You are on page 1of 11

Fast-Track Semester 2022

Technical Answers to Real-World Problems


Digital Assignment 2
Prof. Rajakumar K

Informative Text Summarizer Using NLP and DL

Team Members:
Saumitra Pathak (19BCE2411)
Shivam Bansal (19BCE0930)
Arkaraj Ghosh (19BCE24218)
Debalay Dasgupta (19BCE2423)
Pratyay Piyush (19BCE2364)

Literature Review:
No Paper Title & Method/ Challenges Observations Research Gap
Journal details Algorithm

1. Title: Scene Decision Trees, The challenge In this paper, The dataset
Classification for
Neural Networks, here is they they have tried used here is a
Sports Video Support Vectors are trying to to propose a very small one.
Summarization Machine (SVM), use a smaller model for The Video’s
Using Transfer Deep-Learning, dataset and sports video transcript must
Learning. AlexNet they use data scene be clear and
Convolutional augmentation classification free of
Journal: Sensors Neural Network, to improve the with the grammatical
2020, 20, 1702 Long Short term quality of particular errors only then
memory network scene intention of will the
classification. sports video classifier work
summarization. smoothly.
This article As it takes 3
focuses on models the time
presenting a complexity is
method to high.
achieve
high-quality
scene
classification
using a smaller
dataset.

2. Title: Automatic Deep Cricket Cricket is As most of the Here, the


Video Summarization played in a cricket matches researchers
Summarization Network (DCSN), wide range of are used short data
from Cricket CNN, formats and long-lasting, the for the dataset,
Videos Using bi-directional comprises a audience and video
Deep Learning. Recurrent Neural variety of preferably quality also
Network (RNN), complex needs to have needed to be
Journal: 2020 Determinantal rules. For this the high thus
23rd Point Process complicacy, summarized increasing the
International (DPP). automatically format of the amount of the
Conference on creating large content. dataset.
Computer and skims from In this paper,
Information cricket match they intend to
Technology videos is an speed up
(ICCIT), 19-21 inherently large-scale
December, 2020 difficult task. video
processing by
providing a
concise
synopsis that
still conveys the
story of a
cricket match.

3. Title: Slide classifier, a It is really a Here, they have Training and


Lecture2Notes: deep CNN, challenging proposed a preprocessing
Summarizing LSTM, task to find multimodal unbalanced
Lecture Videos Bidirectional and and locate the automatic training-test
by Classifying Auto-Regressive valuable summarization length. Video
Slides and Transformers piece of approach for used is too
Analyzing Text. (BART) or information, lectures. Their complex then
PEGASUS key points, approach the classifier
Journal: Hayden and other stuff extracts the may take time.
T. Housen from each of voice transcript
Pawling High the lectures. and formatted
School Pawling, text of unique
NY 12564 slides. These
two sources are
combined and
summarized to
create detailed
notes.

4. Title: DSNet: A Summarizes the Variable-lengt Different from


Flexible training h durations of existing
Detect-to-Summ procedure, video supervised
arize Network for performing interests pose video
Video alternate training unique summarization
Summarization. for M epochs with challenges for methods
several video video without
Journal: IEEE sequences summarizatio temporal
Transactions On n. continuity and
Image integrity
Processing, Vol. constraints, the
30, 2021 anchor-based
DSNet regards
video
summarization
as an interest
detection
problem and
achieves very
promising
performance on
two
widely-used
datasets.

5. Title: Video It includes the Different Most of the Here the


Summarization overview of types of video dataset used is
using CNN and implementation methods used summarization too short, and if
Clustering approaches for for video techniques use the video is
Algorithm. an enhanced summarizatio Histogram as a sarcastic in
algorithm for n techniques feature tone then it has
Journal: JASC: video have been extraction difficulty in
Journal of summarization of surveyed and technique. But classifying.
Applied Science cricket video found that the Histogram
and using video most of the feature
Computations. processing. The methods extraction
Volume 5, Issue purpose of this depend on technique is not
10, proposed the type of used for all
October/2018 algorithm is to video (video types of videos.
enhance the formats). In It has less time
video the future, a complexity as
summarization reply can be compared to
method. detected and the CNN
reduce the feature
summary. extraction
Using some technique. But
algorithm in it doesn’t
neural provide an
network time accurate result.
complexity CNN feature
can be extraction
reduced. technique is
common for all
because the
model is trained
by ImageNet
(large visual
database)
dataset. So all
types of videos
can be
summarized
accurately.

6. Title: Deep Use image Because of The algorithm Here the main
Learning Based detection and limitations of currently works issue is with the
Automated localization hardware, for mp4 videos video quality,
Sports Video techniques to video clips only, we plan to the video
Summarization determine the extraction extend the quality needs to
using YOLO. location of the was not algorithm to be high and
scoreboard in a possible and work on clear for the
Journal: Elcvia, video frame, crop only the different video image to be
VOL. 20 NO. 1 the scoreboard, timestamps of formats. classified by
(2021), JUNE run OCR over the key events YOLO.
2021. cropped are the
scoreboard, and output.
compare the
score with the
previous score to
dete, what
important events
are in the video.

7. Title: A Review Tree Based The This paper If the text is


Paper on Text Method, Abstractive discusses sarcastic in
Summarization. Template Based approach extractive and nature or has
Method, should be abstractive incorrect
Journal: Deepali Ontology Based explored summarization wordings then
K. Method, Lead more. with the the classifiers
Gaikwad and C. and Body Phrase techniques have shown to
Namrata Method used, fail.
Mahender. Vol. advantages
5, Issue 3, and
March 2016 disadvantages,
And technique
used in each
approach.
An abstractive
summarization
needs more
learning.
However, it
provides a
more
meaningful and
appropriate
summary
compared to
extractives.

8. Title: Soccer It comprises two This model is Each and every Here, the
Video models. specific to summarized researchers
Summarization 1)A 3D-CNN for soccer. It can video is used short data
using Deep feature extraction be applied to evaluated using for the dataset,
Learning. 2)An LSTM some the Mean and video
Network ball-related Opinion Score quality also
Journal: . In this paper, a sports such (MOS) scale. needed to be
Rockson soccer dataset is as basketball, Where 1=very high thus
Agyeman, Rafiq created by rugby, and poor , increasing the
Muhammad and manually volleyball. 2=poor,3=neutr amount of the
Gyu Sang Choi annotating 744 However, it al,4= good and dataset.
Department of soccer clips. A cannot be 5 = very good .
Information and ResNet based applied to The
Communication 3D CNN model is other sports. summarized
Engineering developed which videos got a
Yeungnam can distinguish score of 4 out
University, April soccer actions. of 5 out of
2019. Then , this action MOS.
recognition
generalization to
soccer clips is
extended . The
model extracts
features from
soccer clips.
Then an LSTM
network is trained
based on the
features. In order
to summarize the
soccer video , the
input is treated
as a sequence of
video segments.
All highlights are
combined into a
summarized
video.

9. Title: Video They represented Dividing the Learned The input video
Summarization the input video as video in features is crucial here,
using Deep a sequence of segments is extracted from deciding that
Semantic deep features in difficult and videos with may take time.
Features. the semantic must be done similar content Also dividing
space, each of manually. to make the video in
Journal: The which clusters in the segments is
13th Asian corresponds to a semantic difficult and has
Conference on video segment. space. The to be done
Computer Vision This input video is manually as the
(ACCV'16), Sep representation represented by model can’t
2016. can encode the deep features divide the video
semantic in the semantic into segment by
transition of the space, and itself by just
video and thus segments seeing the
can be useful for corresponding video into
various tasks to cluster topics.
including video centers are
retrieval, video extracted to
description generate a
generation, etc. video summary.
Some clusters By comparing
can be observed, the summaries
each of which are to manually
expected to created
contain summaries,
semantically they have
similar video shown the
segments. advantage of
Based on this incorporating
assumption, their our deep
approach picks features in a
out a subset of video
video segments summarization
by optimizing an technique.
objective function
involving the
representativene
ss of the subset.

10 Title: Video The proposed They converted The dataset


. Summarization system takes the the subtitles of used over here
using NLP. input of a the video into a was the Ted
YouTube video text document Talks,
Journal:Sanjana link and the time and then sometimes the
R, Sai Gagana duration to which applied the speaker can be
V, Vedhavathi K the video has to summarization informal and
R, Kiran K N, be summarized. algorithm. They make a few
Aug 2021. After generating fitted to the jokes that time
the summarized duration given the classifier
video using NLP by the user and takes it literally.
based LSA then created
algorithm, video the final
output is summarized
displayed on the video.
web page. For
the purpose of
illustration, they
have considered
a Ted talk
available on
YouTube, and the
summary for the
Ted talk is
displayed.

11. Title: A Survey They have Evaluating The extractive If the text is in a
on Extractive proposed an summaries summarization sarcastic tone
Text iterative ranking (either technique then the
Summarization. algorithm automatically focuses on classifier fails.
(variation or manually) choosing how
Journal: IEEE of HITS is paragraphs,imp
International algorithm) which a difficult task. ortant
Conference on is efficient The main sentences, etc
Computer, in selecting problem in produce the
Communication, important evaluation original
and Signal sentences and comes from documents in
Processing also ensures the precise form.
(ICCCSP- 2017) coherency impossibility The implication
in the final of of sentences is
summary. building a determined
standard based on
against which linguistic
the results and statistical
of features. In this
the systems work, a
have to be comprehensive
compared. review of
extractive text
summarization
process
methods has
been
ascertained

12 Title: Digital Three-dimension The main The effective Training and


. Video al Convolutional challenge is implementation preprocessing
Summarization Neural Network the of video unbalanced
Techniques: A (3D-CNN) and 1) training summarization training-test
Survey. Long Short-term and applied in length. Video
Memory (LSTM) preprocessing different used is too
Journal: IJERT Recurrent Neural unbalanced real-world complex then
Volume 09, Network (RNN). training-test scenarios like the classifier
Issue length.2) movie tailor in may take time.
01 (January complexity in the film
2020) application industry, the
and highlight in
development. football soccer,
3) The and anomaly
temporal detection video
relationship surveillance
between system.
video frames
in information
like video
tags,
captions,
comments,
and so on will
need t to be
investigated
in the
future.4)
inexpensiven
ess of training
video mostly
the annotated
dataset.

13 Title: Video This mechanism 1)Processing The video Video transcript


. Summarization firstly divides the the summarization must be
Techniques: A input file into unbalanced is used to provided with
Review. many segments, length of a overcome correct
where the first dataset. these issues grammar, only
Journal: frame of each 2)In-expensiv that deal with then will the
International segment is eness of lengthy videos classifier work
Journal Of considered as a training and condense properly.
Scientific & representative. videos. those, based
Technology However, these 3)Subjectivity on the various
Research frames can be (Variation in features. This
Volume 9, Issue computed by the the selection paper presents
11, histogram. of shot) a review and
November 2020 Finally, the 4)Redundanc comparative
histogram can be y (Minor analysis of
clustered them difference of video
by the k-means frames is summarization
algorithm. considered techniques.
as a key Initially, these
frame). methods have
5)Highly been classified
changes in into different
content make categories
clustering based on their
difficult. characteristics.
6)The
complication
in the
development
of an
application.
7)Integration
of application
and model.

14 Title: Automated Static and Developing Compact Dataset was


. Video Dynamic efficient representations too small, and
Summarization Visualization of representatio of video data the video
Using Speech Video ns for video can enable transcript
Transcripts. Summaries. browsing efficient video should always
presents browsing. Such be provided
Journal: Video some unique representations only then the
and Image algorithmic provide the classifier could
Processing challenges,1, user with identify and
Laboratory 2 as well as information summarize.
(VIPER) School new technical about the
of Electrical and challenges.3 content of the
Computer Video is a particular
Engineering sequential sequence being
Purdue and examined while
University West information-ric preserving the
Lafayette h medium. It essential
includes message.
audio and
motion, and it
carries long
temporal
contextual
relationships
between
shots and
scenes. In
contrast to
images in an
image
database,
manipulation
of video is
inherently
more
complex.

15 Title: A Review Genetic Problem of They have Redundant


. on Text Algorithm (GA), redundancy, introduced data was used.
Summarization Particle Swarm Problem of some concepts
Techniques. Optimization irrelevancy, related to
(PSO), Harmony Problem of extractive text
Journal: Volume Search (HS), loss of summarization
64, Issue 1, Differential coverage, and then
2020 Evolution (DE), Problem of provide a
Journal of and Cat Swarm non-readabilit systematic
Scientific Optimization y and less analysis of
Research. (CSO) for single cohesive various text
as well as content. summarization
multi-document techniques.
summarization

Problem Statement
Text Summarization is a key tool when it comes to understanding a large piece of text and
spontaneously understanding its meaning. This project quantifies factual consistency at the
entity level only and leaves the relation level consistency for future work. We propose a
metric to quantify such hallucinations made by the model as well as propose a couple of
measures and training schemes that could help the model perform better and generate
entity-level factually correct summaries. The main objective is to cut the clutter and deliver a
summarized paragraph of News articles, informative/academic youtube videos, and
Academic Journals.

You might also like