Automatic Text Summarization Methods: A Comprehensive Review

Divakar Yadav1, Jalpa Desai, Arun Kumar Yadav2

[email protected], [email protected], [email protected]

ORCID: 1https://orcid.org/0000-0001-6051-479X , 2https://orcid.org/0000-0001-9774-7917

Abstract:

One of the most pressing issues that have arisen due to the rapid growth of the Internet is known as information overloading. Simplifying the relevant information
in the form of a summary will assist many people because the material on any topic is plentiful on the Internet. Manually summarising massive amounts of text is
quite challenging for humans. So, it has increased the need for more complex and powerful summarizers. Researchers have been trying to improve approaches for
creating summaries since the 1950s, such that the machine-generated summary matches the human-created summary. This study provides a detailed state-of-the-
art analysis of text summarization concepts such as summarization approaches, techniques used, standard datasets, evaluation metrics and future scopes for research.
The most commonly accepted approaches are extractive and abstractive, studied in detail in this work. Evaluating the summary and increasing the development of
reusable resources and infrastructure aids in comparing and replicating findings, adding competition to improve the outcomes. Different evaluation methods of
generated summaries are also discussed in this study. Finally, at the end of this study, several challenges and research opportunities related to text summarization
research are mentioned that may be useful for potential researchers working in this area.

Keyword: Automatic text summarization, Natural Language Processing, Categorization of text summarization system, abstractive text summarization, extractive
text summarization, Hybrid Text Summarization, Evaluation of text summarization system

1. Introduction:

The task of compressing a piece of text into a shorter version, minimizing the size of the original text while keeping crucial informational aspects and content
meaning, is known as summarization. Fig. 1 shows task of summarization in a simple way. A summary is a reductive transformation of a source text into a summary
text by extraction or generation (Radev et al., 2004). According to another definition, “An automatic summary is a text generated by a software that is coherent and
contains a significant amount of relevant information from the source text. Its compression rate τ is less than a third of the length of the original document (Hovy
& Lin, 1996). The ratio between the length of the summary and the length of the source document is calculated by the compression rate τ as shown below:

||
τ=
||

Where | • | indicates the length of the document in characters, words, or Sentences. τ can be expressed as a percentage. In fact, (C. Y. Lin, 1999) study shows that
the best performances of automatic summarization systems are found with a compression rate of τ = 15 to 30% of the length of the source document.

Fig. 1: generating summary from input document

Understanding the source text and creating a brief and abbreviated version of it are two processes in the human generation of summaries. Fig. 2 shows how human
produces the summaries of an original text document. The summarizer's linguistic and extra-linguistic abilities and knowledge are required for both understanding
the material and producing summaries. Although people can write better summaries (in terms of readability, content, form, and conciseness). Automatic text
summarizing is a useful supplement to manual summation rather than a substitute.

Fig. 2: human’s process for generating summaries

1.1 Requirement of Text Summarization

The adage "too much information kills information" is as relevant today as it has ever been. The fact that the Internet is available in various languages only adds
to the aforementioned document analysis challenges. Automatic text summarization aids in the effective processing of an ever-increasing volume of data that
humans are just unable to handle. Let’s look at some eye-opening facts about the world of data provided by Arne von See (2021) as shown in fig. 3. Some facts
about it are: in the previous two years, 90 percent of the world's data has been created. The majority of businesses only look at 12% of their data. Each year, bad
data costs the United States $3.1 trillion. By 2025, the amount of data created will have surpassed 180 zettabytes. To download all of the material from the internet
now, it would take a human around 181 million years.
 Fig. 3. Volume of data/information created, captured, copied, and consumed worldwide from 2010 to 2025 (Arne von See, 2021)

There are several valid reasons in favour of the automatic summarization of documents. Here are listed just a few (Ab & Sunitha, 2013)

i. Summaries saves reading time.

ii. Summaries help in the selection of documents when conducting research.
iii. Indexing is more successful when automatic summarization is used.
iv. When compared to human summarizers, automatic summary systems are less biased.
v. Because they provide personalized information, personalized summaries are important in question-answering systems.
vi. Commercial abstract services can improve the number of texts they can process by using automatic or semi-automatic summarizing techniques.

Automatic Text Summarization (ATS) is a relatively new learning issue that has gotten a lot of interest. As research advances, we hope to see a breakthrough that
will help with this by giving a timely technique of summarising big texts. We present an overview of text summarising techniques in this work to highlight their
usefulness in dealing with enormous data and to assist researchers in using them to address challenges. The fig. 4 shows the number of researcher papers published
in domain of text summarization in a particular time interval staring from 1958.

Fig. 4: Number of Research articles published in the domain of ATS in different time interval

1.2 Main Contribution of study

This work provides a concise, current and comprehensible view in the field of text summarization. The major contributions of this study are as under:

a. Starting from ground level, make the reader comfortable with the ATS system and why we need an ATS system. Provided examples of ATS systems
presented in the literature for each application and illustrated ATS systems' classifications.
b. Provided a detailed analysis of the three ATS approaches extractive, abstractive, and hybrid. Furthermore, the review table is built on factors like dataset,
approach, performance, advantages, and disadvantages.
c. Provided an overview of the standard datasets and provided complete details about evaluation methods available for the ATS system.
d. Detailed analysis of challenges and future scopes for text summarization.

This article is arranged into six sections. Section 1 discusses the introduction of an automatic text summarization system with its requirements and applications.
The automatic text summarization is divided into many categories discussed in detail in section 2. Next, section 3 focused on Extractive, Abstractive and Hybrid
text summarization. The evaluation methods for summaries generated by the system are discussed in section 4. Later that frequently used datasets for summarization
task is listed in section 5. Lastly, the conclusion is given in section 6.

2. Categorization of ATS
There are different classifications for an automatic text summarization (ATS) system based on its input, output, purpose, length, algorithms, domain, and language.
There are many other factors that can be considered while discussing the classification of summarization. Different researchers have considered different factors.
As per our survey, the detailed categorization of an ATS system is given in fig. 5. A detailed explanation of a particular category are discussed in following sub-
sections as under:

2.1 Based on no. of Input documents

Based upon size of input source documents that are used to generate a summary, summarization can be divided in two types:
• Single Document: Single document text summarization is automatic summarization of information a single document (Garner, 1982).
 • Multiple Document: Multi-document text summarization is an automatic summarization of information from multiple document (Ferreira et al., 2014).

Multi-document summarization is important where we must put different types of opinions together, and each idea is written with multiple perspectives within
a single document. Single document text summarization is easy to implement, but multi-document summarization is a complex task. Redundancy is one of the
biggest problems in summarizing multiple documents. Carbonell & Goldstein (1998) has given MMR (Maximal Marginal Relevance) approach, which helps to
reduce redundancy. Another main problem for multi-document summarization is heterogeneity within a large set of documents. It is very complex to summarize
multiple documents with extractive methods where there are so many conflicts and biases in the real world. Here for multiple documents, abstractive summarization
performs far better. However, multi-document summarization also brings issues like redundancy in output summary while working with a huge number of
documents. Single document text summarization is used in a limited field like reading the given comprehension and giving an appropriate title or summary. In
contrast, multi-document text summarization can be used in the field of news summarization from different sites, customer's product reviews from different vendors,
Q&A systems and many more.

SummCoder (Joshi et al., 2019) is a new methodology for generic extractive single document text summarization. The method creates a summary based on three
criteria they developed: sentence content relevance, sentence novelty, and sentence position relevance. The novelty metric is produced by utilizing the similarity
among sentences represented as embedding in a distributed semantic space, and the sentence content relevance is assessed using a deep auto-encoder network. The
sentence position relevance metric is a custom feature that gives the initial few phrases more weight thanks to a dynamic weight calculation method controlled by
the document length. In the extractive multi-document text summarization field, Sanchez-Gomez et al. (2021) shows that all feasible combinations of the most
prevalent term-weighting schemes and similarity metrics have been implemented, compared, and assessed. Experiments with DUC datasets were conducted, and
the model's performance was evaluated using eight ROUGE indicators and the execution time. The TF-IDF weighting scheme and cosine similarity give the best
result of 87.5% ROUGE score as a combination.

Fig. 5: Detailed Categorization of automatic text summarization system

2.2 Based on Summarization Methods

Based on methods that how can summaries are produced, i.e. Just picking up sentences from the source text or generating new sentences after reading source text
or a combination of both, summarization can be divided into three types:

• Extractive Automatic Text Summarization: Extractive text summarization is the strategy of concatenating on extracting summary from a given corpus
(Rau et al., 1989).
• Abstractive Automatic Text Summarization: Abstractive text summarization involves paraphrasing the given corpus and generating new sentences (Zhang
et al., 2019).
• Hybrid Automatic Text Summarization: It combines both extractive and abstractive methods. It means extracting some sentences and generating a new
one from a given corpus (Binwahlan et al., 2010).

Fig. 6: Extractive text summarizer and Abstractive text summarizer

Think of a highlighter used to point out important sentences in a book. It is an example of extractive text summarization. Now think of the notes we prepare
from a book using our own words. It is an example of abstractive text summarization. Extractive text summarization is like copy-pasting some of the important
sentences from the source text, while abstractive text summarization selects some meaningful sentences and generates new sentences from previously selected
sentences. Refer Fig. 6 for a better understanding of Extractive and Abstractive summarization. Hybrid text summarization combines an approach for producing a
summary efficiently. Both Extractive and Abstractive text summarization falls into Machine Learning and NLP domain. Additionally, abstractive text
summarization covers NLG. The survey of both approaches is shown in the later section of this article. Critical areas where extractive text summarization is applied
are news, medical, book, legal document, abstractive text summarization, customer reviews, blog, tweet summarization, etc.
The plus point of the extractive text summarization model is that the sentences in the summaries must adhere to the syntactic structure's constraints. However, that
model's shortcoming is that the summaries' sentences may not be semantically meaningful. This disadvantage arises because adjacent sentences in the summaries
are not always contiguous in the original text. Because ATS models learn the collocation between words and construct a sequence of keywords based on the
collocation between words after training, they have the advantage of inclusive semantics. The downside of ATS models is that it is challenging to meet the criterion
of syntactic structure with this sequence of keywords. Rare words are another major flaw in traditional ATS models. The number of occurrences of a rare word and
its collocation will define its importance, but humans will use other elements to assess whether a word is essential. As a result, in some instances, some words that
appear infrequently might be deemed unimportant, although a portion of these words is critical for summary construction from a human perspective (Song et al.,
2019b).
2.3 Based on Output Summary Nature:
Based on the output summary’s characteristics the ATS system can be divided into two types:
• Generic: Generic text summarizers fetch important information from one or more documents to provide a concise meaning of given document(s) (Aone et
al., 1997).
 • Query-Based: A query-based summarizer is built to handle multi-documents and gives a solution as per the user’s query (Van Lierde & Chow, 2019). The
score of sentences in each document is based on the frequency counts of words or phrases in query-based text summarization. Sentences containing query
phrases receive higher marks than sentences containing single query words. The sentences with the highest scores and their structural context are extracted
for the output summary (Kiyani & Tas, 2017)..
A query-based sentence extraction algorithm is given as below (Pembe & Güngör, 2007):
i. Rank all the sentences according to their score.
ii. Add the main title of the document to the summary.
iii. Add the first level-1 heading to the summary.
iv. While (summary size limit not exceeded)
v. Add the next highest scored sentence.
vi. Add the structural context of the sentence: (if any and not already included in the summary)
vii. Add the highest-level heading above the extracted text (call this heading h).
viii. Add the heading before h in the same level.
ix. Add the heading after h in the same level.
x. Repeat steps 7, 8 and 9 for the subsequent highest-level headings.
xi. End while

A query is not used in generic summaries. Because they do not comprehensively assess the original document, query-based summaries are biased. They are not
ideal for content overview because they solely deal with user queries. Generic summaries are necessary to specify the document's category and to describe the
document's essential points. The key subjects of the documents are considered in the best general summary, which strives to minimize redundancy as much as
possible (Kiyani & Tas, 2017).

2.4 Based on Summary Language

Based on the language of input and output of the ATS system, it can be divided into the following 3 categories:
• Monolingual: In a Monolingual text summarizer, the language of the input document and output summary is the same (Kutlu et al., 2010).
• Multilingual: In a Multilingual text summarizer, input is written in many languages (Hindi, English, and Gujarati), and output summary is generated
likewise in these languages (Hovy & Lin, 1996).
• Cross- Lingual: In a Cross-lingual text summarizer, the input document is in one language (say English), and the output summary is in another language
(say Hindi) (Linhares Pontes et al., 2020).

Most of the research papers studied in this article are based on monolingual text summarization. Compared to monolingual, multilingual and cross-lingual is
challenging to implement. It takes more effort to train a machine on more than one language structure. SUMMARIST (Hovy & Lin, 1996) is a multilingual text
summarization system based on an extraction strategy that generates summaries from English, Indonesian, Spanish, German, Japanese, Korean, and French sources.
Cross-Language Text Summarization (CLTS) (Linhares Pontes et al., 2020) generates a summary in a target language from source language materials. It entails a
combination of text summarising and machine translation methods. Unfortunately, this combination leads to mistakes, which lowers the quality of summaries. CLTS
systems may extract relevant information from both source and destination languages through joint analysis, which improves the development of extractive cross-
lingual summaries. Recent methods for CLTS have offered compressive and abstractive approaches; however, these methods rely on frameworks or tools that are
only available in a few languages, restricting their applicability to other languages.

2.4 Based on Summarization Algorithms

Based on the actual algorithm that is used to generate the summarises, the ATS system is divided into two types as given below:

• Supervised: The supervised summarizer needs to train the sample data by labelling the input text document with the help of human efforts (Neto et al.,
2002).
• Unsupervised: In the Unsupervised summarizer training phase is not needed (Alami et al., 2019).

In order to select important content from documents in a supervised system, training data is required. Training data is a large volume of labelled or annotated
data is required for learning techniques. These systems are approached as a two-class classification issue at the sentence level, with positive samples being sentences
that belong to the summary and negative samples being sentences that do not belong to the summary. On the other hand, unsupervised systems do not require any
training data. They create the summary by just looking at the documents they want to look at for summarization. As a result, they can be used with any newly
observed data without needing extra adjustments. These systems use heuristic methods to extract relevant sentences and construct a summary. Clustering is used in
unsupervised systems (Gambhir & Gupta, 2017).
A single document supervised machine learning-based approach for the Hindi language is given by Nikita (2016). The sentences are divided into four categories:
most influential, important, less important, and insignificant. The summarizer is then trained using the SVM supervised machine learning algorithm to extract
important sentences based on the feature vector. Sentences are then included in the final summary based on the required compression ratio. The experiment was
carried out on news stories from various categories such as Bollywood, politics, and sports, and the results showed 72 percent accuracy at a compression ratio of 50
percent and 60 percent at a compression ratio of 25 percent. Recently, an unsupervised neural network approach has been studied by Meknassi et al. (2021) for
Arabic language text summarization. A new approach using documents clustering, topic modelling, and unsupervised neural networks have been proposed to build
an efficient document representation model to overcome problems raised with Arabic text documents. The proposed approach is evaluated on Essex Arabic
Summaries Corpus and compared against other Arabic text summarization approaches using ROUGE measure.

2.5 Based on Summary Content

Based on the type of the content of the output summaries, the system is categorised into two parts as:
• Inductive: Indicative summary contains only generic idea about the source document (Bhat et al., 2018).
• Informative: Informative summary contains all the main topics about the original document (Bhat et al., 2018).
• Evaluative or Critical: It capture the summary from author's point of view on a given topic (Jezek & Steinberger, 2008).

Inductive summarization is used to indicate what the document is all about, and it aims to give an idea to a user whether to read this original document or not.
The length of this summary is approximately 5% of the original content. The informative summarization system summarises the primary text concisely. The helpful
summary is around 20% of the whole text length (Kiyani & Tas, 2017). A typical example of evaluative summaries are reviews, but they are pretty out of the scope
of nowadays summarizers. It should be emphasized that the three groupings indicated above are not mutually exclusive and are common summaries that have both
an informative and an indicative role. Informative summarizers are frequently used as a subset of indicative summarizers (Jezek & Steinberger, 2008).

2.6 Based on Summary types

Based on the length of the generated summaries, the ATS system is divided into four types as given below:
• Headline: A headline generated from a source document is usually shorter than a sentence (Barzilay & Mckeown, 2005).
• Sentence level: The sentence-level summarizer produces a single sentence from the original text (Y. H. Hu et al., 2017).
• Highlight: Highlights are produced in a compressed form of the original text written in bullet points (Tomek, 1998).
• Full summary: Full summary is generated as per the user’s compression rate or user’s requirements (Koupaee & Wang, 2018)

Headlines, highlights, and sentence level type of summary are generally used in news database or opinion mining or for social media dataset whereas a full
summary is commonly used for all the domains.

2.7 Based on Summarization Domain:

Based on the domain of the input and output of the ATS system, it is divided into following 3 categories:
 • Genre Specific: It accepts only special type of input text format (Hovy & Lin, 1996) .
• Domain dependent: Domain dependent summarization is specific to one domain (Farzindar & Lapalme, 2004).
• Domain independent: Domain independent summarization system is independent of source documents’ domain.

In genre-specific summarization, there is a restriction on the text template. Newspaper articles, scientific papers, stories, instructions, and other types of templates
are available. The summary is generated by the system using the structure of these templates. On the other hand, independent systems have no predefined limitations
and can take a variety of text kinds. Furthermore, some techniques only summarise texts whose subject can be characterized in the system's domain; these systems
are domain-dependent. These systems impose some restrictions on the topic matter of documents. Such systems know everything there is to know about a specific
subject and use that knowledge to summarise. Generally, graph-based techniques are adopted for domain-dependent summarisation as they have sound potential.
The authors of (Moradi et al., 2020) have given an efficient solution to deal with the challenges in graph-based methods. To capture the linguistic, semantic, and
contextual relationships between the sentences, they trained the model by continuous word representation model. i.e., Word2vec's Skiagrams and Continuous Bag
of Words (CBOW) models(Mikolov et al., 2013) and Global Vectors for Word Representation (GloVe) (Mutlu et al., 2020)(Hanson Er, 1971) on a large corpus of
biomedical text. To solve the challenge of ranking the most important nodes in a graph, they adopted undirected and weighted graph ranking techniques like the
PageRank algorithm (Brin & Page, 2012). Newspaper stories and scientific text have distinct qualities than legal text. In a comprehensive document of the news
genre, for example, there is little or no structure. The presence of the same term at different levels of the hierarchy will have distinct effects. The relevance of the
words in a judgement is determined by the source of the ruling (whether it is from a District Court, State Court, Supreme Court, or Federal Court). We can generally
ignore references/citations when summarizing content; however, this may not be possible in legal writings (Kanapala et al., 2019).

2.8 Based on Processing Level

Based on the processing level of the input document, the system is divided into two types:
• Surface-level approaches: In this scenario, data is represented by shallow feature ideas and their combinations.
Statistically salient terms, positionally salient terms, cue phrases, domain-specific or a user's query terms are examples of shallow features. The outcomes are in the
form of extracts (Ježek et al., 2007).
• Deeper-level approaches: Extracts or abstracts may be produced through deeper-level techniques. In the latter scenario, synthesis is used to generate natural
language. It requires some semantic analysis. For example, entity techniques can construct a representation of text entities (text units) and their relationships to
identify salient areas. Entity relationships include thesaural, syntactic, and semantic relationships, among others. They can also use discourse methodologies to
represent the text structure, such as hypertext mark-up or rhetorical structure (Ježek et al., 2007).

3 Detailed about ETS, ABS and HTS

Among all the Automatic text summarization system classification, the most commonly accepted or used categories are Extractive, Abstractive and hybrid
summarization. Thus, this article focused is on these two approaches mainly. Now lets us see a detailed survey on these approaches:

3.1 Extractive text summarization

Since starting with the era when first-time automatic text summarization came into the picture (Luhn, 1958), the text processing task is performed mainly by
using features based on IR (Information Retrieval) measures, i.e., term frequency (TF), inverse term frequency (TF-IDF). Table-1 shows a detailed survey on
extractive text summarization with research papers, the dataset used, the system's accuracy, and its pros and cons.Earlier, the efficiency of the summary was
prepared by the proportion of no. of judged-important points to total no. of words in the summary (Garner, 1982). The immediate summarization result and
relationship to detailed comprehension and recall results were analyzed in that study. The lack of linguistic knowledge is a weak point for extracting helpful
information from a large amount of the data. To overcome these two limitations: (i) One mechanism that deals with unknown words and gaps in linguistic
information. (ii) To extract linguistic information from text automatically, SCISOR (System for Conceptual Information Summarization, Organization and
Retrieval) was developed by Rau et al. (1989). Experiments performed on summarization until 1990 were focused on just extracting (reproduction) the summaries
from original text rather than abstracting (newly generated). SUMMRIST system (Hovy & Lin, 1996) was developed with the help of NLP techniques, in which
one can create a multi-lingual summarizer by modifying some part of the structure.
The challenges with traditional frequency-based, knowledge-based and discourse-based summarization lead to encountering these challenges with robust NLP
techniques like corpus-based statistical NLP (Aone et al., 1997). The summarization system named DimSum consists of a summarization server and summarization
client. The features produced from these powerful NLP algorithms were also used to give the user numerous summary views in an innovative way. Evaluating the
summaries of humans and systems by four parameters; optimistic evaluation, pessimistic evaluation, intersection evaluation, union evaluation(Salton et al., 1997)
and proven that the summaries generated by the two humans are dissimilar for the same article while automatic methods are favourable here. A robust summarization
was practically implemented on online news 'New York Times' by Tomek (1998), which gives summaries very quickly with including significantly less portion of
original lengthy text. The study of effects of headings on text summarization proven (Lorch et al., 2001) that readers depend heavily on organizational signals to
construct a topic structure. The machine learning approach (Neto et al., 2002) considers automatic text summarization as a two-class classification problem, where
a sentence is considered 'correct' if it appears in extractive reference summary or otherwise as 'incorrect'. Here they used two famous ML classification approaches,
Naïve Bayes and C4.5. Lexicon is a salient part of ant textual data. Focusing on an algorithm (Silber & McCoy, 2002) that efficiently makes lexical chains in linear
time is a feasible intermediate representation of text summarization.
As a prior study shows, supervised methods were where human-made summaries helped us find parameters or features of summarization algorithms. Despite
that, unsupervised methods (Nomoto & Matsumoto, 2003) with diversity functionality define relevant features without any help from human-made summaries.
(Yeh et al., 2005b) proposed a trainable summarizer that generates summaries based on numerous factors such as location, positive keyword, negative keyword,
centrality, and resemblance to the title. It uses a genetic algorithm (GA) to train the score function to find a good combination of feature weights. After that, it
employs latent semantic analysis (LSA) to derive a document's or corpus' semantic matrix and semantic sentence representation to build a semantic text relationship
map. Combining three approaches: a diversity-based method, fuzzy logic, and swarm-based methods (Binwahlan et al., 2010), can generate good summaries.
Where diversity-based methods use to figure out similar sentences and get the most diverse sentence and concentrate on reducing redundancy, while swarm-based
methods are used to distinguish the most important and less important sentences then use fuzzy logic to tolerate redundancy, approximate values and uncertainty,
and this combination concentrates on the scoring techniques of sentences. While comparing two-approach Swarm-fuzzy based methods performs well than
diversity-based methods here.
The construction of methods for measuring the efficiency of SAS (Systems of automatic summarization) functioning is an important area in the theory and
practice of automatic summarization. Based on a model vocabulary supplied by subjects, four techniques (ESSENCE (ESS), Subject Search Summarizer (SSS),
COPERNIC (COP), Open Text Summarizer (OTS)) of automatic text summarization are evaluated by Yatsko & Vishnyakov (2007). The distribution of vocabulary
terms in the source text is compared to the distribution of vocabulary terms in summaries of various lengths generated by the systems. (Ye et al. (2007b) contend
that the quality of a summary can be judged by how many concepts from the source documents can be retained after summarization. As a result, summary generation
can be viewed as an optimization task involving selecting a set of sentences with the least answer loss. The proposed document concept lattice (DCL) is a unique
document model that indexes sentences based on their coverage of overlapping concepts. The authors of (Ko & Seo, 2008) suggested method merges two
consecutive sentences into a bi-gram pseudo sentence, allowing statistical sentence-extraction tools to use contextual information. The statistical sentence-extraction
approaches first choose salient bi-gram pseudo sentences, and then each selected bi-gram pseudo sentence is split into two single sentences. The second sentence-
extraction operation for the separated single sentences is completed to create a final text summary.
CN-Summ(Complex Networks-based Summarization) was proposed by (Antiqueira et al., 2009). Nodes relate to sentences in the graph or network representing
one piece of text, while edges connect sentences that share common significant nouns. CN-Summ consists of 4 steps: 1) prepossessing (lemmatization). 2) resulting
text is mapped to a network representation according to adjutancy and weight metrics of order n*n (n is no. of sentences/nodes) .3) compute different network
measurements 4) the first m sentences are selected as summary sentences depending upon compression rate. Alguliev & Aliguliyev (2009) gave a new approach
for unsupervised text summarization. That approach is focused on sentence clustering, where clustering is the technique of detecting interesting distributions and
patterns within multidimensional data by establishing natural groupings or clusters based on some similarity metric. Here the researchers have proposed a new
method to measure similarity named Normalized Google Distance (NGD) and to optimize criterion functions discrete differential evolution algorithm called as
MDDE (Modified Discrete Differential Evolution) Algorithm is proposed.
Swarm Intelligence (SI) is the collective intelligence resulting from the collective behaviours of (unsophisticated) individuals interacting locally and with their
environment, causing coherent functional global patterns to emerge. The primary computational parts of swarm intelligence are Particle Swarm Optimization
(PSO), which is inspired by the social behaviour of bird flocking or fish schooling, and Ant Colony Optimization (ACO), which is inspired by the behaviour of
ants. Binwahlan et al. (2009a) suggested a model based on PSO whose primary goal is to score sentences while focusing on dealing equally with text elements
depending on their value. Combining three approaches: diversity-based method, fuzzy logic, and swarm-based methods (Binwahlan et al., 2010) can generate good
summaries. Where diversity-based methods use to figure out similar sentences and get the most diverse sentence and concentrate on reducing redundancy, while
swarm-based methods are used to distinguish the most important and less important sentences then use fuzzy logic to tolerate redundancy, imprecise values and
uncertainty, and this combination concentrates on the scoring techniques of sentences. While comparing two-approach Swarm-fuzzy based methods performs well
than diversity-based methods here.
In (Mashechkin et al., 2011), the researchers had used LSA(Latent Semantic Analysis) for text summarization. The original text is reproduced as a matrix of
terms and sentences. Text sentences are represented as vectors in the term space, and a matrix column represents each sentence. The resultant matrix is then
subjected to latent semantic analysis to construct a representation of text sentences in the topic space, which is performed by applying one of the matrix factorizations
(singular value decomposition (SVD)) to the text matrix. (Alguliev et al., 2011b) consider text summarization problem as integer linear programming problem
while assuming that summarization is a task of finding a subset of sentences from the original text to represent important detail of the original text. That study
focused on three characteristics (relevance, redundancy, and length) and tried to optimize that by particle swarm optimization algorithm (PSO) and branch and
bound optimization algorithm. In extractive text summarization, sentence scoring is the most commonly used technique. The study (Ferreira et al., 2013) evaluated
15 algorithms available for sentence scoring based on quantitative and qualitative assessments. In conjunction with a graph-based ranking summarizer, Wikipedia
is given by (Sankarasubramaniam et al., 2014a). It has given a unique concept by introducing incremental summarization property, where single and multi-document
both can provide additional content in real-time. So, the users can first see the initial summary, and if willing to see other content, they can make a request.
Using a deep auto-encoder (AE) to calculate a feature space from the term-frequency (tf) input, (Yousefi-Azar & Hamey, 2017b) offer approaches for extractive
query-oriented single-document summarization. Both local and global vocabularies are considered in experiments. The study shows the effect of adding slight
random noise to local TF as the AE's input representation and propose the Ensemble Noisy Auto-Encoder as a collection of such noisy AEs (ENAE). Even though
there is a lot of study on domain-based summarising in English and other languages, there is not much in Arabic due to a lack of knowledge bases. A hybrid, single-
document text summarization approach is proposed in (Al-Radaideh & Bataineh, 2018a) paper (ASDKGA). The method uses domain expertise, statistical traits,
and genetic algorithms to extract essential points from Arabic political documents. For domain or genre-specific summarization (such as for medical reports or
specific news articles), feature engineering-based models have shown to be far more successful, as classifiers can be taught to recognize particular forms of
information. For general text summary, these algorithms produce poor results. To overcome the issue, an entirely data-driven approach for automatic text
summarization is given by (Sinha et al., 2018)
The most challenging difficulties are covering a wide range of topics and providing diversity in summary. New research based on clustering, optimization, and
evolutionary algorithms has yielded promising results for text summarization. A two‐stage sentences selection model based on clustering and optimization
techniques, called COSUM, was proposed by Alguliyev et al. (2019b). The sentence set is clustered using the k - means algorithm in the first stage to discover all
subjects in a text. An optimization approach is proposed in the second step for selecting significant sentences from clusters. The most crucial reason for the lack of
domain shift approaches could be understanding different domain definitions in text summarization. For the text summarization task, (Wang et al., 2019) extended
the traditional definition of the domain from categories to data sources. Then used, a multi-domain summary dataset to see how the distance between different
domains affects neural summarization model performance. Traditional applications have a major flaw: they use high-dimensional, sparse data, making it impossible
to gather relevant information. Word embedding is a neural network technique that produces a considerably smaller word representation than the classic Bag-of-
Words (BOW) method. (Alami et al., 2019) has created a text summarization system based on word embeddings, and it showed that the Word2Vec representation
outperforms the classic BOW representation. Another summarization approach using word embeddings was given by (Mohd et al., 2020). This study also used
Word2Vec as a distributional semantic model that captures the semantics.
Current state-of-art systems produce generic summaries that are unrelated to the preferences and expectations of their users. CTRLsum (He, Kryscinski, et al.,
2020), a unique framework for controlled summarizing, is presented to address that limitation. This system permits users to interact with the summary system via
textual input in a collection of key phrases or descriptive prompts to influence several features of generated summaries. The majority of recent neural network
summarization algorithms are either selection-based extraction or generation-based abstraction. (Xu & Durrett, 2020) introduced a neural model based on joint
extraction and syntactic compression for single-document summarization. The model selects phrases from the document, identifies plausible compressions based
on constituent parses, and rates those compressions using a neural model to construct the final summary. Four algorithms were proposed by (El-Kassas et al., 2020).
The first algorithm uses the input document to create a new text graph model representation. The second and third algorithms look for sentences to include in the
candidate summary in the built text graph. The fourth algorithm selects the most important sentences when the resulting candidate summary exceeds a user-
specified limit. Automatic text summarization is an arduous effort for under-resourced languages like Hindi, and it is still an unsolved topic. Another problem with
such languages is the lack of corpus and insufficient processing tools. For Hindi novels and stories, (Rani & Lobiyal, 2021) developed an extractive lexical
knowledge-rich topic modelling text summarising approach in this study. The standard words-based similarity measure grants weight to most graph-based text
summarising techniques. Belwal et al. (2021) offered a new graph-based summarization technique that considers the similarity between individual words and the
sentences and the entire input text.
Table 1: Research survey on Extractive text summarization method

Citation article Model/methods/te Dataset Used Performance Advantages/Pros Disadvantages/Con

chniques applied s
(Garner, 1982) Efficient Text Extractive text Dutch elm Disease Efficiency: - -
Summarization: summarization (167 words- a proportion of
Costs and single word) number of judged-
Benefits important ideas to
total number of
words in summary
Proportion=0.2 to
0.12
(Rau et al., 1989) Information Extractive text Dataset named as: - SCISOR is robust Size of lexicon.
extraction and summarization Group Offers to and reliable (10,000 words) and
text Sweeten Warnaco extraction of system is limited.
summarization Bid information
using linguistic
knowledge
acquisition
(Bloom et al., Automatic Extractive text 175-page article - robust and In the absence of
1994) Analysis, Theme summarization Entitle "United generally deep linguistic, it is
Generation, and States of America. applicable to a wide not possible to build
Summarization variety of Intellectually
of Machine- texts in many satisfactory text
Readable Texts different summaries.
environments,
(Reimer et al., Formal Model of Extractive text text - attempt was Currently the
1997) Text summarization made to properly summarization
Summarization integrate the text process considers
Based on summarization only activity and
Condensation process to the connectivity
Operators of a formal reasoning patterns m the text
Terminological mechanisms of the knowledge base
Logic underlying
knowledge
representation
language
(Aone et al., A Scalable Extractive text text - The DimSum The results are not
1997) Summarization summarization summarization evaluated with
System Using system advances human generate
Robust NLP summarization summaries.
technology by
applying
corpus-based
statistical NLP
techniques, robust
information
extraction, and
readily available
on-line
resources
(Salton et al., Automatic text Multi document encyclopaedia 4 factors are Original text is Human extracts are
1997) structuring and extractive text article 78 compared with broken into dissimilar for a
summarization summation (Abortion) respect to human constituent pieces same article
extract: Optimistic and combing these
evaluation, pieces according to
Pessimistic their functionality.
evaluation, This text structure
Intersection, and used for producing
Union. comprehensive
summaries by
automatic
paragraph
extraction
-domain
independent
(Lorch et al., Effects of Extractive text The experimental - More topics were Headings interacted
2001) Headings on Text summarization text was a revised included in the in recall, but not
Summarization version of a text on summaries of in the
energy problems participants who summarization task
and read the text with
solutions headings than in the
summaries of
participants who
read the
text without
headings
(Neto et al., 2002) Automatic Text Extractive TIPSTER Comparison Works well with C4.5 gives poor
Summarization summarization on document base. between Naïve Bayes results as compared
Using a Machine multi-document automatically although the dataset to naïve Bayes.
Learning by machine produced and consist of 33,658
Approach learning approach manually produced documents within
summaries:
Compression rate:
10%
Precision/Recall:
30.79 ± 3.96
(Silber & McCoy, Efficiently Multi document Randomly selected Time complexity of Feasible and robust Results are poor
2002) computed lexical extractive text documents which algorithm: linear algorithm with where proper nouns
chains as an summarization are of length time (O(n) where n linear time as proper nouns and
intermediate ranging from 2,247 is nouns in complexity domain specific are
representation for to 26,320 words document) not found in
automatic text each. WordNet.
summarization
(Nomoto & The diversity- Unsupervised text 5080 Information centric -Does not rely on Diversity based
Matsumoto, based approach to summarization news articles in evaluation human-made summarizer perform
2003) open-domain text and open domain Japanese summaries but poorly when exhibit
summarization measuring the high agreement.
performance in
terms of IR.
-Diversity based
summarizer is
superior to tf-idf
based summaries.
(Yeh et al., Text modified corpus- 100 documents in 5 Compression rate: -LSA+T.R.M. The performance of
2005a) summarization based approach sets of domains of 30% performs better both approaches
using a trainable (MCBA) and politics were MCBA: than keyword- vary as compression
summarizer and LSA-based collected from New F-measure: 49% based text rates are changes.
latent semantic T.R.M. approach Taiwan Weekly MCBA+GA: summarization. For good result
analysis F-measure: 52% -The results of chose appropriate
LSA+T.R.M MCBA+GA show compression rate.
F-measure: 44% that using GA in
(for single- training is a good
document) technique to deal
F-measure: 40% with the
(for corpus) circumstance where
we are undecided
about an
appropriate
combination of
feature weights.
(Ye et al., 2007a) Document multi-document DUC-2005 -ROUGE-2 Recall Proposed a -
concept lattice for extractive DUC-2006 = 7.17% & document concept
text summarization ROUGE-SU4 lattice (DCL)
understanding Recall = 13.16% in model and the
and DUC-2005 corresponding
summarization -ROUGE-2 Recall algorithm for
= 8.99% and summarization
ROUGE-SU4
 Recall = 14.75% in
DUC-2006
(Ko & Seo, 2008) An effective Single/multi- 841 news articles F1 score: 47.9 with To tackle the The performance
sentence- document from KOrea 10% summary feature sparseness changes according
extraction extractive Research and F1 score: 50.4 with problem, it used to no. of query
technique using summarization Development 30% contextual words in multi-
contextual Information Centre information (bi- document
information and (KORDIC) dataset gram pseudo
statistical phrase) and the
approaches for combination
text method of
summarization statistical
methodologies to
increase the
performance.
-Independent on
language
(Antiqueira et al., A complex graph-based 100 newspaper Average f-measure Language & There are 15
2009) network approach extractive articles in Brazilian for all 15 version: domain versions of CN-
to text summarization Portuguese 42% independent graph- Summ. Not an
summarization based approach approach that
using complex combine all.
networks (CN-
Summ)
(alguliev & Evolutionary Unsupervised DUC2001 & ROUGE-1: Proposed new -the results is shown
aliguliyev, 2009) Algorithm for extractive text DUC2002 0.45952 method for using 3 criterion
Extractive Text summarization ROUGE-2: measuring functions and
Summarization 0.19338 similarity between among them
ROUGE-L: sentences criterion function 1
0.21763 - proposed a new perform poorly.
algorithm method
to optimize
objective functions
(Binwahlan et al., Swarm Based Swarm DUC2002 The proposed Employed PSO for The experiment
2009b) Text intelligence based model creates sentence scoring results are
Summarization extractive text summaries which based on their compared with
summarization are 43% similar to importance. manually human
the manually made summaries
generated which are only 49%
summaries similar.
(Binwahlan et al., Fuzzy swarm Extractive text First 100 ROUGE-1: Diversity-based and Well accepted
2010) diversity hybrid summarization documents from Fuzzy swarm fuzzy methods method as per
model for text (Diversity- DUC 2002. Average F: 0.45524 reduce ambiguity. results but not made
summarization swarm-fuzzy) Swarm-based for multi-document
methods summarization.
differentiate
important sentences
and less important
sentences
(Mashechkin et Automatic text LSA based DUC2001 & ROUGE-2: A new generic The optimal
al., 2011) summarization extractive text DUC2002 0.19260 summarization numbers of topics
using latent summarization. ROUGE-L: method is proposed for each
semantic analysis 0.37229 that uses summarization
nonnegative matrix problem is
factorization to empirically
estimate sentence obtained.
relevance
-In comparison to
the singular value
decomposition
position, the
proposed
estimation of
semantic
characteristics
(topics) weight
better preserves the
internal structure of
the text.
(Alguliev et al., MCMR: PSO, DUC2005 & ROUGE-2 on Good for both -Results are directly
2011a) Maximum unsupervised DUC2007 DUC-2005: single document depending upon
coverage and based multi- B&B: 0.0790 and multi- optimization
minimum document PSO: 0.0754 document text algorithm
redundant text extractive text ROUGE-2 on summarization -parameter value of
summarization summarization DUC-2007: objective function
model B&B: 0.1221 influences
PSO: 0.1165 performance of
summarization
(Sankarasubrama Text Wikipedia-based DUC 2002 ROUGE-1: 0.46 To extract salient Among results of
niam et al., summarization graph-based ROUGE-2: 0.23 topics from text evaluation
2014b) using Wikipedia multi-document Precision: 0.57 document it Sentences chosen
summarization Recall:0.50 proposed algorithm uniformly at
algorithm F-measure: 0.51 using Wikipedia random performs
with graph-based poor.
summarization
-concentrated on
personalization and
query-focusing of
summaries
(Gupta & Kaur, A Novel Hybrid Extractive text Standard Unicode- Overall score Question- Language
2016) Text summarization based Punjabi ROUGE-2: answering dependent
Summarization (concept, Corpus and Punjabi 0.85 summary also tried
System for statistical, news corpus AZIT F-score with 50% here where
Punjabi Text location, numeric with each having compression rate: proposed method
data, and 150 documents. 0.86 gives 80-90%
linguistic-based accuracy.
features)
(Yousefi-Azar & Text query-oriented Summarization and Improvement of Auto-Encoders -Better to use semi-
Hamey, 2017a) summarization single-document Keyword ROUGE-2 recall on produces a rich supervised
using extractive text Extraction from average 11.2%. concept vector for technique because
unsupervised summarization Emails (SKE) an entire sentence shortage of
deep learning using deep BC3 from British from a bag-of- manually annotated
learning Columbia words input. data
University -the computational
cost of training and
the necessity to
appropriately tune
the training hyper
parameters
(Sinha et al., n.d.) Extractive Text Extractive text DUC 2002 ROUGE-1: 55.1 A data-driven assumed that
Summarization summarization ROUGE-2: 22.6 technique that summary length to
using Neural using Neural leverages a basic be generated
Networks Networks feedforward neural should be less than
network achieves specified length.
good performance
while compared
with state-of-the-art
systems while
being both
implementationally
and
computationally
light.
(Al-Radaideh & A Hybrid Extractive text KALIMAT corpus Average F- Domain -Language
Bataineh, 2018b) Approach for summarization and Essex Arabic measure: 0.605 knowledge dependent
Arabic Text (domain Summaries Corpus with compression enhances results of -training done with
Summarization knowledge, (EASC) ratio of 40%. the proposed only 500 keywords
Using Domain genetic method which lies into 4
Knowledge and algorithm) categories.
Genetic
Algorithms
(Alguliyev et al., COSUM: Text Extractive text DUC2001 & DUC2001: The model Sometimes
2019a) summarization summarization DUC2002 ROUGE-1: 0.4727 additionally limits performance is
based on using clustering ROUGE-2: 0.2012 the length of conceded while
clustering and and optimization sentences selected compared to other
optimization DUC2002: in the candidate methods
ROUGE-1: 0.4908 summary to ensure
ROUGE-2: 0.2309 readability.
(Wang et al., Exploring Extractive multi- MULTI-SUM - Proposed a multi- Excellent results in
2019) Domain Shift in domain text dataset domain dataset cross-dataset
Extractive Text summarization CNN/DailyMail MULTI-SUM (CNN/DailyMail)
Summarization -Given four but results declined
learning schemes as in in-domain and
a preliminary out-of-domain
investigation into settings.
the properties of
various learning
strategies when
dealing with multi-
domain
summarising jobs.
(Xu & Durrett, Neural Extractive Extractive text New York Times CNN: A sentence Best performance
2020) Text summarization corpus ROUGE-1: 32.7 extraction model is on CNN but
Summarization with deep neural CNN/DailyMail ROUGE-2: 12.2 combined with a performance
with Syntactic network ROUGE-L: 29.0 compression somewhere declined
Compression classifier to for another dataset
determine whether while compared to
or not syntax- baseline methods.
derived
compression
choices should be
deleted for each
sentence.
(Mohd et al., Text document Extractive text DUC2007 F-score for 25% summarization Proposed
2020) summarization summarization summary length: technique based on distributed semantic
using word using ROUGE-1: 33% the distributional model is
embedding distributional ROUGE-2: 7% hypothesis to computationally
semantics ROUGE-L: 20% capture the expensive and time
(Word2vec) ROUGE-SU4: 13% semantics of consuming.
better than the Results are given on
baselines. different % of
summary length,
where sometimes
recall values are low
than baseline
systems.
 (He, Kryściński, CTRLsum: Extractive text CNN/DailyMail CNN/DailyMail: proposed a generic During training, the
et al., 2020) Towards Generic summarization arXiv scientific ROUGE-1: 48.75 framework to model is
Controllable Text (keyword-based papers ROUGE-2: 25.98 perform multi- conditioned on
Summarization model) BIGPATENT ROUGE-L: 45.42 aspect controllable keywords to predict
patent articles arXiv: summarization. summaries but one
ROUGE-1: 47.58 keyword can be fall
ROUGE-2: 18.33 into more than one
ROUGE-L: 42.79 aspect.
Success rate: 97.6
Factual
Correctness: 99.0
(El-Kassas et al., EdgeSumm: Extractive single- DUC2001 & Improvement of -EdgeSumm -Results are very
2020) Graph-based document text DUC2002 1.2% and 4.7% in integrates a number promising but not
framework for summarization ROUGE-1 and of extractive ATS extended for multi-
automatic text using Graph- ROUGE-L methods (graph- document
summarization based approach respectively for based, statistical- summarization
DUC2002 dataset. based, semantic- -not domain specific
based, and system
centrality-based -Language
methods) to take dependent
use of their
strengths while
minimising their
weaknesses.
-The use of hyphens
removal and
synonyms
substitution tasks
on the performance
of the generated
summaries is
investigated.
(Rani & Lobiyal, An extractive text Extractive text 114 Hindi novels Perform preferable -Since there was no -Semantic features
2021) summarization summarization including short than the corpus of Hindi are not taken into
approach using based on LDA- stories from baseline algorithms novels and stories, consideration
tagged-LDA model ‘Munshi for 10% - 30% it built one corpus. -Results are
based topic Premchand’s compression ratios -four distinct somewhere
modelling stories’ blog and given sentence weighting improved and
evaluation metrics scheme-based somewhere declined
variants are derived while compared to
by manipulating the baseline systems for
proposed system different-different
compression rates
(Belwal et al., A new graph- Extractive Opinosis dataset CNN/DailyMail: Added an extra -
2021) based extractive summarization CNN/ DailyMail ROUGE-1: 0.428 parameter that
text using Graph- ROUGE-2: 0.201 calculates the
summarization based and Topic- ROUGE-L: 0.392 similarity of the
using keywords based techniques nodes to the entire
or topic Opinosis dataset: content.
modelling ROUGE-1: 0.271 Used topic
ROUGE-2: 0.084 modelling to
ROUGE-L: 0.161 address the issue of
redundancy
associated with
existing
summarising
methods
(Alami, Mallahi, Hybrid method Extractive text The dataset Dataset-1: Remove Language
et al., 2021) for text summarization consists of 153 F1-measure: 59.47 redundancy with dependent
summarization (statistical and Arabic articles with 40% using MMR
based on semantic taken from two compression rate (Maximal Marginal
statistical and methods) Arabic Dataset-2: Relevance)
semantic newspapers and F1-measure:60.79 Improve sentence
treatment the Arabic version 40% compression score by other
of Wikipedia. The rate statistical features
3.2 Abstractive text summarization
The competitions DUC-22003 and DUC-2004 had standardized the task of Abstractive text summarization, in which news articles from various fields with
multiple reference summaries per article generated by humans are used as datasets. The TOPIARY system (Zajic et al., 2004) stood the best performing technique.
Some noticeable work was submitted by Banko et al. (2000) with phrase-table based machine translation techniques and (Woodsend et al., 2010) with quasi-
synchronous grammar techniques. Table-2 shows a detailed survey on Abstractive text summarization with a particular research paper with the dataset used, the
system's accuracy, and its pros and cons.After that, deep learning was introduced as a viable alternative to many NLP problems. Text is a sequence of words where
sequence-to-sequence models can entertain input and output sequences. With the apparent similarities Machine translation (MT) problem may be mapped to text
summarization despite that abstractive summarization is very different from it. MT is lossless while summarization is lossy in the manner, and MT is a one-to-one
word-level mapping between source and target, but that mapping is less in summarization.
In (Rush et al., 2015), the researchers had used convolution models to encode the input and context-sensitive feed-forward network with attentional mechanism
and showed better results for Gigaword and DUC datasets. (Chen, 2015) have produced a sizeable Chinese dataset for short text summarization (LCSTS), which
has given good results on their dataset while using RNN architecture at both encoder and decoder sides. Beyond RNN architecture at both encoder and decoder
sides, (Nallapati et al., 2016) captured keywords, modelled unseen or rare words, and captured the document's hierarchy using a hierarchical attention mechanism.
The authors have also tried to analyse the quality of the output summary. In that case, somewhere models perform well and somewhere poor compared to others.
Human's summaries are more abstractive naturally because they use some inherent structures while writing summaries. The deterministic transformation in a
discriminative model(RNN) used by Nallapati et al. (2016) limits the representation of latent structure information. After that, Miao & Blunsom (2016) gave a
generative model to capture the latent structure, but they did not consider recurrent dependencies in their generative model. The authors of (Li et al., 2017) tried
to find some common structures such as "what", "what happened", "who actioned what" from the source and proposed a deep recurrent generative model for
modelling latent structure.
AMR (Abstract Meaning Representation) was firstly introduced by Banarescu et al. (2013). AMR targets fetching the meaning of the text by giving a special-
meaning representation to the source text. AMR attempts to capture "who is doing what to whom". The work of Liu et al. (2015).' s includes AMR, but they did
not use it at the abstraction level, so their work is limited to extractive summarization only. Also, the approach aims to generate a summary from a story. Producing
a single graph assumes that all the important sentences can be extracted from a single subgraph. Difficulties arise when information is spread out all over. So, Doha
et al. (2017a) worked on multiple summary graphs and explored problems with existing evaluation methods and datasets while doing abstractive summarization.
Combining the advantages of extractive and abstractive and curing the disadvantages (Song et al., 2019b) had implemented a model named ATSDL (ATS using
DL). This model uses a phrase extraction method called MOSP to extract key phrases from the original text after that, learns the collocation of phrases. Following
training, the model will generate a phrase sequence that satisfies the syntactic structure criteria. Furthermore, we leverage phrase location information to overcome
the problem of unusual terms, which practically all abstractive models would face. Regarding sequence-to-sequence models, RNN is not most used because it tends
to low-efficiency problems as they rely on the previous step when training, and it must preserve the hidden state of the whole past, thus not able to perform parallel
operations. To overcome these problems, Zhang et al. (2019) proposed a sequence-to-sequence model based on CNN to create the representation of source text.
As we know, traditional CNN can only encode with fixed size contexts of inputs, but in this study, they increase the compelling text by stacking CNN layers over
each other. The length of the sequence under consideration may thus be readily regulated, and each component of the sequence can be computed in parallel. More
commonly, abstractive summarization problems are that the generated summaries are frequently incompatible with the source content in terms of semantics. WEI
et al. (2018) offer a regularisation strategy for the sequence-to-sequence model in this research, and we use what the model has learnt to regularise the learning
objective to mitigate this problem's influence.
Until now, the model discussed does not consider whether the summaries are factually consistent with source documents. Kryściński et al. (2019a) present a
model-based technique for evaluating factual consistency and detecting conflicts between source documents and the output summary that is a weakly supervised
model. The steps of these models are:
• Determine whether sentences are factually consistent after being transformed,
• Find a span in the source documents to validate the consistency prediction, and
• Find an inconsistent span in the summary phrase if one exists.
Another notable work is done in the sequence-to-sequence encoder and decoder approach by Kryściński et al. (2020). That study makes two main contributions.
First, separate extraction and generation at decoder part. The contextual network is standalone for extraction, and a language model generates paraphrases. Second,
optimizing the n-gram overlap while encouraging abstraction with ground-truth summaries.
Wang et al. (2020) provide a unique Generative Adversarial Network (GAN) for Abstractive Text Summarization with a multitask constraint in this research
(PGAN-ATSMT). Through adversarial learning, this model simultaneously trains a generator G and a discriminator D. The sequence-to-sequence architecture is
the backbone of the generative model G, which takes the source document as input and generates the summary. The model uses a language model to implement D
instead of a binary classifier as the discriminative model D, and the output of the language model is used as the reward to steer the generative model. A minimax
two-player game was used to optimize the generative model G and the discriminative model D. Extended work on the GAN network is done by Yang et al. (2021).
They present a new Hierarchical Human-like deep neural network for ATS (HH-ATS), influenced by how humans interpret articles and produce summaries. HH-
ATS comprises three main components (i.e., a knowledge-aware hierarchical attention module, a multitask learning module, and a dual discriminator generative
adversarial network) that reflect the three phases of human reading cognition (i.e., rough reading, active reading, and post-editing).

Table 2: Research survey on Abstractive text summarization method

Citation article Model/methods/te Dataset Used Performance Advantages/Pros Disadvantages/Cons

chniques applied
(Nallapati et al., Abstractive text Abstractive Gigaword ROUGE-1: 35.46 -Performs a little bit well as -Worked for single
2016) summarization (Sequence-to- corpus, ROUGE-2:13.30 compared to state-of-the-art sentence output
using sequence- sequence DUC, ROUGE-L:32.65 -Propose new dataset for summaries
to-sequence encoder-decoder CNN/daily Qualitative multiple sentence summary -capturing the
RNNs and with RNN) mail corpus evaluation: a few meaning of complex
beyond high quality and sentence is weakness
poor-quality output for this model
(Li et al., 2017) Deep recurrent Abstractive Gigaword ROUGE-1: 36.71 Implement deep recurrent Worked for single
generative (Sequence-to- corpus, ROUGE-2:24.00 generative model to capture latent sentence output
decoder for sequence DUC, ROUGE-L:34.10 structure information. summaries
abstractive text encoder-decoder LCSTS
summarization with deep
recurrent
generative
decoder (DRGN))
(Dohare et al., Text Abstractive CNN, ROUGE-1: 39.53 -Suggested a full pipeline for Not a lot of work has
2017b) Summarization summarization DailyMail ROUGE-2:17.28 summarization with AMR been done to extract
using Abstract with AMR. ROUGE-L:36.38 -proven that ROGUE can’t be used AMR graphs for
Meaning for evaluating the abstractive summaries
Representation summaries
-a novel approach for extracting
multiple summary graphs
(Song et al., Abstractive text Abstractive CNN, ROUGE-1: 34.9 -Implement ATSDL (ATS using -Training of deep
2019a) summarization (CNN-LSTM) DailyMail ROUGE-2:17.8 Deep Learning) system that learning is very time-
using LSTM- dataset Combines CNN and LSTM for consuming process.
CNN based better performance -ROUGE can’t
deep learning -solve problem of rare words evaluate the quality of
summary effectively.
(Zhang et al., Abstract text Abstractive Gigaword ROUGE-1: 42.04 -equip the CNN model with GLU In practise, adding
2019) summarization (CNN) corpus, ROUGE-2: 19.77 and residual connections. more sentences has a
with a DUC, ROUGE-L: 39.42 -hierarchical attention mechanism negative impact on
convolutional CNN/daily to generate the keywords and the performance, which
seq2seq model mail corpus key sentences simultaneously. we explain to the fact
-a copying mechanism to extract that the latter
out-of-vocabulary words from sentences are
source text. unrelated to the
summary.
(Kryściński et Evaluating the Abstractive CNN/Daily Accuracy: 74.15 -Implemented factual -common-sense
al., 2019b) Factual Mail dataset F1-Score: 0.5106 consistency checking model mistakes made by
Consistency of (FactCC). summarization
Abstractive models.
Text - stemming errors
Summarization from dependencies
between different
sentences within the
summary
(WEI et al., Regularizing Abstractive Large-Scale ROUGE-1: 36.2 - suggest a method for Language dependent
2018) output Chinese ROUGE-2:24.3 regularising the output word (Chinese dataset)
distribution of Short Text ROUGE-L: 33.8 distribution so that semantic
abstractive Summarizati Accuracy of inconsistency in the training data,
Chinese social on Dataset Human such as terms not linked to the
media text (LCSTS) Evaluation source content, is
summarization approach: 53.6% underrepresented in the model.
for improved -Proposed a simple human
semantic evaluation approach for
consistency determining the generated
summary's semantic
compatibility with the original
information.
 (Kryściński et Improving Abstractive CNN/Daily ROUGE-1: 40.19 -Separate the decoder into a Matches score when
al., 2018) abstraction in (Sequence-to- Mail dataset ROUGE-2:17.38 contextual network and a compared to the state-
text sequence ROUGE-L: 37.52 pretrained language model. of-the-art in relevance
summarization encoder-decoder) Qualitative -Optimized metric through but slightly low in
evaluation: policy learning terms of readability of
Readability: summaries.
6.76/10
Relevance: 6.73/10
(Yang, Wang, et Plausibility- Abstractive Gigaword ROUGE-1: 40.19 -Developed PGAN-ATSMT, an -External Common-
al., 2020) promoting (GAN) corpus, ROUGE-2:17.38 adversarial framework for sense knowledge
generative CNN/daily ROUGE-L: 37.52 abstractive text summarization from the language is
adversarial mail corpus Perplexity:10.21 with multi-task constraint. missing.
network for Qualitative -PGAN-ATSMT jointly trains -Time consuming as
abstractive text evaluation: the task of abstractive text deep learning
summarization Relevance: 3.29/5 summarization and two other algorithm
with multi-task Fluency:3.38/5 related tasks: text classification
constraint and syntax generation.
(Yang, Li, et al., Hierarchical Abstractive Gigaword ROUGE-1: 43.16 HH-ATS extends the seq2seq To evaluate human
2020) Human-Like (GAN) corpus, ROUGE-2:20.32 framework by replicating the judgement better it is
Deep Neural CNN/daily ROUGE-L: 39.14 process of writing a summary for required to explore
Networks for mail corpus Qualitative a piece of text by humans. different automatic
Abstractive evaluation: metrics
Text Informativeness:
Summarization 3.41/5
Fluency:3.32/5
3.3 Hybrid text summarization:
When discussing very precious approaches of Automatic text summarization that are Extractive and Abstractive, both come with their pros and cons. Extractive
summarization is comparatively easier to implement than abstractive summarization, but extractive summarization is not as efficient as user perception. Combining
these methods by strengthening their pros and weakening their cons leads to hybrid methods for text summarization.
Experiments were done on summarization until 1990 were focused on just extracting (reproduced) the summaries from original text rather than abstracting
(newly generated).SUMMRIST system (Hovy & Lin, 1996) was developed with the help of NLP techniques. We can develop a multi-lingual summarizer by
modifying some parts of the structure.
Semantic and statistical features combine extracting and abstracting. The authors of Bhat et al. (2018) used emotions of the text as a semantic feature. Emotions
play a significant role in defining the user's emotional affinity, so lines with implicit emotional content are crucial to the writer and should be included in the
summary. The extracted summary is then put into the Novel language generator, a hybrid summarizer that combines WordNet, Lesk algorithm, and POS to
transform extractive summary into an abstractive summary. Table-3 shows a detailed survey on hybrid text summarization with a particular research paper with
the dataset used, accuracy of the system and its pros and cons.

Table 3. Research survey on hybrid text summarization method

Citation article Model/methods/te Dataset Used Performance Advantages/Pros Disadvantages/Cons

chniques applied
(Hovy & Lin, Automated text Hybrid text Text SUMMARIST By avoiding some language- Lack of good training
n.d.) summarization summarization scores 30% higher specific methods in data in variety of
and the than human SUMMARIST it is possible to domains
SUMMARIST abstracts create multi-lingual summarizer
system easily.
(Bhat et al., SumItUp: A Hybrid Single- DUC-2007 Precision:0.4 to 0.8 -hybrid approach to a single- Not proper evaluation
2018) Hybrid Single- Document Text with 15 news Compression rate: document summarization is done. The results
Document Text summarization articles 35% combining semantic and are compared to MS
Summarizer statistical features word (no ROUGE
-emotional feature metrics)

4 ATS System Evaluation and Evaluation Programs

There have been many efforts to solve summary evaluation issues in the past two decades. NIST (National Institute of Standards and Technology) leads the effort
by organizing the DUC and TAC challenges. (Huang et al., 2010) has given four pillars that should be considered to generate summaries:

• Information Coverage
• Information Significance
• Information Redundancy
• Text Coherence

In the discipline of automatic text summarization, evaluating the summary is a critical task. Evaluating the summary and increasing the development of reusable
resources and infrastructure aids in comparing and replicating findings, adding competition to improve the outcomes. However, carefully evaluating many texts to
acquire an unbiased perspective is impossible. As a result, accurate automatic evaluation measures are required for quick and consistent evaluation. It is difficult
for people to recognize what information should be included in a summary; therefore, evaluating it is difficult. Information changes depending on the summary's
purpose, and mechanically capturing this information is a challenging undertaking (Gambhir & Gupta, 2017). Evaluation of the ATS system is given in fig.7 below:

a) Extrinsic Evaluation: An extrinsic evaluation looks at how it influences the accomplishment of another task (Text classification, Information retrieval,
Question answering). i.e., a summary is termed a good summary if it provides help to other tasks. Extrinsic evaluations have looked at how summarization
affects tasks such as relevance assessment, reading comprehension, etc.

- Relevance evaluation: Various methods are used to analyse a topic's relevance in the summary or the original material.

- Reading comprehension: After reading the summary, it assesses whether it is possible to answer multiple-choice assessments.

b) Intrinsic Evaluation: An Intrinsic evaluation looks at the summarization system on its own. The coherence and informativeness of summaries have been the
focus of intrinsic evaluations. Evaluations based on comparisons with the model summary/summaries and evaluations based on comparisons with the source
document are the two types of intrinsic techniques (Steinberger & Ježek, 2009).

It assesses the quality of a summary by comparing the coverage of a machine-generated and a human-generated summary. The two most significant aspects of
judging a summary are its quality and informativeness. A summary's informativeness is usually assessed by comparing it to a human-made summary, such as a
reference summary. There is also faithfulness to the source paradigm, which examines if the summary contains the same or similar material as the original document.
This method has a flaw: how it can be determined which concepts in the document are relevant and which are not?
 Fig. 7: The evaluation Techniques for Automatic Text Summarization

4.1 Content Evaluation

• Co-selection: Only identical sentences can be used in co-selection measures. It ignores the reality that even though two sentences are written differently, they
can contain the same information. In addition, summaries provided by two separate authors rarely contain similar sentences. Co-selection can be calculated by
precision, recall and F-measure.

a. Precision: Precision is the fraction of relevant instances among the retrieved instances.

|{ } ∩{ }|

 = |{ }|
(1)

 
 = (2)
 + 

b. Recall: Recall is the fraction of relevant instances that were retrieved.

|{ } ∩{ }|

 = |{ }|
(3)

 
 = (4)
 + 

c. F-measure: It is computed by combining recall and precision.

∗
 −  =  ∗ (5)
+

• Content-based: Drawbacks of co-selection methods are handled by content-based methods.

a. Cosine Similarity: Cosine Similarity can be measured as,

∑  
cos(, ) = (6)
√∑() 2 √∑( ) 2

Where, X and Y are representations of a system summary and its reference document based on the vector space model.

b. Unit Overlap: Unit Overlap can be calculated as,

‖∩‖
overlap(, ) = ‖‖+‖‖‖∩‖ (7)

Where, X and Y are representations based on sets of words or lemmas. ‖‖ is the size of set X.

c. Longest Common Subsequence (LCS): the LCS formula is defined as shown in equation (8),

ℎ()+ℎ() (,)
LCS(, ) = (8)


Where, X and Y are representations based on sequences of words or lemmas, LCS (X, Y) is the length of the longest common subsequence between X and Y,
length(X) is the length of the string X, and edit di(X, Y) is the edit distance of X and Y.

d. ROUGE (Recall-Oriented Understudy for Gisting Evaluation): It was firstly introduced by C. Lin & Rey (2001). It contains measures for automatically
determining the quality of a summary by comparing it to other (ideal) summaries generated by people. The measures count the number of overlapping units
such as n-grams, word sequences, and word pairs between the computer-generated summaries to be evaluated and the ideal summaries written by humans.
ROUGE includes five measures like ROUGE-N, ROUGE-L, ROUGE-W, ROUGE-S and ROUGE-SU.

• ROUGE-N counts the number of N-gram units shared by a given summary and a group of reference summaries, where N is the length of the N-gram. i.e.,
ROUGE-1 for unigrams and ROUGE-2 for bi-grams.
• ROUGE-L calculates the LCS (Longest Common Subsequence) statistic. LCS is the maximum size of a common subsequence for two given sequences, X
and Y. ROUGE-L estimates the ratio of the LCS of two summaries to the LCS of the reference summary.
• ROUGE-W is the weighted longest common subsequence metric. It is a step forward from the basic LCS strategy. ROUGE-W prefers LCS with successive
common units. Dynamic programming can be used to compute it efficiently.
• ROUGE-S (Skip-Bigram co-occurrence statistics) calculates the percentage of skip bigrams shared between a single summary and a group of reference
summaries. The skip bigrams are word pairs in the sentence order with random gaps.
• ROUGE-SU is a weighted average of ROUGE-S and ROUGE-1 that expands ROUGE-S to include a unigram counting unit. This is a step forward from
ROUGE-S.
e. LSA-based method: This method was developed by Steinberger & Ježek (2009). If there are m terms and n sentences in the document, we will obtain an m*n
matrix A. The next step is to apply Singular Value Decomposition (SVD) to matrix A. The SVD of an m*n matrix A is defined as given in equation (9):

A = UΣVT (9)

In terms of NLP, SVD (Singular Value Decomposition) is used to generate the document's latent semantic structure, represented by matrix A: that is, a breakdown
of the original document into r linearly-independent basis vectors that express the document's primary 'Topics'. SVD can record interrelationships between terms,
allowing concepts and sentences to be clustered on a 'semantic' rather than a 'word' basis.

4.2 Text Coherence or Quality Evaluation:

a. Grammaticality: The text should not contain non-textual items (i.e., markers), punctuation errors or incorrect words.
b. Non-redundancy: The text should not contain redundant information.
c. Reference clarity: The nouns and pronouns should be referred to in summary. For example, the pronoun he has to mean somebody in the context of the
summary.
d. Coherence and structure: The summary should have good structure, and the sentences should be coherent.

The linguistic characteristics of the summary are properly considered here. Non-redundancy, focus, grammaticality, referential clarity, and structure and coherence
are five questions based on linguistic quality used in DUC (Document Understanding Conference) and TAC (Text Analysis Conference) conferences to evaluate
summaries not needed to be reviewed against the reference summary. Expert human assessors manually score the summary based on its quality, awarding a score
to the summary according to a five-point scale (Gambhir & Gupta, 2017).

The text quality of a summary can also be checked by examining several readability variables. Text quality is analysed using various criteria such as vocabulary,
syntax, and discourse to estimate a correlation between these characteristics and previously acquired human readability ratings. Unigrams represent vocabulary,
while the average number of verbs or nouns represent syntax.
4.3 Automatic Text Summarization Evaluation Programs

SUMMAC (TIPSTER Text Summarization Evaluation) was the first conference where automatic summarization systems were reviewed, and it was held at the end
of the 1990s, where text summaries were assessed using two extrinsic and intrinsic criteria. DUC (Document Understanding Conferences), which took place every
year from 2001 to 2007, is another notable conference for text summarizing. Initially, activities at DUC conferences like DUC 2001 and DUC 2002 featured generic
summarizing of single and multiple documents, which was later expanded to include a query-based summary of multiple documents in DUC 2003. Topic-based
single and multi-document cross-lingual summaries were assessed in DUC 2004. Multi-document, query-based summaries were examined in DUC 2005, and DUC
2006, whilst multi-document, update, query-based summaries were evaluated in DUC 2007. However, in 2007, DUC conferences were no longer held because
they were absorbed into the Text Analysis Conference (TAC), which featured summarization sessions. TAC is a series of evaluation workshops designed to promote
research in the domains of Natural Language Processing and related fields. The TAC QA program arose from the TREC QA program. The Summarization track
aids in the development of methods for producing concise, cohesive text summaries. Every year TAC workshops have been held since 2008.

5. Frequently used Dataset for ATS

There are applications of ATS systems that are widely spread worldwide and know the available data globally. So, to perform the text summarization task essential
thing is the data. Not all data can be directed feed to the system. It required prepossessing and other treatments. The machine learning-based approaches need a
huge training dataset with ideal summaries to train the model. Also, the ideal or sample Dataset is needed to evaluate a particular ATS system. That sample data is
manually generated or created by human researchers. The list of the Dataset available for the ATS task is very long. A very few datasets are given below:

• DUC: The National Institute of Standards and Technology (NIST) provides these datasets, the most prevalent and widely used datasets in text summarization
research. The DUC corpora were distributed as part of the DUC conference's summarizing shared work. The most recent DUC challenge took place in 2007.
Datasets for DUC 2001 through DUC 2007 are available on the DUC website.
• Text Analysis Conference (TAC) Datasets: DUC was added to the TAC as a summary track in 2008. To gain access to the TAC datasets, you must first fill
out the application forms available on the TAC website.
• Gigaword: Created by (Rush et al., 2015), Headline-generation on a corpus of article pairs from Gigaword consisting of around 4 million articles in the English
language.
• LCTCS: Created by Chen (2015), the LCSTS Dataset was constructed from the Chinese microblogging website Sina Weibo. It consists of over 2 million real
Chinese short texts with short summaries given by the author of each text. Requires application in the Chinese language.
• wikiHow: Created by Koupaee & Wang (2018), the WikiHow Dataset contains article and summary pairs extracted and constructed from an online knowledge
base written by different human authors in the English language. There are two features: - text: wikiHow answers texts. - headline: bold lines as summary.
• CNN: CNN/DailyMail non-anonymized summarization dataset. The CNN / Daily Mail Dataset is an English-language dataset containing just over 300k unique
news articles written by journalists at CNN and the Daily Mail. The current version supports both extractive and abstractive summarization, though the original
version was created for machine-reading and comprehension and abstractive question answering.
6. Application, Challenges and future scope
6.1 Applications of ATS

There are numerous uses for automatic text summarizing. As we see how text summarization is divided into many more categories. All these categories further
lead us to treasure of ATS’s applications. This subsection includes some of the applications of ATS system. Table 4 shows the research survey on application of
ATS system. The table includes article name, the method & the dataset used in a particular article, the performance of the system proposed in that particular research
study and advantages & disadvantages of that article.

Here are a few examples:

• Improving the performance of classic IR and IE systems (using a summarization system in conjunction with a Question-Answering (QA) system); (De
Tre et al., 2014) (S. Liu et al., 2012) (Perea-Ortega et al., 2013)
• News Summarization and Newswire generation (Tomek,1998) (Bouras & Tsogkas, 2010)
• Rich Site Summary (RSS) feed summarization (Zhan et al., 2009)
• Blog Summarization (Y. H. Hu et al., 2017)
• Tweet Summarization (Chakraborty et al., 2019)
• Web page Summarization (Shen et al., 2007)
• Email and email thread Summarization (Muresan et al., 2001)
• Report Summarization for business men, politicians, researchers, etc. (Lloret et al., 2013)
• Meeting Summarization.
• Biographical extracts
• Legal Document Summarization (Farzindar & Lapalme, 2004)
• Books Summarization (Mihalcea & Ceylan, 2007)
• Use of Summarization in medical field (Feblowitz et al., 2011)(Ramesh et al., 2015)
 Table 4: Research survey on Application of ATS System

Citation Article Model/methods Dataset Used Performance Advantages/Pros Disadvantages/Con

applied s
(Tomek,19 A Robust Multi- New York Length of - The algorithm is very robust that Quality of
98) Practical Text Document, times online summaries: efficiently process a large range of summaries not that
Summarization Extractive Text news The summaries are documents, domain independent and much good. It can
summarization only 5% to 10% of can be easily accepted by most of the be improved by
original text so it European languages. additional
can be quickly read - It can work in two modes: Generic and paragraph scoring
and understood. Topical function.
-Worked on DMS (Discourse Macro
Structure) to conquer the shortcomings
of sentence-based summarization by
working on paragraph level
(Muresan et Combining Extractive text Emails Precision: 83% Linguistic Knowledge Enhances Deep linguistic
al., 2001) linguistic and summarization Recall:85.7% Machine Learning knowledge is
machine learning with machine required.
techniques for learning
email
summarization
(Mckeown Tracking and Multi-document News sites - These research Personalization of
et al., 2002) summarizing summarization achievements are incorporated into Newsblaster and
news on a daily with extractive “Newsblaster” restricting it to user
basis with approach preferred topics or
Columbia's questions is still
Newsblaster having to be done
(Farzindar Legal Text Extractive Text Corpus Preliminary results -Summary presented as table style The System is not
& Lapalme, Summarization summarization contains are very promising. -focused on many categories of properly for many
2004) by Exploration 3500 -F-measures: 0.935 judgements i.e., Copyright, Air low, other remaining
of the Thematic judgements on average for all human rights etc. categories of
Structure and of Federal stages judgements
Argumentative court of
Roles Canada
(Mihalcea Explorations in Extractive text gold F-Measure: 0.404 it introduced a new summarization Exhaustive method
& Ceylan, automatic book summarization standard” benchmark, specifically targeting the when we have short
2007) summarization with data set of 50 evaluation length book
TEXTRANK books of systems for book summarization
approach
(Ling et al., Generating gene Multi- test set with ROUGE-N Semi- structured summaries are - High quality of
2007) summaries from Document, 20 genes performs better than generated which consist of sentences data required. (Data
biomedical Extractive Text existing methods regarding specific semantic aspects of a dependent)
literature: A summarization gene. -Redundant
study of semi- information in
structured generated summary
summarization
(Shen et al., Noise reduction Extractive Text 2 million F-measures for Removing noise from web pages while - Focus only
2007) through summarization Web pages hybrid approach preserving most relevant features to isolated web pages.
summarization crawled (supervised and increase accuracy of web classification Does not include
for Web-page from the unsupervised): hyperlinks.
classification LookSmart Naïve Byes- 72.0 ±
Web 0.3
directory SVM- 72.9 ± 0.3
(Zhan et al., Gather customer Extractive Text Five datasets Average find and extracts important topics from - For different
2009) concerns from summarization from Hu’s responsiveness a set of online reviews and then ranks online sites the style
online product corpus (M. scores: 4.3 these retrieved topics of reviews is written
reviews – A text Hu & Liu, in differently. This
summarization 2004) and 3 difference is not
approach sets from efficiently
Amazon entertained.
(Bouras & Noun retrieval Multi- Numerous Precision is boosted -Enhance personalization algorithm Language
Tsogkas, effect on text Document news portals by using noun with help of various features extracted dependent
2010) summarization Extractive Text around the retrieval effect. from user’s profile and viewed history
and delivery of summarization internet With new of articles.
personalized personalization - A stable system for day-to-day use is
news articles to scheme increase to build named as PeRSSonal.
the user's around 17% for
desktop precision and 14%
for recall.
(Feblowitz Summarization Extractive Text Day to day Both computer supported and computer Not standardised
et al., 2011) of clinical summarization clinical data independent clinical tasks are analysed. and optimized
information: A - clinical summary
conceptual
model
(Lloret et Text Extractive Text Most Query-focused Efficiently combines text The summary for a
al., 2011) Summarization summarization relevant 20 summaries gives summarization for semantic question particular question’s
Contribution to documents 58% improvement answering while focusing on query- answer was built
Semantic from google in accuracy than based summaries rather than generic from only 20
Question search generic summaries. summaries. documents which
Answering: New engine for shorter summaries were retrieved from
Approaches for particular gives 6.3% google search
Finding Answers Question. improvement in engine.
on the Web accuracy than long
Elena summaries.
(S. Liu et TIARA: Multi- IBM Improvement in -An Interactive visual text analysis tool Application –
al., 2012) Interactive, Document employees’ TIARA with TIARA which produces a visual specific features are
topic-based Extractive Text Email regarding to summary of text analytic results not added in the
visual text summarization usefulness and automatically and enhanced with tool.
summarization satisfaction of significant, time-sensitive topic-based
and analysis summary than summary method.
TheMail. (Viégas et
al., 2006)
(Kavila et An automatic Extractive text corpus 90% results in Hybrid system of Keyword/ Key phrase Future work will a
al., 2013) legal document summarization presently automatic search matching technique and hybrid system to
summarization consists of Case based technique find the expected
and search using 100 legal judgement for a new
hybrid system documents case from the past
related to cases.
criminal and
civil
collected
(Lloret et COMPENDIUM Hybrid Text 50 research -quantitative and -COMPENDIUM produces summaries Extractive
al., 2013) : A text summarization papers from qualitative for biomedical research papers summaries
summarization journal of evaluation automatically with both approaches: generated by system
system for medicine approach extractive and abstractive. are contain less
generating -Precision: 40.53 -abstractive are far better as per user similar information
abstracts of Recall:44.02 perspective. to human written
research papers summaries
(Perea- Application of Extractive Text 169,477 GeoCLEF: Two types of summaries are generated - approach can be
Ortega et text summarization documents Compression rate (generic & geographical) with only applied to
al., 2013) summarization collections ranging from 20% improvement in recall and compression single document
techniques to the of stories and to 90% rates. -discarding all the
geographical newswires Recall sentences which do
information MAP not contain any
retrieval task geographical
information may
leads to loss of
information.
(Sankarasu Text Single and DUC ROUGE-1: 0.46 -Works for both single document and Evaluation is done
bramaniam summarization Multi- (2002)-567 ROUGE-2: 0.23 multi document against human-
et al., 2014) using Wikipedia Document English - For selecting summary phrases, generated summary
Extractive Text news articles created a bipartite sentence–concept which naturally
summarization graph and presented an iterative ranking favours the leading
algorithm. line sin new article.
(Ramesh et Figure- Extractive Text 94 annotated F1 score of 0.66 and Unsupervised FigSum+ system that -It requires
al., 2015) Associated Text summarization figures ROUGE-1 score of automatically fetches associated texts, annotated figures.
Summarization selected 0.97 remove repetitions, and generate text - Experiments
and Evaluation from 19 summary for a particular figure. results are generated
different by only 94 figures.
journals
(Y. H. Hu Opinion mining Machine Reviews of Mean, standard -Proposed technique fetches top-k - TripAdvisor.com
et al., 2017) from online hotel learning based Red Roof deviation and informative sentences from online contains reviews in
reviews – A text Extractive Text Inn and significance level. reviews many more
summarization summarization Gansevoort -Proposed method focused on critical languages but the
approach Meatpacking factors such as usefulness of a review, proposed approach
Hotel credibility of authors, review time, works for only
selected conflicts in reviews. English.
from - Small sample of
TripAdvisor. participants with
com. same backgrounds.
- Experiments are
done for only 2
hotels.
(Lovinger Gist: general Extractive text Movie Average F- High F-score compared to TextRank High computational
et al., 2017) integrated summarization reviews, measure: 0.276 and LexRank time and cost
summarization with news articles Average running
of text and optimization- time: 0.067235
reviews Based approach
(Chakrabor Tweet Extractive text News article Rouge-1, F1-score: Capturing the diverse opinions helps in classification of
ty et al., summarization summarization dataset and 0.664 better identification of the relevant opinion from tweets
2019) of news articles: with tweet dataset Rouge-2, F1-score: tweet set. are not implemented
An objective LEXRANK 0.548
ordering-based approach
perspective
(Kumar & Factual instance Extractive text IPL 2017 Sentiment classifier classification of sub-events of tweets -opinion
Reddy, tweet summarization challenger 1, time: -Logistic classification is not
2019) summarization eliminator, regression derived.
and opinion challenger 2, Train: 0.27s
analysis of sport and final Test: 0.025
competition competition
with each
having
10,000
tweets

6.2 Challenges and future scope

While generating an automated text summary, one faces a lot of challenges. The first challenge is defining what constitutes a decent summary, or more precisely,
how a summary might be constructed. Our requirements for a summary provide good clues as to what it should be: extractive or abstractive, general, or query-
driven, etc. Even if we figure out how humans normally summarise, putting it into practice will be difficult. Creating a powerful automatic text summarizer
necessitates many resources, either in terms of tools or corpora. Another challenge is summary in formativeness; how can a machine emulate human people when
it comes to summarising? One of the long-standing issues has been the coherence of the summary. The shortage of resources is another most challenging problems
in ATS.
Compared to the past, there are numerous powerful tools for stemming, parsing, and other tasks available now. Despite this, determining which ones are
appropriate for a particular summarization problem is the issue. Furthermore, annotated corpus for ATS can be considered a challenge. The evaluation process is
also a significant difficulty. Both intrinsic and extrinsic evaluation approaches were explored in this work. The language shared by a machine-generated reference
summary is generally the focus of current intrinsic evaluation methods. Intuitive evaluation can create new ways to evaluate the summary based on the information
it includes and how it is presented. The process of evaluation is highly subjective. First, a reasonable criterion must be defined to understand what is important and
what is not. It is also unclear whether this procedure can be fully automated.
Text summarising has been around for more than fifty years, and the academic community is very interested in it; therefore, they continue to improve existing
text summarization ways or invent new summary approaches to provide higher-quality summaries. However, text summarization performance is still average, and
the summaries created are not perfect. As a result, by merging this system with other systems, it can be made more intelligent, allowing the combined system to
perform better.
• Conclusion

Automatic Text summarization reduces the size of a source text while maintaining its information value and overall meaning. Automatic Text summarization has
become a powerful technique for analysing text information due to a large amount of information we are given and the growth of Internet technologies. The
automatic summarization of text is a thriving-known task in natural language processing (NLP). Automatic text summarization is an exciting research area, and it
has a treasure of applications. This paper aims to make readers understand automatic text summarization from ground level and familiarise them with all detailed
types of ATS systems. After that all different types are distinguished deeply and clearly in this study. The summarization task is mainly divided into extractive and
abstractive. The study shows numerous techniques for extractive summarization, but the summaries generated by extractive summarizers are far from human-made
summaries. On the other hand, abstractive summarizer is close to human summaries but not practically implemented with high performance. The combination of
both extractive and abstractive is hybrid text summarization. This paper includes research survey on Extractive, Abstractive and Hybrid Text Summarization. Also,
this survey article tried to cover all major application areas of ATS system and provided detailed survey on the same. There are so many methods to evaluate
summarizing system and generated summaries that are included in this paper. Further it gives brief idea about frequently used datasets, conferences and programs
that held every year for automatic text summarization system.
. The future study is to build a robust, domain and language independent extractive text summarization that works well with multi-documents. Similarly, because
the quality evaluation of the summary is done manually by experienced assessors, it is highly subjective. There are specific quality assessment criteria, such as
grammaticality and coherence, but the results are different when two experts evaluate the same summary.

Conflicts of Interest: On behalf of all authors, I state that there is no conflict of interest.
References
N. D. (2016). Automatic Text Summarization Using Supervised Machine Learning Technique for Hindi Langauge. International Journal of Research in
Engineering and Technology, 05(06), 361–367. https://doi.org/10.15623/ijret.2016.0506065
Ab, A., & Sunitha, C. (n.d.). An Overview on Document Summarization Techniques. 113–118.
Al-Radaideh, Q. A., & Bataineh, D. Q. (2018a). A Hybrid Approach for Arabic Text Summarization Using Domain Knowledge and Genetic Algorithms. Cognitive
Computation, 10(4), 651–669. https://doi.org/10.1007/s12559-018-9547-z
Al-Radaideh, Q. A., & Bataineh, D. Q. (2018b). A Hybrid Approach for Arabic Text Summarization Using Domain Knowledge and Genetic Algorithms. Cognitive
Computation, 10(4), 651–669. https://doi.org/10.1007/s12559-018-9547-z
Alami, N., Mallahi, M. El, Amakdouf, H., & Qjidaa, H. (2021). Hybrid method for text summarization based on statistical and semantic treatment. Multimedia
Tools and Applications. https://doi.org/10.1007/s11042-021-10613-9
Alami, N., Meknassi, M., & En-nahnahi, N. (2019). Enhancing unsupervised neural networks based text summarization with word embedding and ensemble
learning. Expert Systems with Applications, 123, 195–211. https://doi.org/10.1016/j.eswa.2019.01.037
Alami, N., Meknassi, M., En-nahnahi, N., El Adlouni, Y., & Ammor, O. (2021). Unsupervised neural networks for automatic Arabic text summarization using
document clustering and topic modeling. Expert Systems with Applications, 172. https://doi.org/10.1016/j.eswa.2021.114652
ALGULIEV, R., & ALIGULIYEV, R. (2009). Evolutionary Algorithm for Extractive Text Summarization. Intelligent Information Management, 01(02), 128–138.
https://doi.org/10.4236/iim.2009.12019
Alguliev, R. M., Aliguliyev, R. M., Hajirahimova, M. S., & Mehdiyev, C. A. (2011a). MCMR: Maximum coverage and minimum redundant text summarization
model. Expert Systems with Applications, 38(12), 14514–14522. https://doi.org/10.1016/j.eswa.2011.05.033
Alguliev, R. M., Aliguliyev, R. M., Hajirahimova, M. S., & Mehdiyev, C. A. (2011b). MCMR: Maximum coverage and minimum redundant text summarization
model. Expert Systems with Applications, 38(12), 14514–14522. https://doi.org/10.1016/j.eswa.2011.05.033
Alguliyev, R. M., Aliguliyev, R. M., Isazade, N. R., Abdi, A., & Idris, N. (2019a). COSUM: Text summarization based on clustering and optimization. Expert
Systems, 36(1), 1–17. https://doi.org/10.1111/exsy.12340
Alguliyev, R. M., Aliguliyev, R. M., Isazade, N. R., Abdi, A., & Idris, N. (2019b). COSUM: Text summarization based on clustering and optimization. Expert
Systems, 36(1). https://doi.org/10.1111/exsy.12340
Antiqueira, L., Oliveira, O. N., Costa, L. da F., & Nunes, M. das G. V. (2009). A complex network approach to text summarization. Information Sciences, 179(5),
584–599. https://doi.org/10.1016/j.ins.2008.10.032
Aone, C., Okurowski, M. E., Gorlinsky, J., & Larsen, B. (1997). A Scalable Summarization System Using Robust NLP. Proceedings of the Intelligent Scalable
Text Summarization Workshop, 66–73.
Banko, M., Mittal, V. O., & Witbrock, M. J. (2000). Headline generation based on statistical translation. 318–325. https://doi.org/10.3115/1075218.1075259
Barzilay, R., & Mckeown, K. R. (2005). Sentence Fusion for Multidocument News Summarization.
Belwal, R. C., Rai, S., & Gupta, A. (2021). A new graph-based extractive text summarization using keywords or topic modeling. Journal of Ambient Intelligence
and Humanized Computing, 12(10), 8975–8990. https://doi.org/10.1007/s12652-020-02591-x
Bhat, I. K., Mohd, M., & Hashmy, R. (2018). SumItUp: A Hybrid Single-Document Text Summarizer. Advances in Intelligent Systems and Computing, 583(April),
619–634. https://doi.org/10.1007/978-981-10-5687-1_56
Binwahlan, M. S., Salim, N., & Suanmali, L. (2009a). Swarm Based Text Summarization. 2009 International Association of Computer Science and Information
Technology - Spring Conference, IACSIT-SC 2009, 145–150. https://doi.org/10.1109/IACSIT-SC.2009.61
Binwahlan, M. S., Salim, N., & Suanmali, L. (2010). Fuzzy swarm diversity hybrid model for text summarization. Information Processing and Management, 46(5),
571–588. https://doi.org/10.1016/j.ipm.2010.03.004
Binwahlan, M. S., Salim, N., & Suanmali, L. (2009b). Swarm Based Text Summarization. 2009 International Association of Computer Science and Information
Technology - Spring Conference, IACSIT-SC 2009, 145–150. https://doi.org/10.1109/IACSIT-SC.2009.61
Bloom, ; E H, Fischer, H., Charbonneau, N. K., Tonks, ; J, Mirkovitch, J. E., Sadowski, D. ; H. B., Shuai, K., Darnell, J. E., & Gilman, M. Z. (n.d.). Automatic
Analysis, Theme Generation, and Summarization of Machine-Readable Texts. In Interferon: Principles and Medical Applications (Vol. 13).
Bouras, C., & Tsogkas, V. (2010). Noun retrieval effect on text summarization and delivery of personalized news articles to the user’s desktop. Data and Knowledge
Engineering, 69(7), 664–677. https://doi.org/10.1016/j.datak.2010.02.005
Brin, S., & Page, L. (2012). Reprint of: The anatomy of a large-scale hypertextual web search engine. Computer Networks, 56(18), 3825–3833.
https://doi.org/10.1016/j.comnet.2012.10.007
Carbonell, J., & Goldstein, J. (1998). The use of MMR, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st
annual international ACM SIGIR conference on Research and development in information retrieval, 335-336.
Chakraborty, R., Bhavsar, M., Dandapat, S. K., & Chandra, J. (2019). Tweet summarization of news articles: An objective ordering-based perspective. IEEE
Transactions on Computational Social Systems, 6(4), 761–777. https://doi.org/10.1109/TCSS.2019.2926144
Chen, Q. (2015). LCSTS: A Large Scale Chinese Short Text Summarization Dataset.
De Tre, G., Hallez, A., & Bronselaer, A. (2014). Performance optimization of object comparison. International Journal of Intelligent Systems, 29(2), 495–524.
https://doi.org/10.1002/int
Dohare, S., Karnick, H., & Gupta, V. (2017a). Text Summarization using Abstract Meaning Representation. ArXiv.
Dohare, S., Karnick, H., & Gupta, V. (2017b). Text Summarization using Abstract Meaning Representation. http://arxiv.org/abs/1706.01678
El-Kassas, W. S., Salama, C. R., Rafea, A. A., & Mohamed, H. K. (2020). EdgeSumm: Graph-based framework for automatic text summarization. Information
Processing and Management, 57(6), 102264. https://doi.org/10.1016/j.ipm.2020.102264
Farzindar, A., & Lapalme, G. (2004). Legal Text Summarization by Exploration of the Thematic Structure and Argumentative Roles. In Text Summarization
Branches Out Conference Held in Conjunction with ACL 2004, 27–38.
Feblowitz, J. C., Wright, A., Singh, H., Samal, L., & Sittig, D. F. (2011). Summarization of clinical information: A conceptual model. Journal of Biomedical
Informatics, 44(4), 688–699. https://doi.org/10.1016/j.jbi.2011.03.008
Ferreira, R., De Souza Cabral, L., Freitas, F., Lins, R. D., De França Silva, G., Simske, S. J., & Favaro, L. (2014). A multi-document summarization system based
on statistics and linguistic treatment. Expert Systems with Applications, 41(13), 5780–5787. https://doi.org/10.1016/j.eswa.2014.03.023
Ferreira, R., De Souza Cabral, L., Lins, R. D., Pereira E Silva, G., Freitas, F., Cavalcanti, G. D. C., Lima, R., Simske, S. J., & Favaro, L. (2013). Assessing sentence
scoring techniques for extractive text summarization. In Expert Systems with Applications (Vol. 40, Issue 14, pp. 5755–5764).
https://doi.org/10.1016/j.eswa.2013.04.023
Gambhir, M., & Gupta, V. (2017). Recent automatic text summarization techniques: a survey. Artificial Intelligence Review, 47(1). https://doi.org/10.1007/s10462-
016-9475-9
Garner, R. (1982). Efficient text summarization: Costs and benefits. Journal of Educational Research, 75(5), 275–279.
https://doi.org/10.1080/00220671.1982.10885394
Gupta, V., & Kaur, N. (2016). A Novel Hybrid Text Summarization System for Punjabi Text. Cognitive Computation, 8(2), 261–277.
https://doi.org/10.1007/s12559-015-9359-3
HANSON ER. (1971). Musicassette Interchangeability. the Facts Behind the Facts. AES: Journal of the Audio Engineering Society, 19(5), 417–425.
He, J., Kryscinski, W., McCann, B., Rajani, N., & Xiong, C. (2020). CTRLsum: Towards generic controllable text summarization. ArXiv, 1–35.
He, J., Kryściński, W., McCann, B., Rajani, N., & Xiong, C. (2020). CTRLsum: Towards Generic Controllable Text Summarization.
http://arxiv.org/abs/2012.04281
Hovy, E., & Lin, C.-Y. (n.d.). AUTOMATED TEXT SUMMARIZATION AND THE SUMMARIST SYSTEM.
Hovy, E., & Lin, C.-Y. (1996). Automated text summarization and the SUMMARIST system. 197. https://doi.org/10.3115/1119089.1119121
Hu, M., & Liu, B. (2004). Mining opinion features in customer reviews. Proceedings of the National Conference on Artificial Intelligence, 755–760.
Hu, Y. H., Chen, Y. L., & Chou, H. L. (2017). Opinion mining from online hotel reviews – A text summarization approach. Information Processing and
Management, 53(2), 436–449. https://doi.org/10.1016/j.ipm.2016.12.002
Huang, L., He, Y., Wei, F., & Li, W. (2010). Modeling Document Summarization as Multi-objective Optimization. 2–6. https://doi.org/10.1109/IITSI.2010.80
Ježek, K., Katedra, J. S., & Steinberger, J. (2007). Automatic Text Summarization (The state of the art 2007 and new challenges).
Jezek, K., & Steinberger, J. (2008). Automatic summarizing: (The state of the art 2007 and new challenges). Proceedings of Znalosti, 1–12.
Joshi, A., Fidalgo, E., Alegre, E., & Fernández-Robles, L. (2019). SummCoder: An unsupervised framework for extractive text summarization based on deep auto-
encoders. Expert Systems with Applications, 129, 200–215. https://doi.org/10.1016/j.eswa.2019.03.045
Kanapala, A., Pal, S., & Pamula, R. (2019). Text summarization from legal documents: a survey. Artificial Intelligence Review, 51(3), 371–402.
https://doi.org/10.1007/s10462-017-9566-2
Kavila, S. D., Puli, V., Raju, G. S. V. P., & Bandaru, R. (2020). An Automatic Legal Document Summarization and Search Using Hybrid System An Automatic
Legal Document Summarization. January 2013. https://doi.org/10.1007/978-3-642-35314-7
Kiyani, F., & Tas, O. (2017). A survey automatic text summarization. Pressacademia, 5(1), 205–213. https://doi.org/10.17261/pressacademia.2017.591
Ko, Y., & Seo, J. (2008). An effective sentence-extraction technique using contextual information and statistical approaches for text summarization. Pattern
Recognition Letters, 29(9), 1366–1371. https://doi.org/10.1016/j.patrec.2008.02.008
Koupaee, M., & Wang, W. Y. (2018). WikiHow: A Large Scale Text Summarization Dataset. http://arxiv.org/abs/1810.09305
Kryściński, W., McCann, B., Xiong, C., & Socher, R. (2019a). Evaluating the factual consistency of abstractive text summarization. ArXiv.
https://doi.org/10.18653/v1/2020.emnlp-main.750
Kryściński, W., McCann, B., Xiong, C., & Socher, R. (2019b). Evaluating the Factual Consistency of Abstractive Text Summarization.
http://arxiv.org/abs/1910.12840
Kryściński, W., Paulus, R., Xiong, C., & Socher, R. (2018). Improving Abstraction in Text Summarization. http://arxiv.org/abs/1808.07913
Kryściński, W., Paulus, R., Xiong, C., & Socher, R. (2020). Improving abstraction in text summarization. Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing, EMNLP 2018, 1808–1817. https://doi.org/10.18653/v1/d18-1207
Kumar, N. V., & Reddy, M. J. (n.d.). Factual Instance Tweet Summarization and Opinion Analysis of Sport Competition. Springer Singapore.
https://doi.org/10.1007/978-981-13-3393-4
Kutlu, M., Ciǧir, C., & Cicekli, I. (2010). Generic text summarization for Turkish. Computer Journal, 53(8), 1315–1323. https://doi.org/10.1093/comjnl/bxp124
Li, P., Lam, W., Bing, L., & Wang, Z. (2017). Deep Recurrent Generative Decoder for Abstractive Text Summarization. http://arxiv.org/abs/1708.00625
Lin, C., & Rey, M. (2001). R OUGE : A Package for Automatic Evaluation of Summaries.
Lin, C. Y. (1999). Training a selection function for extraction. International Conference on Information and Knowledge Management, Proceedings, 55–62.
https://doi.org/10.1145/319950.319957
Ling, X., Jiang, J., He, X., Mei, Q., Zhai, C., & Schatz, B. (2007). Generating gene summaries from biomedical literature: A study of semi-structured summarization.
Information Processing and Management, 43(6), 1777–1791. https://doi.org/10.1016/j.ipm.2007.01.018
Linhares Pontes, E., Huet, S., Torres-Moreno, J. M., & Linhares, A. C. (2020). Compressive approaches for cross-language multi-document summarization. Data
and Knowledge Engineering, 125. https://doi.org/10.1016/j.datak.2019.101763
Liu, F., Flanigan, J., Thomson, S., Sadeh, N., & Smith, N. A. (2015). Toward abstractive summarization using semantic representations. NAACL HLT 2015 - 2015
Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Proceedings of the Conference,
1077–1086. https://doi.org/10.3115/v1/n15-1114
Liu, S., Zhou, M. X., Pan, S., Song, Y., Qian, W., Cai, W., & Lian, X. (2012). TIARA: Interactive, topic-based visual text summarization and analysis. ACM
Transactions on Intelligent Systems and Technology, 3(2). https://doi.org/10.1145/2089094.2089101
Lloret, E., Llorens, H., Moreda, P., Saquete, E., & Palomar, M. (2011). Text summarization contribution to semantic question answering: New approaches for
finding answers on the web. International Journal of Intelligent Systems, 26(12), 1125–1152. https://doi.org/10.1002/int.20502
Lloret, E., Romá-Ferri, M. T., & Palomar, M. (2013). COMPENDIUM: A text summarization system for generating abstracts of research papers. Data and
Knowledge Engineering, 88, 164–175. https://doi.org/10.1016/j.datak.2013.08.005
Lorch, R. F., Lorch, E. P., Ritchey, K., McGovern, L., & Coleman, D. (2001). Effects of Headings on Text Summarization. Contemporary Educational Psychology,
26(2), 171–191. https://doi.org/10.1006/ceps.1999.1037
Lovinger, J., Valova, I., & Clough, C. (2017). Gist : general integrated summarization of text and reviews. Soft Computing. https://doi.org/10.1007/s00500-017-
2882-2
Mashechkin, I. V., Petrovskiy, M. I., Popov, D. S., & Tsarev, D. V. (2011). Automatic text summarization using latent semantic analysis. Programming and
Computer Software, 37(6), 299–305. https://doi.org/10.1134/S0361768811060041
Mckeown, K. R., Klavans, J. L., & Schiffman, B. (n.d.). Tracking and Summarizing News on a Daily Basis with Columbia ’ s Newsblaster.
Miao, Y., & Blunsom, P. (2016). Language as a latent variable: Discrete generative models for sentence compression. EMNLP 2016 - Conference on Empirical
Methods in Natural Language Processing, Proceedings, 319–328. https://doi.org/10.18653/v1/d16-1031
Mihalcea, R., & Ceylan, H. (2007). Explorations in automatic book summarization. EMNLP-CoNLL 2007 - Proceedings of the 2007 Joint Conference on Empirical
Methods in Natural Language Processing and Computational Natural Language Learning, June, 380–389.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space. 1st International Conference on Learning
Representations, ICLR 2013 - Workshop Track Proceedings, 1–12.
Mohd, M., Jan, R., & Shah, M. (2020). Text document summarization using word embedding. Expert Systems with Applications, 143.
https://doi.org/10.1016/j.eswa.2019.112958
Moradi, M., Dashti, M., & Samwald, M. (2020). Summarization of biomedical articles using domain-specific word embeddings and graph ranking. Journal of
Biomedical Informatics, 107(May), 103452. https://doi.org/10.1016/j.jbi.2020.103452
Muresan, S., Tzoukermann, E., & Klavans, J. L. (2001). Combining linguistic and machine learning techniques for email summarization. 1–8.
https://doi.org/10.3115/1117822.1117837
Mutlu, B., Sezer, E. A., & Akcayol, M. A. (2020). Candidate sentence selection for extractive text summarization. Information Processing and Management, 57(6).
https://doi.org/10.1016/j.ipm.2020.102359
Nallapati, R., Zhou, B., dos Santos, C., Gulçehre, Ç., & Xiang, B. (2016). Abstractive text summarization using sequence-to-sequence RNNs and beyond. CoNLL
2016 - 20th SIGNLL Conference on Computational Natural Language Learning, Proceedings, 280–290. https://doi.org/10.18653/v1/k16-1028
Neto, J. L., Freitas, A. A., & Kaestner, C. A. A. (2002). Automatic text summarization using a machine learning approach. Lecture Notes in Computer Science
(Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2507(i), 205–215. https://doi.org/10.1007/3-540-36127-8_20
Nomoto, T., & Matsumoto, Y. (2003). The diversity-based approach to open-domain text summarization. Information Processing and Management, 39(3), 363–
389. https://doi.org/10.1016/S0306-4573(02)00096-1
Pembe, F., & Güngör, T. (2007). Automated Querybiased and Structure-preserving Text Summarization on Web Documents. … on Innovations in Intelligent
Systems and ….
https://pdfs.semanticscholar.org/7db2/fde4c2e4ba7aa7949780435ef4120f9a24ce.pdf%0Ahttp://www.cmpe.boun.edu.tr/~gungort/papers/automated query-
biased and structure-preserving text summarization on web documents.pdf
Perea-Ortega, J. M., Lloret, E., Alfonso Ureña-López, L., & Palomar, M. (2013). Application of text summarization techniques to the geographical information
retrieval task. Expert Systems with Applications, 40(8), 2966–2974. https://doi.org/10.1016/j.eswa.2012.12.012
Proceedings of the 2009 2nd International Conference on Computer Science and Its Applications : CSA 2009. (2009). IEEE.
Radev, D. R., Jing, H., Styś, M., & Tam, D. (2004). Centroid-based summarization of multiple documents. Information Processing and Management, 40(6), 919–
938. https://doi.org/10.1016/j.ipm.2003.10.006
Ramesh, B. P., Sethi, R. J., & Yu, H. (2015). Figure-associated text summarization and evaluation. PLoS ONE, 10(2). https://doi.org/10.1371/journal.pone.0115671
Rani, R., & Lobiyal, D. K. (2021). An extractive text summarization approach using tagged-LDA based topic modeling. Multimedia Tools and Applications, 80(3),
3275–3305. https://doi.org/10.1007/s11042-020-09549-3
Rau, L. F., Jacobs, P. S., & Zernik, U. (1989). INFORMATION EXTRACTION AND TEXT SUMMARIZATION USING LINGUISTIC KNOWLEDGE
ACQUISITION*. In Informarron Processrng & Managemenr (Vol. 25, Issue 4).
Reimer, U., Hahn, U., Life, S., & Unlverslty, F. (n.d.). A Formal Model of Text Summarization Based on Condensation Operators of a Terminological Logic.
Rush, A. M., Chopra, S., & Weston, J. (2015). A neural attention model for sentence summarization. Conference Proceedings - EMNLP 2015: Conference on
Empirical Methods in Natural Language Processing, 379–389.
Salton, G., Singhal, A., Mitra, M., & Buckley, C. (1997). Automatic text structuring and summarization. Information Processing and Management, 33(2), 193–
207. https://doi.org/10.1016/S0306-4573(96)00062-3
Sanchez-Gomez, J. M., Vega-Rodríguez, M. A., & Pérez, C. J. (2021). The impact of term-weighting schemes and similarity measures on extractive multi-document
text summarization. Expert Systems with Applications, 169(December 2020). https://doi.org/10.1016/j.eswa.2020.114510
Sankarasubramaniam, Y., Ramanathan, K., & Ghosh, S. (2014a). Text summarization using Wikipedia. Information Processing and Management, 50(3), 443–461.
https://doi.org/10.1016/j.ipm.2014.02.001
Sankarasubramaniam, Y., Ramanathan, K., & Ghosh, S. (2014b). Text summarization using Wikipedia. Information Processing and Management, 50(3), 443–461.
https://doi.org/10.1016/j.ipm.2014.02.001
See, A. V., (2021).Volume of data/information created, captured, copied, and consumed worldwide from 2010 to 2025.
https://www.statista.com/statistics/871513/worldwide-data-created/, accessed on 27 Jan 2022.
Shen, D., Yang, Q., & Chen, Z. (2007). Noise reduction through summarization for Web-page classification. Information Processing and Management, 43(6),
1735–1747. https://doi.org/10.1016/j.ipm.2007.01.013
Silber, H. G., & McCoy, K. F. (2002). Efficiently computed lexical chains as an intermediate representation for automatic text summarization. Computational
Linguistics, 28(4), 486–496. https://doi.org/10.1162/089120102762671954
Sinha, A., Yadav, A., & Gahlot, A. (n.d.). Extractive Text Summarization using Neural Networks.
Song, S., Huang, H., & Ruan, T. (2019a). Abstractive text summarization using LSTM-CNN based deep learning. Multimedia Tools and Applications, 78(1), 857–
875. https://doi.org/10.1007/s11042-018-5749-3
Song, S., Huang, H., & Ruan, T. (2019b). Abstractive text summarization using LSTM-CNN based deep learning. Multimedia Tools and Applications, 78(1), 857–
875. https://doi.org/10.1007/s11042-018-5749-3
Steinberger, J., & Ježek, K. (2009). Evaluation measures for text summarization. Computing and Informatics, 28(2), 251–275.
Van Lierde, H., & Chow, T. W. S. (2019). Query-oriented text summarization based on hypergraph transversals. Information Processing and Management, 56(4),
1317–1338. https://doi.org/10.1016/j.ipm.2019.03.003
Viégas, F. B., Golder, S., & Donath, J. (2006). Visualizing email content: Portraying relationships from conversational histories. Conference on Human Factors in
Computing Systems - Proceedings, 2, 979–988.
Wang, D., Liu, P., Zhong, M., Fu, J., Qiu, X., & Huang, X. (2019). Exploring domain shift in extractive text summarization. ArXiv.
WEI, B., REN, X., SUN, X., ZHANG, Y., CAI, X., & SU, Q. (2018). Regularizing output distribution of abstractive chinese social media text summarization for
improved semantic consistency. ArXiv, 5, 1–14.
Woodsend, K., Feng, Y., & Lapata, M. (2010). Title generation with quasi-synchronous grammar. EMNLP 2010 - Conference on Empirical Methods in Natural
Language Processing, Proceedings of the Conference, October, 513–523.
Xu, J., & Durrett, G. (2020). Neural extractive text summarization with syntactic compression. EMNLP-IJCNLP 2019 - 2019 Conference on Empirical Methods
in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Proceedings of the Conference, 3292–3303.
https://doi.org/10.18653/v1/d19-1324
Yang, M., Li, C., Shen, Y., Wu, Q., Zhao, Z., & Chen, X. (2020). Hierarchical Human-Like Deep Neural Networks for Abstractive Text Summarization. IEEE
Transactions on Neural Networks and Learning Systems, 1–14. https://doi.org/10.1109/tnnls.2020.3008037
Yang, M., Li, C., Shen, Y., Wu, Q., Zhao, Z., & Chen, X. (2021). Hierarchical Human-Like Deep Neural Networks for Abstractive Text Summarization. IEEE
Transactions on Neural Networks and Learning Systems, 32(6), 2744–2757. https://doi.org/10.1109/TNNLS.2020.3008037
Yang, M., Wang, X., Lu, Y., Lv, J., Shen, Y., & Li, C. (2020). Plausibility-promoting generative adversarial network for abstractive text summarization with multi-
task constraint. Information Sciences, 521, 46–61. https://doi.org/10.1016/j.ins.2020.02.040
Ye, S., Chua, T. S., Kan, M. Y., & Qiu, L. (2007a). Document concept lattice for text understanding and summarization. Information Processing and Management,
43(6), 1643–1662. https://doi.org/10.1016/j.ipm.2007.03.010
Ye, S., Chua, T. S., Kan, M. Y., & Qiu, L. (2007b). Document concept lattice for text understanding and summarization. Information Processing and Management,
43(6), 1643–1662. https://doi.org/10.1016/j.ipm.2007.03.010
Yeh, J. Y., Ke, H. R., Yang, W. P., & Meng, I. H. (2005a). Text summarization using a trainable summarizer and latent semantic analysis. 41(1), 75–95.
https://doi.org/10.1016/j.ipm.2004.04.003
Yeh, J. Y., Ke, H. R., Yang, W. P., & Meng, I. H. (2005b). Text summarization using a trainable summarizer and latent semantic analysis. Information Processing
and Management, 41(1), 75–95. https://doi.org/10.1016/j.ipm.2004.04.003
Yousefi-Azar, M., & Hamey, L. (2017a). Text summarization using unsupervised deep learning. Expert Systems with Applications, 68, 93–105.
https://doi.org/10.1016/j.eswa.2016.10.017
Yousefi-Azar, M., & Hamey, L. (2017b). Text summarization using unsupervised deep learning. Expert Systems with Applications, 68, 93–105.
https://doi.org/10.1016/j.eswa.2016.10.017
Zajic, D., Dorr, B., & Schwartz, R. (2004). BBN / UMD at DUC-2004 : Topiary. Proceedings of the HLT-NAACL 2004 Document Understanding Workshop,
Boston, 112--119.
Zhan, J., Loh, H. T., & Liu, Y. (2009). Gather customer concerns from online product reviews - A text summarization approach. Expert Systems with Applications,
36(2 PART 1), 2107–2115. https://doi.org/10.1016/j.eswa.2007.12.039
Zhang, Y., Li, D., Wang, Y., Fang, Y., & Xiao, W. (2019). Abstract text summarization with a convolutional seq2seq model. Applied Sciences (Switzerland), 9(8).
https://doi.org/10.3390/app9081665