Arun Krishna Chitturi et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(6), November - December 2019, 2956- 2964
2956
ABSTRACT
Text summarization is the core aspects of Natural Language
processing. Summarized text should consist of unique
sentences. It is used in many situations in today’s Information
technological word, one of the best examples is in
understanding customer feedbacks in companies. This job can
be done by humans, but if the text or data that has to be
summarized then it will consume lot of time and work force.
This situation lead to birth of different approaches in
summarization. This paper addresses and concentrates on
various methods and approaches and their results in
abstractive text summarization. This survey gives an insight
about different types of text summarization and various
methods used in abstractive text summarization in recent
developments.
Key words : abstractive summarization, decoder, encoder,
multi document summarization
1. INTRODUCTION
Summarization is very well useful to us in today’s world.
The main aim of abstractive text summarization is to produce
shortened version of input text with relevant meaning[7]. The
adjective abstractive is utilized because it denotes that the
generated summary is not a combination or selection of some
repeated sentences, but it a paraphrasing of core contents of
the input document [8]. Abstractive summarization is a very
difficult problem apart from Machine translation. The main
challenge in ATS is to compress the matter of input document
in an optimized way so that the main concepts of the
document are not missed [8]. In current technologically
advancing world, volumes of data is increasing and it is very
difficult to read the required data in short time[6]. It is a pretty
task to collect the required information and then convert into
summarized form. Therefore, text summarization came into
demand. Summarized text saves time and helps in avoiding
retrieving massive text. Abstractive Text summarization can
be combined with numerous intelligent systems on the basis
of NLP technologies like information retrieval, question
answering, and text classification to find the particular
information [9]. If latent structure information of the
summaries can be incorporated into abstractive
summarization model, then the quality of summaries
generated can be improved [10]. In some research works,
topic models are used to capture the latent information from
the input paragraph or documents. Despite having many
hurdles abstractive text summarization faces core issues like
(i) Neural sequence-to-sequence models which try to produce
generic summary, which include mostly used phrases (ii) The
generated summaries are less readable and are not
grammatically perfect [11]. Summarization is divided into
following types: (a) Extractive text summarization (b)
Abstractive text summarization [6]. Extractive
summarization extracts the frequently used or only precise
phrases without modifying them and generates the summary.
Whereas abstractive summarization generates new sentences
and also optimally decreases the length of the document.
Abstractive is better and qualitative than extractive as it takes
data from multiple documents and then generate precise
information of summary. Abstractive summarization is again
achieved in two ways. They are: (a) Structure based approach
(b) Semantic based approach. Neural network models on the
basis of encoder decoder for machine translation achieved
good ROGUE scores [12]. Abstractive approaches generate
summary similar to summary generated by humans but they
are more expensive [13]. On the basis of current state of RNN
in Attentive RNN the encoder computes score over the input
sentences [14]. The main problem in ATS are (a) Long
document summarization (b) Abstractive metric (c)
Controlling output length. F1 scores are evaluated generally
using ROUGE metrics [15]. Recall-Oriented Understudy for
Gisting Evaluation (ROUGE) metric was proposed by (Lin,
2004) [24]. Named Entity Recognition is also one of the core
application in NLP which helps in removing ambiguity [28].
Information Retrieval is also highly difficult and it requires
quality documents[37].
2. SURVEY
2.1 Semantic Link Network For Summarization[1]:
SLN is a semantics self-formulated for semantically
organizing resources to support advanced information
services like Abstractive Text Summarization [1]. According
the author the semantic link network, which is used in
Abstractive text summarization, has following important
components:
Survey on Abstractive Text Summarization using various approaches
Arun Krishna Chitturi
, Saravanakumar Kandaswamy
1
Vellore Institute of Technology, Vellore, India, chitturiarunkrishna@gmail.com
2
Professor at Vellore Institute of Technology, Vellore, India, ksaravanakumar@vit.ac.in
ISSN 2278-
Volume 8, No.6, November – December 2019
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse45862019.pdf
https://doi.org/10.30534/ijatcse/2019/45862019