J. Reddy, “Factual instance tweet summarization
and opinion analysis of sport competition,” in Soft Computing and Signal
Most researchers focus on the extractive approach. So, the Processing. Springer, 2019, pp. 153–162.
literature for extractive summarization is more extensive than [21] D. Patel, S. Shah, and H. Chhinkaniwala, “Fuzzy logic based multi
abstractive summarization. This survey has reviewed different document summarization with improved sentence scoring and redun-
approaches and methods. We conclude that combining two dancy removal technique,” Expert Systems with Applications, vol. 134,
pp. 167–177, 2019.
approaches or methods is likely to produce promising results, [22] H. Kobayashi, M. Noguchi, and T. Yatsuka, “Summarization based on
increasing the quality of the summaries compared to using any embedding distributions,” in Proceedings of the 2015 Conference on
one approaches. Empirical Methods in Natural Language Processing, 2015, pp. 1984–
1989.
R EFERENCES [23] L. Chen and M. Le Nguyen, “Sentence selective neural extractive
summarization with reinforcement learning,” in 2019 11th International
[1] N. Nazari and M. Mahdavi, “A survey on automatic text summarization,” Conference on Knowledge and Systems Engineering (KSE). IEEE,
Journal of AI and Data Mining, vol. 7, no. 1, pp. 121–135, 2019. 2019, pp. 1–5.
[2] M. Joshi, H. Wang, and S. McClean, “Dense semantic graph and its [24] M. Mohd, R. Jan, and M. Shah, “Text document summarization us-
application in single document summarisation,” in Emerging Ideas on ing word embedding,” Expert Systems with Applications, vol. 143, p.
Information Filtering and Retrieval. Springer, 2018, pp. 55–67. 112958, 2020.
[3] S. Modi and R. Oza, “Review on abstractive text summarization tech- [25] N. Alami, M. Meknassi, and N. En-nahnahi, “Enhancing unsupervised
niques (atst) for single and multi documents,” in 2018 International neural networks based text summarization with word embedding and
Conference on Computing, Power and Communication Technologies ensemble learning,” Expert Systems with Applications, vol. 123, pp. 195–
(GUCON). IEEE, 2018, pp. 1173–1176. 211, 2019.
[4] S. Gholamrezazadeh, M. A. Salehi, and B. Gholamzadeh, “A compre- [26] X. Mao, H. Yang, S. Huang, Y. Liu, and R. Li, “Extractive summariza-
hensive survey on text summarization systems,” in 2009 2nd Interna- tion using supervised and unsupervised learning,” Expert Systems with
tional Conference on Computer Science and its Applications. IEEE, Applications, vol. 133, pp. 173–181, 2019.
2009, pp. 1–6. [27] R. Z. Al-Abdallah and A. T. Al-Taani, “Arabic single-document text
[5] W. S. El-Kassas, C. R. Salama, A. A. Rafea, and H. K. Mohamed, “Au- summarization using particle swarm optimization algorithm,” Procedia
tomatic text summarization: A comprehensive survey,” Expert Systems Computer Science, vol. 117, pp. 30–37, 2017.
with Applications, p. 113679, 2020. [28] K. Krishnakumari and E. Sivasankar, “Scalable aspect-based summa-
[6] A. Mahajani, V. Pandya, I. Maria, and D. Sharma, “A comprehensive rization in the hadoop environment,” in Big Data Analytics. Springer,
survey on extractive and abstractive techniques for text summarization,” 2018, pp. 439–449.
in Ambient Communications and Computer Systems. Springer, 2019, [29] I. K. Bhat, M. Mohd, and R. Hashmy, “Sumitup: A hybrid single-
pp. 339–351. document text summarizer,” in Soft Computing: Theories and Appli-
[7] A. Nenkova and K. McKeown, “A survey of text summarization tech- cations. Springer, 2018, pp. 619–634.
niques,” in Mining text data. Springer, 2012, pp. 43–76. [30] M. J. Mohan, C. Sunitha, A. Ganesh, and A. Jaya, “A study on ontology
[8] J. Zhu, L. Zhou, H. Li, J. Zhang, Y. Zhou, and C. Zong, “Augmenting based abstractive summarization,” Procedia Computer Science, vol. 87,
neural sentence summarization through extractive summarization,” in pp. 32–37, 2016.
National CCF Conference on Natural Language Processing and Chinese [31] S. Chitrakala, N. Moratanch, B. Ramya, C. R. Raaj, and B. Divya,
Computing. Springer, 2017, pp. 16–28. “Concept-based extractive text summarization using graph modelling
[9] V. Gupta and G. S. Lehal, “A survey of text summarization extractive and weighted iterative ranking,” in International Conference on Emerg-
techniques,” Journal of Emerging Technologies in Web Intelligence, ing Research in Computing, Information, Communication and Applica-
vol. 2, no. 3, pp. 258–268, 2010. tions. Springer, 2016, pp. 149–160.
[10] A. Tandel, B. Modi, P. Gupta, S. Wagle, and S. Khedkar, “Multi-
[32] R. Sun, Z. Wang, Y. Ren, and D. Ji, “Query-biased multi-document
document text summarization-a survey,” in 2016 International Confer-
abstractive summarization via submodular maximization using event
ence on Data Mining and Advanced Computing (SAPIENCE). IEEE,
guidance,” in International Conference on Web-Age Information Man-
2016, pp. 331–334.
agement. Springer, 2016, pp. 310–322.
[11] L. Hou, P. Hu, and C. Bei, “Abstractive document summarization via
[33] D. Sakhare, R. Kumar, and S. Janmeda, “Development of embed-
neural model with joint attention,” in National CCF Conference on
ded platform for sanskrit grammar-based document summarization,” in
Natural Language Processing and Chinese Computing. Springer, 2017,
Speech and Language Processing for Human-Machine Communications.
pp. 329–338.
Springer, 2018, pp. 41–50.
[12] N. Moratanch and S. Chitrakala, “A survey on extractive text summariza-
tion,” in 2017 International Conference on Computer, Communication [34] S. Wang, X. Zhao, B. Li, B. Ge, and D. Tang, “Integrating extractive
and Signal Processing (ICCCSP). IEEE, 2017, pp. 1–6. and abstractive models for long text summarization,” in 2017 IEEE
[13] G. A. Miller, “Wordnet: a lexical database for english,” Communications International Congress on Big Data (BigData Congress). IEEE, 2017,
of the ACM, vol. 38, no. 11, pp. 39–41, 1995. pp. 305–312.
[14] Y. Ko and J. Seo, “An effective sentence-extraction technique using con- [35] S. Gupta and S. Gupta, “Abstractive summarization: An overview of the
textual information and statistical approaches for text summarization,” state of the art,” Expert Systems with Applications, vol. 121, pp. 49–65,
Pattern Recognition Letters, vol. 29, no. 9, pp. 1366–1371, 2008. 2019.
[15] M. Gambhir and V. Gupta, “Recent automatic text summarization [36] H. Lin and V. Ng, “Abstractive summarization: A survey of the state of
techniques: a survey,” Artificial Intelligence Review, vol. 47, no. 1, pp. the art,” in Proceedings of the AAAI Conference on Artificial Intelligence,
1–66, 2017. vol. 33, 2019, pp. 9815–9822.

1508
[37] K. Ganesan, C. Zhai, and J. Han, “Opinosis: A graph based approach
to abstractive summarization of highly redundant opinions,” 2010.
[38] N. Okumura and T. Miura, “Automatic labelling of documents based on
ontology,” in 2015 IEEE Pacific Rim Conference on Communications,
Computers and Signal Processing (PACRIM). IEEE, 2015, pp. 34–39.
[39] N. Moratanch and S. Chitrakala, “A survey on abstractive text sum-
marization,” in 2016 International Conference on Circuit, Power and
Computing Technologies (ICCPCT). IEEE, 2016, pp. 1–7.
[40] A. Khan, N. Salim, and Y. J. Kumar, “A framework for multi-document
abstractive summarization based on semantic role labelling,” Applied
Soft Computing, vol. 30, pp. 737–747, 2015.
[41] P. Kouris, G. Alexandridis, and A. Stafylopatis, “Abstractive text sum-
marization based on deep learning and semantic content generalization,”
in Proceedings of the 57th Annual Meeting of the Association for
Computational Linguistics, 2019, pp. 5082–5092.
[42] S. Dutta, V. Chandra, K. Mehra, S. Ghatak, A. K. Das, and S. Ghosh,
“Summarizing microblogs during emergency events: A comparison of
extractive summarization algorithms,” in Emerging Technologies in Data
Mining and Information Security. Springer, 2019, pp. 859–872.
[43] E. Lloret, M. T. Romá-Ferri, and M. Palomar, “Compendium: A text
summarization system for generating abstracts of research papers,”
in International Conference on Application of Natural Language to
Information Systems. Springer, 2011, pp. 3–14.
[44] ——, “Compendium: A text summarization system for generating
abstracts of research papers,” Data & Knowledge Engineering, vol. 88,
pp. 164–175, 2013.
[45] A. P. Patil, S. Dalmia, S. A. A. Ansari, T. Aul, and V. Bhatnagar,
“Automatic text summarizer,” in 2014 International Conference on
Advances in Computing, Communications and Informatics (ICACCI).
IEEE, 2014, pp. 1530–1534.

1509