Wasit Journal of Computer and Mathematic Science Vol. (1) No.

(4) (2022)

Natural Language Processing For Automatic text summarization

[Datasets] - Survey
https://doi.org/10.31185/wjcm.72

Alaa Ahmed AL-Banna()

Department of Computer Science, College of Science ,Al-Nahrain University, Baghdad, Iraq.
[email protected]

Abeer K AL-Mashhadany
Department of Computer Science, College of Science, Al-Nahrain University, Baghdad, Iraq.

Abstract— Natural language processing has developed significantly recently,

which has progressed the text summarization task. It is no longer limited to re-
ducing the text size or obtaining helpful information from a long document only.
It has begun to be used in getting answers from summarization, measuring the
quality of sentiment analysis systems, research and mining techniques, document
categorization, and natural language Inference, which increased the importance
of scientific research to get a good summary. This paper reviews the most used
datasets in text summarization in different languages and types, with the most
effective methods for each dataset. The results are shown using text summariza-
tion matrices. The review indicates that the pre-training models achieved the
highest results in the summary measures in most of the researchers' works for the
datasets. Dataset English made up about 75% of the databases available to re-
searchers due to the extensive use of the English language. Other languages such
as Arabic, Hindi, and others suffered from low resources of dataset sources,
which limited progress in the academic field.

Keywords— Natural Language Processing, Automatic Text Summarization,

Abstractive Text Summarization, Extractive Text Summarization, Text Summa-
rization Datasets

1 Introduction

Automatic summarization is a challenge in Natural Language Processing

(NLP) that involves developing algorithms that can reduce an input text, such as a sci-
entific journal article, to a compacted version containing just the relevant information.
The process of summarizing is complex because it involves not only the selection, as-
sessment, collection, and rearrangement of information but also compression, general-
ization, and paraphrasing [1]. There are several types of summarizations based on the
desired type of input and summary. The two major types of summarizations are Extrac-
tive and Abstractive summarizations [2]. In extractive summarization, text segments
such as sentences or phrases of the most important are extracted from the original text
without being modified words and concatenated to generate the summary [3].

159
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

Extractive summarization aims to identify the importance of each text sentence and
develop a shorter version of the original text that represents it accurately. On the other
hand, text summarization based on abstraction uses linguistic methods for understand-
ing and a deeper analysis of the text [4]. An abstract text summary type consists of new
sentences generated by paraphrasing or reformulating the extracted content [5]. Ab-
stractive summaries are usually resembling human-written ones because they tend to
represent the content and the meaning of the original text more naturally. Although this
type of summarization performs more efficiently, its implementation requires great
knowledge of deep learning techniques [6]. Also, the type of input documents used for
summarizing may vary. The target of summarizing may be to obtain a summary from
several documents or one document to obtain a summary text [7]. Single-document
summarizers deal with a single source text and generate summaries from it Inde-
pendently of other documents [8]. On the other hand, multi-document summarizing is
viewed as an extension of a single-document summary. It compiles many documents
on the same subject into a single summary. The multi-document summarization task is
more complex than summarizing a single document, even if it is lengthy. The difficulty
comes from handle within a large set of documents with thematic diversity [9].

1.1 Automatic Text Summarization Evaluation

The quality of the summarized text is evaluated by human assessment and using
Natural Language Text Summarization metrics. ROUGE, or" Recall-Oriented Under-
study for Gisting Evaluation," is the most popular matrices containing software pack-
ages for evaluating automatic summarization and translation in natural language pro-
cessing. By comparing an automatically produced summary with a reference or a group
of contacts (human-produced) [10]. The overlap of unigrams (per word) between the
system and reference summaries is referred to as ROUGE-1. In contrast, ROUGE-2
refers to the bigram overlap between the system and reference summaries. And
ROUGE-L: Statistics based on the "Longest Common Subsequence (LCS) " The long-
est common sub-sequence naturally considers sentence-level structural similarities and
automatically determines the longest co-occurring in sequence n-grams [11]. In addi-
tion to these matrices, human evaluation is adopted based on the correct linguistic rules
for the summary, the coherence of the text, etc. [12].

2 Overview of the Datasets

The dataset is the first step to getting a well-trained model to perform a specific task
in artificial intelligence [13]. When summarizing texts, it is necessary to look at the
available datasets, languages, types of functions, and the latest methods that led to ob-
taining a well-trained model. Below we review the most famous datasets that have been
worked with a summary table of information:-

160
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

1) Multi-Document Summarization on "Wikipedia Current Events Portal

(WCEP)"
The WCEP dataset for multi-document summarization comprises human-written
summaries of news events from the ("WCEP"), each linked with a cluster of news arti-
cles. These stories are based on sources mentioned by WCEP editors and automatically
collected from Common Crawl News and supplemented with items acquired automat-
ically from t (WCEP) [14]. The best result in the dataset shown in [15] is by using the
PRIMERA model by result Rouge-1=46.1 Rouge-2=25.2 Rouge-L=37.9. Multi-Docu-
ment Summarization captures useful information and filters out superfluous infor-
mation to summarize a series of documents. Extractive and abstractive multi-document
summarization are popular. Extractive summarizing systems extract significant sam-
ples, phrases, or sections from texts, whereas abstractive summarization systems para-
phrase the information [9].

2) DUC 2004 (Document Understanding Conferences)

The DUC2004 dataset is a multi-document summarization dataset used for testing.
It comprises 500 news stories, each with four human-written summaries. It comprises
50 clusters of "Text REtrieval Conference (TREC) " documents drawn from the fol-
lowing collections: "AP newswire, 1998-2000; New York Times newswire, 1998-
2000; and Xinhua News Agency (English version) "through 1996-2000. Each cluster
had an average of ten papers [16]. in [17] Summarization values got their highest levels:
ROUGE-1 = 28.18 ROUGE-2 = 8.49 ROUGE-L = 23.81.

3) CNN/Daily Mail
The "CNN/Daily Mail" dataset is used for summarizing text. Questions (with one of
the entities obscured) and associated portions from CNN and Daily Mail news articles
were manually constructed to train the algorithm to answer a fill-in-the-blank inquiry.
Authors have made available scripts to crawl, extract, and produce passage and question
pairs from various resources. The hands specify 11,487 test pairs, 13368 validation
pairs, and 286,817 training pairs. Documents in training average 766 words and 29.74
sentences in length, whereas their summaries only use 53 comments and 3.72 sentences
[18].in the latest studies in 2022 [19] using “Pegasus 2B + SLiC”, Y. Zhao and others
get the best result to text summarization task in this data set, Rouge-1= 47.97, Rouge-
2= 24.18, Rouge-L= 44.88.

4) PubMed ("Public/Publisher MEDLINE")

There are a total of 19717 studies on diabetes in the PubMed dataset, which may be
broken down into three categories. There are a total of 44338 references in the citation
web. A word vector from a vocabulary with 500 different terms is used to define each
article in the dataset according to its TF/IDF importance [20]. By using transformers
architectures [21] to get the best value to abstractive text summarization in ROUGE-1
=50.95, ROUGE-2= 21.93, and ROUGE-L= 45.61.

5) arXiv ("Arxiv HEP-TH (high energy physics theory) citation graph")

161
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

There are 27,770 publications included in the dataset that makes up the "Arxiv HEP-
TH" citation graph, and the total number of edges is 352,807. A directed edge connects
two nodes in the network if and only if paper I refers to paper J. There is no indication
in the graph of any papers outside the dataset being referenced or cited by any docu-
ments inside it. Specifically, the data includes articles published between January 1993
and April 2003. (124 months) [22]. the transformer was applied to summarize the text
in [21] and reached a ratio of ROUGE-1 =50.95, ROUGE-2 =21.93 ROUGE-L =45.61.
6) XSum (Extreme Summarization)
Extreme Summarizing (XSum) is a dataset for testing abstract single-document sum-
marization methods. The objective is to write a brand-fresh, catchy one-sentence sum-
mary that explains the article's subject. A one-sentence summary is provided for each
of the 226,711 news stories in the collection. Many different topics are covered in this
compilation of BBC stories that span from 2010 to 2017. (e.g., "News, Politics, Sports,
Weather, Business, Technology, Science, Health, Family, Education, Entertainment,
and Arts"). 204,045 (90%), 11,332 (5%) and 11,334 (5%) documents make up the of-
ficial training, validation, and test sets, respectively [23]. In [19], researchers work on
this dataset and introduce sequence likelihood calibration (SLiC). This makes Decoding
unnecessary, and the quality of decoding candidates rises dramatically independent of
the decoding method. Exceed SOTA results on various generation tasks, including ab-
stractive summarization, question creation, abstractive question answering, and data-
to-text generation, even with small models. Pegasus 2B +SLiC achieved result up to
ROUGE-1 =49.77, ROUGE-2 =27.09, and ROUGE-L =42.08.

7) MentSum ("Mental Health Summarization Dataset")

Mental health is still major one in the field of public health. Many people today are
turning to internet forums and social media to discuss their struggles with mental health,
vent their emotions, and connect with like-minded and trained professionals. The length
of the articles may vary, but it is helpful to offer a brief yet relevant description so that
the counselors can go through it quickly. MentSum includes over 24k hand-picked user
posts from Reddit and their short user-written summary (called TL; DR) in English
from 43 mental health subreddits to facilitate research into the summarization of online
posts related to mental health. This domain-specific dataset could be of interest not only
for generating short summaries on Reddit but also for generating summaries of posts
on dedicated mental health forums such as Reachout. in 2022[24]. This paper [25]works
on this dataset and evaluates extractive and abstractive state-of-the-art summarization
baselines [24]. Using the BART model, better values in this research are "Rouge-1
=29.13, Rouge-2 =7.98, and Rouge-L =20.27".

8) OrangeSum
OrangeSum is an extreme summarizing dataset that focuses on a single document to
summarize it. It includes two tasks, title and abstract. Title and abstract assignments
have an average ground truth summary length of 11.42 and 32.12 words, whereas 315-
and 350-word documents have similar average sizes. Creating a French language ver-
sion of the XSum dataset was the impetus behind OrangeSum. For models to do well

162
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

in OrangeSum, they need to be more abstract than they need to be on the historical
CNN, Daily Mail, and NY Times datasets. Extracting article titles and abstracts from
the orange Actu website led to the development of OrangeSum. Pages scraped span
over a decade, from February 2011 to September 2020, and may be roughly categorized
under five broad headings: France, the globe, politics, automobiles, and society. Health,
environment, people, culture, media, high technology, unusual ("insolite" in French),
and miscellaneous are the eight divisions of the society event [26] .in the 2021 paper
[27] provides "BARThez", the first large-scale pre-trained seq2seq model for French,
in this paper. "BARThez" is particularly well-suited for generative jobs because of its
BART foundation.BARThez is very competitive with cutting-edge BERT-based
French language versions such as CamemBERT and FlauBERT.and also proceed to
train a multilingual BART on BARThez' corpus and show that the resultant model,
mBARThez, greatly improves BARThez' generative performance.

9) BookSum (Books Summarization)

BookSum is a data set library for summarizing books and other lengthy texts. Novels,
dramas, and tales are some of the literary works included in this dataset, which also has
highly abstractive, human-written summaries at the paragraph, chapter, and book
levels. Longer texts, non-trivial causal, temporal relationships, and complex discourse
structures are only some of the obstacles that summarizing algorithms must overcome
in this dataset. So far, BookSum has summarized 142,753 paragraphs, 12,293 chapters,
and 436 volumes [28]. Using BART-LS in [29], the measurements ROUGE-1=38.5,
ROUGE-2=10.3, and ROUGE-L=36.4.

10) arXiv Summarization Dataset

For more than 30 years, ArXiv has served the public and scientific communities by
offering open access to scholarly publications in fields ranging from physics to com-
puter science and everything in between, including "math, statistics, electrical engineer-
ing, quantitative biology, and economics". This vast repository of data provides sub-
stantial but, at times, bewildering depth. This dataset is a free, open pipeline on Kaggle
to the machine-readable arXiv dataset: a library of 1.7 million articles containing im-
portant information such as article titles, authors, categories, abstracts, full-text PDFs,
and more, to help make the arXiv more accessible. To enable new use cases that can
lead to the exploration of more profitable machine learning techniques that combine
multi-modal features for applications such as trend analysis, paper recommender en-
gines, category prediction, co-citation networks, knowledge graph construction, and
semantic search interfaces. ArXiv is a community-supported, jointly financed resource
created by Paul Ginsparg in 1991 and maintained and controlled by Cornell University
[30]. Savelieva and others In [31], using Facet-Aware Modeling Improving Unsuper-
vised Extractive Summarization, the best result is ROGUE-1=40.92 ROUGE-2=13.75
ROUGE-L =35.56.
11) WikiHow Dataset
"WikiHow" is a dataset of over 230,000 article-summary pairs mined from a wiki-
based knowledge repository and created by various human writers. The topics covered

163
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

and the writing styles represented in these pieces have high diversity [32]. Using state-
of-the-art NLP in [33] makes abstractive summaries of "recorded instructional films
ranging from gardening and cooking to software configuration and sports". The model
is pre-trained on a few large cross-domain datasets in written and spoken English using
transfer learning. ROUGE measurement is used to evaluate, and this paper has achieved
the best results and used BertSum by relying on this dataset, as it was ROUGE-1
=35.91, ROUGE-2 =13.9, ROUGE-L =34.82.
12) Urdu News Dataset
Over a million news articles from the fields of business, economics, science and
technology, entertainment, and sports are included in this collection. This dataset is
helpful for numerous Urdu NLP applications since its four unique categories were care-
fully selected to eliminate ambiguity. For many NLP, Machine/Deep Learning tasks,
including "text processing, classification, summarization, named entity recognition,
topic modeling, and text generation", the dataset A Large-Scale News Dataset for Urdu
Text Processing is the only dataset in the Urdu language that is currently available—
created this dataset.in 2021 and so far, no studies have been done on this dataset in text
summarization [34].
13) Bengali News Articles ("IndicNLP")
Since the previous several decades, "natural language processing (NLP)" has been
used extensively in studying Western languages, especially the English language. Lan-
guage processing research on the eastern counterpart, particularly the languages of the
Indian subcontinent, needs to be increased. Western languages have access to a wealth
of dictionaries, WordNet, and related resources. This data collection, which has been
cleaned and comes with a train and test set to compare your classification and summa-
rization models against, contains 14k news items. A collection of Bengali news items
make up this dataset. It may be applied to problems involving classification and lan-
guage modeling [35]. Although there has been a substantial amount of critical study on
abstractive summary in English, just a few works have been done on Bengali abstractive
news summarization. Paper [36] described a seq2seq-based Long Short-Term Memory
(LSTM) network architecture focused on the encoder-decoder. The suggested system
uses the attention-based model to construct a long sequence of words. The summary
was evaluated subjectively and statistically, and its results were compared to other pub-
lished results. Mechanism of attention demonstrated a considerable improvement in
state-of-the-art human assessment ratings.

14) Hindi Text Short and Large Summarization Corpus

The "Hindi Text Short and Large Summarization Corpus" is a collection of 180k
articles with headlines and summaries taken from Hindi news websites. This is a first-
of-its-kind Hindi Dataset that may be used to benchmark algorithms for Hindi text sum-
marization. This does not include articles from this dataset, which is being released
concurrently with this Dataset. The dataset preserves the articles' original punctuation,
numerals, and other formatting [37]. In 2022 [38] executed the Hindi text summariza-
tion, which had received relatively little attention. A machine learning model has been
developed and evaluated using around 100,000 data samples, resulting in highly

164
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

accurate summaries that benefit society. The model has an F-Score of 58% and a Rouge
Score of 67.5%. Pandas, NumPy, sklearn, and other libraries are utilized.LSTM, word
embedding and seq2seq are used to train the data model.

15) Arabic News articles from Aljazeera.net

Natural Language Processing (often known as NLP) is a highly researched area of
machine learning. Recent years have seen significant development, paving the way for
this field's expansion into widespread use across various contexts. Today, natural lan-
guage processing (NLP) is used in several contexts, including but not limited to: social
media platforms, search engines, translation apps, Chabot helpers, and many more. Pro-
gress and outcomes, however, vary from language to language. Most ML systems only
care about English and ignore other languages, particularly Arabic. A primary cause of
this is the scarcity of relevant datasets. This data collection comprises around 5870
Arabic-language news stories extracted from the aljazeera.net website [39]. An abstrac-
tive Arabic text summarization model based on RNN is proposed and used in this da-
taset in this study [40]. A multilayer encoder and a single-layer decoder are included.
Encoder layers use bidirectional long short-term memory, whereas decoder levels use
unidirectional long short-term memory. The evaluation shows that the model produced
is the best, achieving ROUGE1=38.4.

16) COVID-19 Open Research Dataset Challenge (CORD-19)

In response to COVID-19, the White House and prominent academic organizations
created the COVID-19 Open Research Dataset (CORD-19). CORD-19 has almost
1,000,000 academic publications on COVID-19, SARS-CoV-2, and similar corona-
viruses, including over 400,000 in full text. This free dataset is supplied to the world-
wide research community to utilize natural language processing, and other AI ap-
proaches to develop new insights in the battle against this deadly illness. The tremen-
dous increase in coronavirus literature makes it challenging for medical researchers to
stay up [41]. In 2022 paper [42] proposes a hybrid, unsupervised, abstractive approach
that walks through a document, generating salient textual fragments representing its
key points. We then select the most important sentences of the paper by choosing the
most similar sentences to the generated texts, calculated using BERTScore.and this
method achieved ROUGE-1 =41.02, ROUGE-2=13.79, ROUGE-L =37.25, in the text
summarization measures’, which are considered the best among the works submitted
on this dataset.

17) Scientific Document Summarization (SciTLDR)

A new multi-target dataset of 5.4K TLDRs over 3.2K papers. SciTLDR contains
both author-written and expert-derived TLDRs in computer science. The latter are col-
lected using a novel annotation protocol that produces high-quality summaries while
minimizing the annotation burden [43]. In this dataset train =1992, valid =618,
test=619.The first paper published in this data set is [44] introducing TLDR generation
for scientific papers and releasing SCITLDR, a multi-target dataset of TLDR-paper
pairs. The result is RPOGE-1=43.8, ROUGE-2=20.9 ROUGE-L=35.5 Using pre-
trained models. This data set is considered one of the difficult types due to the small

165
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

number of samples and the difficulty of training them, which is why researchers often
resort to pre-training and fine-tuning methods.

18) ScisummNet Corpus

This massive corpus may be used to train citation-based summarization algorithms
for scientific papers, opening new avenues of inquiry into supervised approaches. This
dataset includes 1,000 samples. Has been organized since 2014 in computational lin-
guistics and NLP papers[45]. Group of researchers in [46] using BertSum, Continual
BERT, Adapter-based BERT, and SummaRuNNer in PubMed and ScisummNet da-
tasets. Bertrum outperforms other pre-trained models used on the ScisummNet dataset,
archives result to ROUGE-1= 33.0, ROUGE-2 = 13.4, ROUGE-L= 31.6, but PubMed
is better In terms of assessment scores because of phrases use of complex and special-
ized medical terminologies rather than ScisummNet's general scientific phrases.

19) SUMARABIC
SumArabic is a dataset for abstractive text summarization in Arabic. The infor-
mation comes from two Arabic news websites: emaratalyoum.com and ema-
ratalyoum.com. - www.almamlakatv.com.The data is divided into four sets: training,
testing, validation, and out-of-domain. Each split has the following examples: 75,817
training, 4,121 validations, and 4,174 tests. 652 out-of-domain, Total: 84,764 [47].This
dataset is one of the latest additions to 2022 in the Arabic language, which does not
contain academic studies on it yet.

20) TalkSumm
TalkSumm is a dataset that contains 1705 automatically-generated summaries pa-
pers in science from ACL, NAACL, EMNLP, SIGDIAL (2015-2018), and ICML
(2017-2018). The dataset contains titles, URLs, and corresponding summaries.
[48]This study presents a unique way for automatically producing summaries for sci-
entific publications based on recordings of lectures at scientific conferences, suggesting
that such presentations provide a cohesive and short summary of the paper's content
and can serve as the foundation for effective summaries. It compiled a collection of
paper summaries from 1716 publications and their accompanying videos. A model
trained on this dataset outperforms models trained on a manually constructed dataset of
summaries. Furthermore, human specialists confirmed the quality of our this summary
[49].

166
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

Table 1. Summarizes the basic information of the dataset used to text summarization from
Language of dataset ،content ،Input document Single or Multiple and Referring to the
best research and methods that achieved the highest percentages in summary matrices

Dataset Da- Da- Input Output Author Ref. Result

Name taset taset Type types /Methods
Lan- Field Sin- Ab-
guage gle/ strac-
Multi- tive
ple /Ex-
tractive
1. Multi- Eng- news Multi Extrac- W. Xiao et ROUGE-
Document lish articles docu- tive al.[14] 1=46.1
Summari- ments \ Primer ROUGE-
zation on 2=25.2
WCEP ROUGE-
[14] L=37.9
2. DUC Eng- news Multi Ex- S. Shen et ROUGE1=
2004 [16]. lish articles docu- treme/ al.[17] 28.18
ments Ab- \ wise sen- ROUGE-
stractive tence 2=8.49
ROUGEL=
23.81
3. Eng- news Single Ab- Y. Zhao et ROUGE-
CNN/Dail lish articles docu- stractive al.[19] 1=47.97
y Mail [18] ment \Pegasus 2B ROUGE-
+SLiC 2=24.18
ROUGE-L
44.88
4. Pub- Eng- scien- Single Ab- B. Pang et al. ROUGE-
Med[20] lish tific docu- stractive [21] 1=50.95
ment \Trans- ROUGE-
former 2=21.93
ROUGE-
L=45.61
5.arXiv Eng- scien- Single Ab- B. Pang et al. ROUGE-
("Arxiv lish tific docu- stractive [21] 1=50.95
HEP-TH ment \Trans- ROUGE-
(high en- former 2=21.93
ergy phys- ROUGE-
ics theory) L=45.61
citation

167
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

graph) "
[22]
6. XSum Eng- News Single Ex- Y. Z o et ROUGE-
[23] lish ("Poli- docu- treme/ al.[19] 1=49.77
tics, ment Ab- \Pegasus 2B ROUGE-
Sports, stractive +SLiC 2=27.09
Weathe ROUGE-
r, Busi- L=42.08
ness,
Tech-
nology,
Sci-
ence,
Health,
Family,
Educa-
tion,
and
Arts")
7.MentSu Eng- Mental Single Ex- S. Sotudeh et ROUGE-
m ("Men- lish health docu- treme/ al.[25] 1=29.13
tal Health ment Ab- \ BART ROUGE-
Summari- stractive 2=7.98
zation Da- ROUGE-
taset") [24] L=20.27
8. Or- French society Single Ab- M. K. Ed- ROUGE-
angeSum docu- stractive dine et al. 1=15.49
[26] ment [27]\BARTh ROUGE-
ez' 2=5.82
ROUGE-
L=13.05
9.Book- Eng- literary Single Ab- W. Xiong et ROUGE-
Sum [28] lish works docu- stractive al. [29] \ 1=38.5
ment BART-LS ROUGE-
2=10.3
ROUGE-
L=36.4
10.arXiv Eng- scien- Single Extrac- Liang et al. ROGUE-
Summari- lish tific docu- tive [31] 1=40.92
zation Da- ments \Facet- ROUGE-
taset [30] Aware Mod- 2=13.75
eling ROUGE-
L=35.56

168
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

11.Wiki- Eng- article Single Ab- Savelieva et ROUGE-1

How [32] lish docu- stractive al. [33] \ =35.91
ments BertSum ROUGE-2
=13.9
ROUGE-L
=34.82
12. Urdu Urdu news single - - -
News Da- articles docu-
taset [34] ments
13. Ben- Ben- news Single Ab- Bhattachar- Humans
gali News gali articles docu- stractive jee et al.[ 36] Evaluation
Dataset ments \ LSTM
[35]
14. Hindi Hindi Hindi Single Ab- Shah et F-Score
Text Short news docu- stractive al.[38] \ =58%
and Large ments LSTM+ Rouge
Summari- word embed- =67.5 %
zation ding
Corpus
[37]
15. Arabic Arabic News Single Ab- Suleiman et ROUGE-
News arti- articles docu- stractive al .[40]\ RN 1=38.4
cles from ments Ns
Aljazeera.
net [39]
16. Eng- Scien- Single Ab- Bishop et al. ROUGE-
COVID- lish tific docu- stractive [42] \ Gen- 1= 41.02
19 Open ments Com- ROUGE-2
Research pareSum = 13.7
Dataset ROUGE-
Challenge L= 37.25
(CORD-
19) [41]
17.Scien- Eng- Scien- Single Ex- Cachola et -
tific Docu- lish tific docu- treme/ al.[44] \
ment Sum- ments Ab- CATTS
marization stractive
("SciTLD
R") [43]
18.Scisum Eng- Scien- Single Ab- Park et al. ROUGE-
mNet Cor- lish tific docu- stractive [46]\ 1= 33.0
pus [45] ments BertSum ROUGE-2
= 13.4
ROUGE-
L= 31.6

169
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

19. Arabic News Single Ab- - -

SumAr- docu- stractive
abic [47] ments
20. Eng- Scien- Single Ab- Lev et -
Talk- lish tific docu- stractive al.[49] \
Summ ments
[48]

3 Conclusion

This review concludes that the dataset's field in summarizing texts is focused on
news and scientific aspects, especially the abstract type. The available datasets lack
diversity in topics, primarily literary, artistic, and some fields of science. The dataset
available in a large size can be used to train the model, fine-tune a small emerging data
set, and get good results. After the significant development in text summarization tech-
niques, getting a good summary has become the beginning of other tasks related to
NLP.

4 References

[1] Suleiman, Dima, and Arafat Awajan. "Deep learning based abstractive text summarization:
approaches, datasets, evaluation measures, and challenges." Mathematical problems in en-
gineering 2020 (2020).
[2] Munot, Nikita, and Sharvari S. Govilkar. "Comparative study of text summarization meth-
ods." International Journal of Computer Applications 102, no. 12, 2014.
[3] V. Gupta and G. S. Lehal, "A survey of text summarization extractive techniques," Journal
of emerging technologies in web intelligence, vol. 2, no. 3, pp. 258-268, 2010.
[4] Sinha, Aakash, Abhishek Yadav, and Akshay Gahlot. "Extractive text summarization using
neural networks." arXiv preprint arXiv: 1802.10137, 2018.
[5] W.-T. Hsu, C.-K. Lin, M.-Y. Lee, K. Min, J. Tang, and M. Sun, "A unified model for ex-
tractive and abstractive summarization using inconsistency loss," arXiv preprint
arXiv:1805.06266, 2018.
[6] Nallapati, Ramesh, Bowen Zhou, Caglar Gulcehre, and Bing Xiang. "Abstractive text sum-
marization using sequence-to-sequence rnns and beyond." arXiv preprint arXiv:1602.06023,
2016.
[7] [C.-Y. Lin and E. Hovy, "From single to multi-document summarization," in Proceedings
of the 40th annual meeting of the association for computational linguistics, 2002, pp. 457-
464.
[8] Christian, Hans, Mikhael Pramodana Agus, and Derwin Suhartono. "Single document auto-
matic text summarization using term frequency-inverse document frequency (TF-IDF)."
ComTech: Computer, Mathematics and Engineering Applications 7, no. 4 (2016): 285-294.
[9] Mutlu, Begum, Ebru A. Sezer, and M. Ali Akcayol. "Multi-document extractive text sum-
marization: A comparative assessment on features." Knowledge-Based Systems 183 (2019):
104848.

170
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

[10] Lin, Chin-Yew. "Rouge: A package for automatic evaluation of summaries." In Text sum-
marization branches out, pp. 74-81. 2004.
[11] Ozsoy, Makbule Gulcin, Ferda Nur Alpaslan, and Ilyas Cicekli. "Text summarization using
latent semantic analysis." Journal of Information Science 37, no. 4 (2011): 405-417.
[12] Gong, Yihong, and Xin Liu. "Generic text summarization using relevance measure and la-
tent semantic analysis." In Proceedings of the 24th annual international ACM SIGIR con-
ference on Research and development in information retrieval, pp. 19-25. 2001.
[13] Wissner-Gross, A. .” Edge.com. Retrieved 8 January 2016.
[14] Demian Gholipour Ghalandari; Chris Hokamp; Nghia The Pham; John Glover; Georgiana
Ifrim, WCEP Dataset [Dataset]. https://paperswithcode.com/dataset/wcep
[15] W. Xiao, I. Beltagy, G. Carenini, and A. Cohan, "Primer: Pyramid-based masked sentence
pre-training for multi-document summarization," arXiv preprint ar X iv: 2110.08499, 2021.
[16] DUC 2004 Dataset [Dataset]. https://paperswithcode.com/dataset/duc-2004
[17] S. Shen, Y. Zhao, Z. Liu, and M. Sun, "Neural headline generation with sentence-wise op-
timization," arXiv preprint arXiv:1604.01904, 2016.
[18] Ramesh Nallapati; Bo-Wen Zhou; Cicero Nogueira dos santos; Caglar Gulcehre; Bing
Xiang, CNN/Daily Mail Dataset [Dataset]. https://paperswithcode.com/dataset/cnn-daily-
mail-1.
[19] Y. Zhao, M. Khalman, R. Joshi, S. Narayan, M. Saleh, and P. J. Liu, "Calibrating Sequence
likelihood Improves Conditional Language Generation," arXiv preprint ar X iv: 2210.00045,
2022.
[20] datadiscovery.nlm.nih.gov (2021). PubMed [Dataset]. https://healthdata.gov/dataset/Pub-
Med/h5mw-dwr6 .
[21] B. Pang, E. Nijkamp, W. Kryściński, S. Savarese, Y. Zhou, and C. Xiong, "Long Document
Summarization with Top-down and Bottom-up Inference," arXiv preprint
arXiv:2203.07586, 2022.
[22] ArXiv Dataset [Dataset]. https://paperswithcode.com/dataset/arxiv.
[23] Shashi Narayan; Shay B. Cohen; Mirella Lapata, XSum Dataset [Dataset]. https://pa-
perswithcode.com/dataset/xsum
[24] Sajad Sotudeh; Nazli Goharian; Zachary Young (2022). Mental Health Summarization
(MentSum) [Dataset]. https://ir.cs.georgetown.edu/resources/mentsum.html.
[25] S. Sotudeh, N. Goharian, and Z. Young, "MentSum: A Resource for Exploring Summariza-
tion of Mental Health Online Posts," arXiv preprint arXiv:2206.00856, 2022.
[26] Moussa Kamal Eddine; Antoine J. -P. Tixier; Michalis Vazirgiannis, OrangeSum Dataset
[Dataset]. https://paperswithcode.com/dataset/orangesum.
[27] M. K. Eddine, A. J.-P. Tixier, and M. Vazirgiannis, "Barthez: a skilled pretrained french
sequence-to-sequence model," arXiv preprint arXiv:2010.12321, 2020.
[28] Wojciech Kryściński; Nazneen Rajani; Divyansh Agarwal; Caiming Xiong; Dragomir
Radev, BookSum Dataset [Dataset]. https://paperswithcode.com/dataset/booksum .
[29] W. Xiong, A. Gupta, S. Toshniwal, Y. Mehdad, and W.-t. Yih, "Adapting Pretrained Text-
to-Text Models for Long Text Sequences," arXiv preprint arXiv: 2209.10052, 2022.
[30] Arman Cohan; Franck Dernoncourt; Doo Soon Kim; Trung Bui; Seokhwan Kim; Walter
Chang; Nazli Goharian (2021). arXiv Summarization Dataset Dataset [Dataset]. https://pa-
perswithcode.com/dataset/arxiv-summarization-dataset.
[31] Liang, Xinnian, Shuangzhi Wu, Mu Li, and Zhoujun Li. "Improving unsupervised extractive
summarization with facet-aware modeling." The Association for Computational Linguistics
Findings: ACL-IJCNLP 2021, pp. 1685-1697. 2021.
[32] Mahnaz Koupaee; William Yang Wang (2021). WikiHow Dataset [Dataset]. https://pa-
perswithcode.com/dataset/wikihow

171
Wasit Journal of Computer and Mathematic Science Vol. (1) No. (4) (2022)

[33] Savelieva, Alexandra, Bryan Au-Yeung, and Vasanth Ramani. "Abstractive summarization
of spoken and written instructions with BERT." arXiv preprint arXiv:2008.09676 (2020).
[34] Saurabh Shahane (2021). Urdu News Dataset [Dataset].
https://www.kaggle.com/saurabhshahane/urdu-news-dataset
[35] PrithwirajSust (2020). Bengali News Summarization Dataset [Dataset].
https://www.kaggle.com/prithwirajsust/bengali-news-summarization-dataset
[36] Bhattacharjee, Prithwiraj, Avi Mallick, and Saiful Islam. "Bengali abstractive news sum-
marization (BANS): a neural attention approach." In Proceedings of International Confer-
ence on Trends in Computational and Cognitive Engineering, pp. 41-51. Springer, Singa-
pore, 2021.
[37] Gaurav (2020). Hindi Text Short Summarization Corpus [Dataset].
https://www.kaggle.com/disisbig/hindi-text-short-summarization-corpus
[38] Shah, Aashil, Devam Zanzmera, and Kevan Mehta. "Deep Learning based Automatic Hindi
Text Summarization." In 2022 6th International Conference on Computing Methodologies
and Communication (ICCMC), pp. 1455-1461. IEEE, 2022.
[39] Abdelkader Rhouati (2020). Arabic News articles from Aljazeera.net [Dataset].
https://www.kaggle.com/arhouati/arabic-news-articles-from-aljazeeranet .
[40] Suleiman, Dima, and Arafat Awajan. "Multilayer encoder and single-layer decoder for ab-
stractive Arabic text summarization." Knowledge-Based Systems 237 (2022): 107791.
[41] Allen Institute for AI (2020). cord-19 [Dataset]. https://allenai.org/data/cord-19 .
[42] Bishop, Jennifer, Qianqian Xie, and Sophia Ananiadou. "GenCompareSum: a hybrid unsu-
pervised summarization method using salience." In Proceedings of the 21st Workshop on
Biomedical Language Processing, pp. 220-240. 2022.
[43] Aaditya Raj (2022). Scientific Document Summarization (SciTLDR-A) [Dataset].
https://www.kaggle.com/datasets/adityawithdoublea/scitldra .
[44] Cachola, Isabel, Kyle Lo, Arman Cohan, and Daniel S. Weld. "TLDR: Extreme summari-
zation of scientific documents." arXiv preprint arXiv:2004.15011, 2020.
[45] Michihiro Yasunaga; Jungo Kasai; Rui Zhang; Alexander R. Fabbri; Irene Li; Dan Fried-
man; Dragomir R. Radev (2021). ScisummNet Dataset [Dataset]. https://paperswith-
code.com/dataset/scisummnet .
[46] Park, Jong Won. "Continual bert: Continual learning for adaptive extractive summarization
of covid-19 literature." arXiv preprint arXiv:2007.03405, 2020.
[47] Mohammad Bani Almarjeh (2022). SumArabic [Dataset].
http://doi.org/10.17632/7kr75c9h24.1 .
[48] Guy Lev; Michal Shmueli-Scheuer; Jonathan Herzig; Achiya Jerbi; David Konopnicki,
TalkSumm Dataset [Dataset]. https://paperswithcode.com/dataset/talksumm .
[49] Lev, Guy, Michal Shmueli-Scheuer, Jonathan Herzig, Achiya Jerbi, and David Konopnicki.
"Talksumm: A dataset and scalable annotation method for scientific paper summarization
based on conference talks." arXiv preprint arXiv:1906.01351 , 2019.

Article submitted 23 October 2022. Published as resubmitted by the authors 31 December 2022.

172