=Paper=
{{Paper
|id=Vol-2823/Paper17
|storemode=property
|title=Context-Aware Model of Abstractive Text Summarization for Research Articles
|pdfUrl=https://ceur-ws.org/Vol-2823/Paper17.pdf
|volume=Vol-2823
|authors=Gopinath Dineshnath, Selvaraj Saraswathi
}}
==Context-Aware Model of Abstractive Text Summarization for Research Articles==
https://ceur-ws.org/Vol-2823/Paper17.pdf
Context-Aware Model of Abstractive Text Summarization for
Research Articles
Gopinath Dineshnatha, Selvaraj.Saraswathib
a
Department of Computer Science Engineering, Pondicherry Engineering College, Pillaichavady,
Puducherry, 605014, India.
b
Department of Information Technology, Pondicherry Engineering College, Pillaichavady,
Puducherry, 605014, India.
Abstract
Research article comprises of different sections each holds its own characteristic domain information.
Summarization of entire article from multiple documents of multiple sections in precise form with
special focus to contextual information is tedious. We proposed context-aware model to extract
contextual texts from research article by utilizing multi-document directed graph for contextual
matching phrases We customized extractive summarization for abstractive text summarization with
lucid-information as prime criterion. Decision matrix with elitism identification further fine tunes the
abstractive text summary and outperforms at sentence level Rouge-L measures 9.32 and summary level
measures 89.65.
Keywords1
Glowworm Swarm Optimization, Analytical Hierarchical Processing (AHP), Skip Gram Model,
Word2vectors, Maximum Marginal Relevance, Log-Likelihood Rate,N-Gram.
1. Introduction in two ways; single-document summaries
produce a summary from a given single source
and multi-document summaries in which
Modern days, internet becomes the integral part
different but related documents are summarized
of human and act as information highway. The
by comprises only the vital materials or main
primary source of information in digital world
ideas in a document in less space.
is Internet and it is boon for academicians,
There is a vast difference between automatic
bloggers, students and researcher fraternity.
multi-document summarization of generic texts
Information available in Internet comprises of
to that of scientific articles. The major
massive flow of information, which makes
difference [2] between generic-text and research
retrieval process complex with respect to
article summarization is; research article
context-specific content. Scientific article
consists of different section namely abstract,
prevailing now with ocean of research domains
introduction, literature survey, methods, results
makes difficult to scholar cope-up, grasp and
and discussions, whereas generic text’s scope is
streamline documents relevant to their interest.
extracted from first few sentences in first
Query based search [1] for specific domain also
paragraphs and entire section holds at
fetch many relevant articles that is difficult task
maximum 500 words.
to categorize surpass human processing
In general, abstract and citation texts in
capabilities. In such scenario, automatic text
scientific articles are considered for automatic
summarization of articles is fruitful solution in
summarization system.
terms of reducing time effort for reviewing
Abstract section[3] is biased to author
entire articles and grab gist of information
findings, author’s own contribution, and
enclosed in it. Basically, summaries generation
evaluation metrics. In simpler way, abstract
ACI’21: Workshop on Advances in Computational Intelligence
at ISIC 2021, February 25-27,2021, Delhi, India.
Email: dinesh.gopinath60@gmail.com(G.Dineshnath);
swathi@pec.edu (S.Saraswathi);
:0000-0003-0026-1932(A.1)
©️2021 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
outlines the domain and list of findings in crisp techniques. Our main focus is Context aware
manner depends upon the type of articles information inference from multiple
(review/original). documents.
Citation sensitive or citation-based summaries
[4]
is another type of Scientific article 2. Literature review
summarization, major task in summary
production is clear-cut distinction of cited and
Lloret et al. [7] have applied both extractive and
non-cited text is performed. Citation summary
abstractive summarization procedures for
system performs categorization of every
sentence and labeled to citation or non-citation scientific article abstracts. The extractive
one. Later, evaluation measures based on summarizer (compendium E) is developed to
similarity between each sentence in the perform conventional preprocessing such as
reference article and citation sentences and then breaking sentences, assigning tokens,
grouped it into one of two classes: cited or non- stemming, lemmatization and PNG markers,
cited. Abstractive Multi document tagging and removing duplicates at various
sentence levels. [8]A mixture of both extractive
summarization [5] selectively picks up either
and abstractive technique (compendium E−A)
first sentence of abstract or introduction of a
is developed to support compendium E as base
paper since it is comprised of background
to incorporates sorted information which are
information of research topic. Construction of
appropriate title for article involves relevant. Relevancy identification with respect
interpretation and integration of the concepts to every sentences, assigns a score that
from multiple sentences of the abstract. Apart emphasize its importance based on code
from that there exist multiple challenging issues quantity principle (CQP)[9] compendium E−A
like content organization, sentence derives abstractive summary by utilizing top
compression and fusion, and paraphrasing ranked sentences with chronological ordering.
Saggion [11] utilized pretrain models
sentences.
All summarization system should meet the for learning and transformations for the
summary length constraints as and other problem of abstract generation. The initial
parameters specified by the user or summary generation from abstracts are
generated and transformed to model based
summarization system is known as controllable
learning. The learning models assists with
summarization. Controllable summarization [6]
is the main criteria in summarization system examples from corpus. Further, abstracts are
which specifies the length of summary gathered from GATE [12]and Weka [13]
generation in accordance to the entities on environment. Abstractive text summarization
which it focuses on and mimics the source’s also known as natural language generation in
natural language processing Paraphrasing of
style. User may define the high-level attributes
for summary generation. Summary generation sentences is also another important criterion in
is controlled by specific control variables, natural language generation Paraphrasing of
length, and source style, entities of interest and sentences [14] involves substitution of relevant
summarizing only remaining portions of the verbatim and modifying the direct to indirect
document. For instance, blog summarization, speech or vice versa. The vector representation
the primary thing is to derive representative is purely focus on various sources of features
words from comments and then selection of namely LDA, D2V, W2V and encoding
schemes[15]..LDA[16] explores semantic
paramount sentences from the blog post which
consists of representative words. associations, D2V vectors finds contextual
Context aware components are usually word vectors along with documents. Contextual
meant task of inferring contextual information. aware model phase is concerned with
Contextual information detection might be contextual theme and dependency phrase
detection ranging from topic community, extraction from multi documents using directed
paragraphs analysis, sentences and words by graph. minimal spanning tree is constructed for
statistical computation measures. The most edges algorithm using Chu Liu Edmonds.
well-known computation is Set-Calculus (CLE)[17] Knowledge base or ontology- based
techniques to find sentence similarity via context aware model useful for domain
Union, Jaccard Co-efficient, Cosine -Similarity classification and annotation.[18]Evaluation of
document summarization, Document
measures followed by normalization
Understanding conference (DUC)[19] is used to slide over word vectors for prediction
benchmark datasets is used generally. Various of context vectors. Skip gram models skips the
datasets such as TAC[20],shared task for text selection of common words rather than
processing and document summarization contextual words. N gram models hold
Similarly, DPIL[21] for paraphrasing Indian predefined size ‘n’ that triggers the selection of
languages for text summarization. contextual words upon size limited to ‘n’. Both
skip gram and N gram models are desirable
notion to pick keywords in context based on
W(Vi) = (1 – d) + d ∗ their lexicography collocations. The Proposed
overlap(si ,sj ) (1)
∑Vj∈ln(Vi) W(Vj) diagram is given in figure-1.
∑ Vk∈Out(Vj )overlap(sj ,sk)
.
The edge weights for nodes are assigned
using status score on the basis of inwards and
outwards edges, W(Vi) represents status score
assigned to vertex Vi. In (Vi) and Out (Vj) are
inwards and outwards edges, points from
particular node. After several iterations, each
sentence in the document is assigned with a
score. The top-n sentences are selected and
ranked which constructs the summary for the
document. There exist some dependent phrases
in graph. The dependent phrases are cyclic in
nature and some disjoint nodes. Such, disjoint
nodes in graph are connected using CLE
algorithm. On basis of lexical contexts,
proposed procedure performs well than
traditional keyword-based algorithms. We
enhance the extractive summary production by
adding co-occurrence measures to ensure
Concept-based Point wise Mutual Information.
(CPMI) CPMI weighs the different section in
paragraphs gradually it weights decreases from
beginning of paragraph to end of paragraph in
document. CPMI measures support
distributional semantics among phrases. CPMI
weightage scheme H(ƥ) is expressed in
equation 2.
H(ƥ)
log B (2)
C − positive constant ƥ ∗ B, if ƥ < − ( )
log C
= ,
{ 1 otherwise }
Figure 1: Proposed Context aware model for
Abstractive Summarization
3. Context Aware Model Phase
Context aware component model proposed 3.1. Contextual matching phrases
follows contextual keyword interpretation,
topic detection and topic clusters formation.
Contextual matching phrase from multi
Context aware component model also
documents is proposed to retrieve thematic
determines vector manipulation by bag of
portions with similar sentence phrases
words and skip gram key terms with respect to
sequentially from one document to another. The
specific documents. One hot encoding scheme
graph based contextual word is intuitive way to
Formula for Extraction of Eq customized to produce extractive to abstractive
Features. n.o: summary.in Table-1. The optimal features and
Number of title Features (3) their convergence in Adaptive Glowworm
𝑇𝑖 =
average length(Title) Optimization (AGWO) is discussed in section
1 − e –α (4)
Tag sum = 3.4.
1 + eα
∑L(Si) (5)
(Sbi) =
Lmax Table 1 Optimal Features Extraction.
TW(𝑆𝑖) (6)
Dsm(𝑆𝑖) =
𝑇𝑀𝑎𝑥
represent sentences as node and corresponding
contextual word as vertices. The path projecting Features Glowworm
from node to node via outcoming and incoming optimization
vertices provides a notion either matching Title feature Luciferin update
target phrases or discriminating phrases. phase
subsequent dependent phrases need to be Named entity Movement phase
included in directed graph. The procedure for recognition and
contextual theme and dependency phrases Tagging
extraction is shown below. Sentence boundary Neighborhood Phase
Step 1 Accept text and store it in text buffer. Distributional Words and conceptual
Step 2 new word falls into below categories.
Semantic Analysis level
Step 2 a): If first word then adds to graph G.
Step 2 b): If fresh word then appends to G.
Step 3 Go to step 2 until words overlap. 3.3. Vector Formulation
Step 4 If overlap, status score using (1).
Step 5 Extract similar texts and update in G. Word embeddings are feature vectors that
Step 6 Construct Digraph using CLE. represent words holds the property that similar
Step 7 Do updates to infer adjacent edges. words have similar feature vectors. The
Step 8: Output the phrases. question might rise in mind where the
embeddings come from. The response is
3.2. Feature Extraction (again): they are learned from data. Several
algorithms exist to learn word embeddings. We
The feature extraction for document consider only one of them: word2vec, and sole
summarization includes title feature, proper version of word2vec called skip-gram, which is
nouns, Named Entities Recognition (NER) and well-known and currently utilized in practice.
parts of speech tagging, sentence boundary Word embedding learning, our goal is to build
analysis and distributional semantic analysis. a model which we can use to convert a one-hot
Title feature scoring scheme is based on the encoding of a word into a word embedding. Let
ratio of mean number of titles present to that of our dictionary contain 10,000 words or Giga-
average length of title. The formula for title word Corpus. Skip-gram model performs for
feature is expressed in equation. (3). The proper given sentence, selection of a word is feed into
nouns are generally recognized as title words classifier, and predict words before and after the
and minimum number of words to accept as selected word in a fixed window. Negative
title. NER marks or labels the salient sentences sampling provides better vectors for frequent
which is considered for summary. The scoring words with low dimension.
scheme is expressed in equation. (4). where α = Latent Dirichlet (LD) allocation is a
(t(s) − µ) /σ (sigmoid function) aggregates possibility-based mechanism viable for
mean count of regular expressions, case level assortments, for example, text assortments. LD
and numeric literals. Sentence boundary consolidates the documents as a blend of shifted
calculation is expressed in equation (5). topics; every unit involves words that have a
Distributional Semantic Analysis weighs spun affiliation that exists between them. Also,
thematic concepts to find word co-occurrence. word choice simply dependent on the numerical
It is scored using formula. (6). Column one idea of likelihood. Recursively determining the
represents features, similarly column two interaction of themes and words is done for the
represents formulae used for computation and phase of a lonely record or a large number of
column three indicates optimal features
documents. At long last, yields the record Glowworm Optimization (GO)[22]
which comprises of different subjects. comprises of three phases namely Luciferin
LD allocation algorithm performs the update phase, Neighborhood phase and
following: movement phase. Adaptive Glowworm
1) Determine the number N of words in the Optimization (AGO) is proposed for tailor-
document concerning probability distribution is made features to acquire vectors or extract
Poisson. features to frame the summary. The
2) Pick a merge of focuses for the report optimization principle is based on five features
from a predefined set of K subjects as and their application phase is listed in Table-1.
demonstrated by the Dirichlet movement. Sentence position is additional feature to
3) Produce individually word in the list of revisit sentence with appropriate ordering.
terminology as follows language Vocabulary Positioning of sentences which is most vital
(V). part in the summary generation have higher
a) Choose a subject; weights. The feature associated with sentence
b) Choose a word in this subject. length; hence we have minimal set of 25 words
to accept as a sentence. F5 feature is in
3.4. Sentence Ranking movement phase of Glow worm optimization
with lucerifin value or luminous quotient, affine
towards the similar topics.
Sentence ranking phase, chiefly performs Luciferin update phase, sentences are
identification of prominent sentences with concatenated with respect to the relevancy.
pertinent information and free from Relevancy is determined by feature with
redundancy. It selects topmost sentences from respect to title and all sentences in document.
documents and produces summary with Luciferin update phase, movement phase and
application of traditional maximization neighbor phase are expressed as equation in (7)
algorithm like EM.[23]. finally produces
(8) and (9) respectively Luciferin
extractive summary. However, Extractive enhancement(ʋ) depends upon Proper Nouns.
summary lacks positional placement of
sentences. hence there is a need to revisit J(xi(t))-objective function which maximize
sentence positions.[24].
weights of every proper nouns. Luciferin decay
constant gradually decreases when common
3.5. Extractive Summary noun exists.
Movement Phase, forms local clusters based
Extractive summary is created with top n on decision range. Sentences are of similar
sentences for research articles summarization contexts likely to move based on entity-tagging
Latter summary is transformed to decision features. Finally, neighbor Phase performs
matrix. Decision Matrix will keep track of chronological sorting of clustered sentences to
extractive summary to make compatible for produces summary.
abstractive text summary with usage measure
and penalty measure. Hence, optimization 4. Results and Discussions
algorithm is used to remove exact replica of
original text produced in extractive summary. The extractive summarization shows better
Co-reference resolution is also handled in results with proposed procedure, the summary
abstractive text summary generation. produced purely relies on lexical features and
Meanwhile construction of cosine similarity surpass traditional keyword ranking schemes.
was rapid, useful, and seems reasonable. yellow color denotes dependency phrases and
green color denotes contextual theme. The
3.6. Optimization output of context aware component of
extractive summarization is shown as well as
output of abstractive text summarization is also
li(t + 1) = (1 − 𝜌)li(t) + ʋJ (xi (t + 1)); (7) shown.
Ni(t) = {j: dij < rid(t); li (t) < lj (t)} (8)
𝑟𝑖𝑑(𝑡 + 1) = 𝑚𝑖𝑛{𝑟𝑠, 𝑚𝑎𝑥{0, (9)
𝑟𝑖𝑑(𝑡) + 𝛽(ǹ𝑡
− |𝑁𝑝(𝑡)|}};
4.1. Training and Testing
Output of Context aware
For training purpose, Document understanding Extractive and Abstractive
Conference (DUC) data set taken into Summarization
considerations. The precise explanation of
DUC data sets and DUC data is customized,
which is free from least significant words or Extractive Summary
stop words according to port-stemmer’s A large number of methods applied in
algorithm. Recall oriented understudy Gist the field of extractive summarization
evaluation, Rouge(R) is also considered for over the years. Scoring sentences for
evaluation. R falls into many variants like R- such summary is tedious task. Many
unigram, R-bigram, R-Longest common
researchers putting so much effort to
Subsequence and R-N gram classes.
Multiple documents of artificial intelligence
improve the quality of summary.
domain for testing and performed various Document summarization focus both
measures like R-1, R-2, and R-L (Longest quality and coverage of content.
Common Subsequence) scoring. At Sentence- Clustering of sentences in document
level, computes longest common subsequence summarization shown promising results
(LCS) between two pieces of text ignores new to discover topics. A Fuzzy oriented
lines and summary-level, newlines in the text clustering for summarization of multi-
are interpreted as sentence boundaries, and the documents. Compendium- a summarizer
LCS is computed between each pair of tool, generates relevant summary free
reference and candidate sentences, and their from redundancy Collabsum clustering
results are tabulated below in table-2.
process both inter and intra document
The proposed AGO performs well in sentence
and summary level than traditional methods.
relationship and forms clusters. Clusters
similarly, contextual theme detection also out in turn apply graph based ranking
performs than traditional schemes like lexrank, methodology. FEOM…genetic
maximum relevance, loglikelihood ranking algorithm…graph -based
[25],[26],[27]
and other centrality measures as approach…probabilistic model.
baseline evaluation.
Table 2 Results Comparison Abstractive summary
Methods R-1 R-2 R-L R-L Various types of Sentence clustering
Sentence Summary techniques applied to document
level level summarization Sentence scoring, topic
Proposed 0.4611 0.1342 0.932 0.89 coverage, relevant sentences and
AGO summarization quality are main
MMR 0.3716 0.0757 0.16 0.80 components in summary production.
LEX 0.4030 0.0913 0.69 0.53 Clustering algorithms for sentence
scoring and grouping similar sentences
Proposed 0.4297 0.0835 0.12 0.83 according to topics conveyed in
contextual document. A fuzzy based, evolutionary
theme based clustering also successfully
Tf-idf 0.3639 0.0736 0.14 0.81 applied in conjunction with other graph-
LLR 0.3975 0.0850 0.084 0.64 based approaches to provide summary.
5. Conclusion References
Abstractive text summarization for research [1] Shafiei Bavani, Elaheh,
articles generates sentences individually using Mohammad Ebrahimi, Raymond
glowworm optimization with six associated Wong, and Fang Chen. "A query-
features. In addition, decision matrix with based summarization service from
elitism identification is formulated to choose multiple news sources." In 2016
summary sentences from both extractive IEEE International Conference on
summary sentences and abstractive summary Services Computing (SCC), IEEE,
sentences with consistency as necessary (2016): pp.42-49. doi:
condition. Extractive summary is reduced to 10.1109/SCC.2016.13.
more than 80% to generate abstractive [2] Bharti, Santosh Kumar, Korra
summary. Extractive summary with 661 word Sathya Babu, Anima Pradhan, S.
tokens is produced as output in first phase. Devi, T. E. Priya, E. Orhorhoro, O.
Later, decision matrix with Elitism Orhorhoro, V. Atumah, E. Baruah,
identification produces abstractive summary and P. Konwar. "Automatic
with 84 tokens is obtained as final output. keyword extraction for text
Proposed multi-document directed graph summarization in multi-document
contextual matching phrases, Rouge-L e-newspapers articles." European
measures in sentence level is 12.08 and Rouge- Journal of Advances in
L measures in summary level is 83.95 for Engineering and Technology
extractive summary. Similarly, Rouge-L (2017): 4, pp.410-427.
measures in sentence level is 9.32 and Rouge-L [3] Chowdhury, S. M., and Mazharul
measures in summary level is 89.68 for Hoque. "A Review Paper on
abstractive summary. A novel model has been Comparison of Different
implemented to be ample enough to provide Algorithm Used in Text
multi objectives and to convince the Summarization." Intelligent Data
instantaneous needs. Ultimately, this study will Communication Technologies and
inspire many researchers to further explore and Internet of Things: ICICI 2019 38
apply the various types of Swarm intelligence (2019): 114.
while solving the summarization tasks, [4] Cohan, Arman, and Nazli
specifically in the abstractive text Goharian. "Scientific article
summarization (ATS) field. summarization using citation-
context and article's discourse
6. Future works structure." (2017). arXiv preprint
arXiv:1704.06619
[5] Nouf Ibrahim Altmami, Mohamed
Decision matrix performs combination of El Bachir Menai, “Automatic
Sentences from extractive summary are summarization of scientific
assessed and deemed to be fit for abstractive articles: A survey.” Journal of King
summary are analyzed in conjunction with Saud University-Computer and
input from optimization algorithm with Information Sciences, (2020)
associated six features. Selection of best doi.:10.1016/j.jksuci.2020.04.020.
sentences and worst sentences based on their [6] Fan, Angela, David Grangier, and
usage and penalty is awarded to compose Michael Auli. "Controllable
summary. Global decision matrix performs abstractive summarization." arXiv
elitism identification (algorithm) and outputs preprint arXiv:1711.05217 (2017).
sentences with sentence flow as criterion. [7] Lloret, Elena, María Teresa Romá-
However, decision matrix follows Analytical Ferri, and Manuel Palomar.
Hierarchical Processing (AHP) [28] with user "COMPENDIUM: A text
defined decision values and their decisions are summarization system for
normalized. We can extend the normalized generating abstracts of research
vectors by using fuzzy [29] based membership papers." Data & Knowledge
assessment as stated by Charugupta,et.al.[30]
Engineering (2013):88, pp.164- [17] Nizami, Muhammad, and Ayu
175. Purwarianti. "Modification of Chu-
[8] Ferrández, Oscar, Daniel Micol, Rafael Liu/Edmonds algorithm and MIRA
Munoz, and Manuel Palomar. "A learning algorithm for dependency
perspective-based approach for solving parser on Indonesian language." In
textual entailment recognition." In 2017 International Conference on
Proceedings of the ACL-PASCAL Advanced Informatics, Concepts,
workshop on textual entailment and Theory, and Applications (ICAICTA),
paraphrasing, (2007). pp.66-71. IEEE (2017) pp.1-6.
[9] Gardenfors, Peter. The geometry of [18]Malik, Sonika, and Sarika Jain.
meaning: Semantics based on "Ontology based context aware
conceptual spaces. MIT press, 2014. model." In 2017 International
[10]Luhn, Hans Peter. "The automatic Conference on Computational
creation of literature abstracts." IBM Intelligence in Data Science (ICCIDS),
Journal of research and development IEEE (2017). pp.1-6.
(1958): 2, pp.159-165. [19]Sanchez-Gomez, Jesus M., Miguel A.
[11]Saggion, Horacio. "Learning predicate Vega-Rodríguez, and Carlos J. Pérez.
insertion rules for document "Extractive multi-document text
abstracting." In International summarization using a multi-objective
Conference on Intelligent Text artificial bee colony optimization
Processing and Computational approach." Knowledge-Based Systems
Linguistics, Springer (2011) pp.301- (2018): 159, pp.1-8.
312. [20]ShafieiBavani, Elaheh, Mohammad
[12]Maynard, Diana, Valentin Tablan, Ebrahimi, Raymond Wong, and Fang
Hamish Cunningham, Cristian Ursu, Chen. "A graph-theoretic summary
Horacio Saggion, Kalina Bontcheva, evaluation for rouge." In Proceedings
and Yorick Wilks. "Architectural of the 2018 Conference on Empirical
elements of language engineering Methods in Natural Language
robustness." Natural Language Processing (2018): pp.762-767.
Engineering (2002): 8, pp.257-274. [21]Anand Kumar M., Singh S., Kavirajan
[13]Witten, Ian H., and Eibe Frank. "Data B. and Soman K.P. “Shared Task on
mining: practical machine learning Detecting Paraphrases in Indian
tools and techniques with Java Languages (DPIL): An Overview”. In:
implementations." ACM Sigmod Majumder P., Mitra M., Mehta P.,
Record (2002):31, pp.76-77. Sankhavara J. (eds) Text Processing.
[14]Sethi, Nandini, Prateek Agrawal, Vishu FIRE 2016. Lecture Notes in Computer
Madaan, and Sanjay Kumar Singh. "A Science, Springer (2018):10478.
novel approach to paraphrase Hindi [22]Alphonsa, MM Annie, and P.
sentences using natural language Amudhavalli. "Genetically modified
processing" Indian Journal of Science glowworm swarm optimization based
and Technology (2016):9(28), pp.1-6. privacy preservation in cloud
[15]Alguliyev, Rasim M., Ramiz M. computing for healthcare sector."
Aliguliyev, Nijat R. Isazade, Asad Evolutionary Intelligence (2018):11
Abdi, and Norisma Idris. “COSUM: pp: 101-116.
Text summarization based on [23]Janani, R., and S. Vijayarani. "Text
clustering and optimization.” Expert document clustering using spectral
Systems (2019):36, clustering algorithm with particle
doi:10.1111/exsy.12340. swarm optimization." Expert Systems
[16]Gupta, Monika, and Parul Gupta. with Applications, Elsevier (2019) 134,
"Research and implementation of event pp.192-200.
extraction from twitter using LDA and [24] Xu, Song, Haoran Li, Peng Yuan,
scoring function." International Journal Youzheng Wu, Xiaodong He, and
of Information Technology (2019):11, Bowen Zhou. "Self-Attention Guided
pp.365-371. Copy Mechanism for Abstractive
Summarization." In Proceedings of the
58th Annual Meeting of the
Association for Computational
Linguistics (2020): pp.1355-1362.
[25] Weng, Shi-Yan, Tien-Hong Lo, and
Berlin Chen. "An Effective Contextual
Language Modeling Framework for
Speech Summarization with Augmented
Features." In 2020 28th European Signal
Processing Conference (EUSIPCO),
IEEE (2021): pp.316-320.
[26] Mallick, Chirantana, Ajit Kumar Das,
Madhurima Dutta, Asit Kumar Das, and
Apurba Sarkar. "Graph-based text
summarization using modified
TextRank." In Soft computing in data
analytics, Springer (2019): pp.137-146.
[27] Sabbah, Thabit, Ali Selamat, Md Hafiz
Selamat, Fawaz S. Al-Anzi, Enrique
Herrera Viedma, Ondrej Krejcar, and
Hamido Fujita. "Modified frequency-
based term weighting schemes for text
classification." Applied Soft Computing
(2017):206, pp.58 193.
[28] Tofighy, Seyyed Mohsen, Ram Gopal
Raj, and Hamid Haj Seyyed Javad. "AHP
techniques for Persian text
summarization." Malaysian Journal of
Computer Science (2013): 26, pp. 1-8.
[29] Bansal, Neha, Arun Sharma, and R. K.
Singh. "Fuzzy AHP approach for legal
judgement summarization." Journal of
Management Analytics (2019):6,
pp.323-340.
[30] Gupta, Charu, Amita Jain, and Nisheeth
Joshi. "Fuzzy logic in natural language
processing–a closer view." Procedia
computer science (2018):132 pp.1375-
1384.