INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 01, JANUARY 2020 ISSN 2277-8616

Automatic Answer Script Evaluator

M. Venkateshwara Rao, I. Sri Harshitha, Y. Sukruthi, T. Sudharshan

Abstract: The way of correcting or evaluating a answer script in the present days is becoming so adverse nowadays and also a heavy burden for the
faculty to correct the papers this is the major problem found nowadays in the educational institutions and also by the human evaluator there might be
some mistakes over taken or even some overridden the problem arises if the hand writing might not be understood by the evaluator or even the mood of
the person hence we came up with the idea of machine correcting the papers in this project we are having two parts converting the hand written letters to
a text document and then evaluating it . In evaluating we use the techniques of computer vision and artificial intelligence m ethods for converting into text
from hand written and for the later we use a algorithm called as natural language processing which understands the natural language of the human and
tokenizes the words required and then checks the words with the key already uploaded by the lecturers hence the evaluation is perfectly done and the
burden also decreases. The evaluation processes mostly depends on the key words used in the answer script and the matching in the key paper given
by the faculty

Keyword: Artificial Intelligence, Machine learning , Natural language Processing

——————————  ——————————

1. INTRODUCTION Here there are two parts in it

The era of technological development today, without realizing
the technology has become a need that is often used for the 1) Hand written character recognition
life of today's society. [1]IoT is a structure in which objects,
people are provided with an exclusive identity and the ability to 2) Evaluating the answer scripts
move data through a network without requiring two-way inter- Hand written recognition:
human to human source either destination or human In hand written recognition we have a tendency to ar
interaction to the computer. IoT is a very promising victimization optical character recognition. The handwritten
development to optimize life based on intelligent sensors and detection algorithm takes input images containing a line of
smart appliances that work together over the internet. text and words and returns a matrix with the chance of every
The basic way of assessing the students capability is character showing in the matrix. Specially, the matrix which is
evaluating the answers in the exam written by them this way of the size (sequence_wordlength × vocabulary. of type
we can assess the students learning ability .[2]The automated characters).
system if it does the evaluating there might be a accurate Our pipeline to mechanically acknowledge written text
measurement of the capability of the students instead of the includes: page segmentation and line segmentation
human correcting the paper which takes a lot of effort and cost
and it might not be totally accurate . here unlike other answer Evaluating the solution scripts:
script evaluators we evaluate from the hand written answer Tokenization may be a technique employed in tongue outfit
script which is more comfortable and easily accessible for the (NLTK) to divide the sentences into words. the explanation
students and also easy instead of the text document .Here the behind the tokenization is to perform some things to be
faculty or the institution just need to upload the kay for the enforced on the tokenized words. [4]The component of
answers.[5] And just uploading the answer scripts directly to Speech tagging in NLTK is employed to assign the tag for the
the system it generates the marks for each student separately tokenized words. Then between these important keys are to
be taken down and to are checked with the tokens within the
________________________ key words
 M.Venkateswararao is currently working as Assistant
Professor,Department of Electronics and Communication
Engineering, Koneru Lakshmaiah Education Foundation, 2 METHODOLOGY:
Vaddeswaram, AP, India, PH-9441536316. E-mail: HandwritteRecognition:
[email protected] Briefly, utilizing a CNN for image classification and feature
 I.Sri Harshitha is currently pursuing bachelors degree program extraction and fed the options into a bitface
Department of Electronics and Communication Engineering, Koneru
LongShortTermMemory and trained the network to optimize
Lakshmaiah Education Foundation, Vaddeswaram, AP, India,
PH,8977799873, the Connectionist Temporal Classification (CTC) loss
E-mail: [email protected] Intuitively, the CNN generates image options that square
 Y.Sukruthi is currently pursuing bachelors degree program measure spatially aligned to the input image. The
Department of Electronics and Communication Engineering, Koneru image options square measure then sliced on the direction of
Lakshmaiah Education Foundation, Vaddeswaram, AP, India,PH- the text ANd consecutive fed into an LSTM. This network is
9030015799,
E-mail :[email protected]
denoted as CNN-biLSTM. the most reason that the CNN-
 T.Sudharshan is currently pursuing bachelors degree program biLSTM was elect is as a result of flat LSTMs well a lot
Department of Electronics and Communication Engineering, Koneru of computationally dearly-won compared to the CNN-biLSTM.
Lakshmaiah Education Foundation, Vaddeswaram, AP, India,PH-
7731998477,E-mail:[email protected]

4023
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 01, JANUARY 2020 ISSN 2277-8616

2.5 Select the suggested word with the shortest weighted

edit distance
Given a list of words to be changed, the weighted edited
distance between the suggested words and the given word
were calculated. Specifically, the edit distance was weighted on
the differences between the handwritten characters. The
Character similarity was modeled using the CNN-bitLSTM
performance, given the training examples.The number of
variations between actual words and the expected words
(insertions, deletions, and replacements) was counted. The
greater the frequencies of a mistake, the more characters are
visually similar. Therefore, the weight should be smaller. The
main problem is the weighted
figure1: Block diagram for Hand written recognition by CNN distance is limited to one character being replaced by another.
It is common for compounds to look alike (e.g., "rn" and "m") in
Multiple down samples were provided to help in handwriting recognition.
recognizing pictures of written text that adjust in size (e.g.,
lines that contain only one word vs lines that 3. Contract texts back to its original form
contain seven words). Note that a pre-trained res-net was This step merely detokenizes and contracts the words (if it
used as a image feature extractor. absolutely was initially decontracted). All in all, the vocabulary
Language models spell checker works best for anticipated strings that a awfully
We explored 3 ways to extract a string of words given a matrix preparing to the specific worth. In any case, if the yield of the
of chances square measure explored: 1) greedy search, 2) covetous algorithmic program doesn't give shut enough string
lexicon looking, and 3) beam looking + lexicon looking with an proposition, the vocabulary spell checker will truly downsize
language mode correct nesses as opposed to up it. A last check to ensure that
2.1Greedy_search: The ravenous inquiry technique just the information and yield sentences ar comparative was
repeats over whenever step and gets the preeminent plausible performed. Shaft watching out + vocabulary search +
character at each time step: This approach doesn't utilize any
language model In request to ease the issues of acquiring
outer language models or vocabulary to change the outcomes. poor proposition from the avaricious algorithmic program, one
2.2Lexicon_searching: The dictionary search model utilize the
will rehash through the probability framework to get different
yield of the covetous technique what's more, makes an proposition. Nonetheless, positioning every single achievable
endeavor to coordinate each word with a lexicon. The proposition is computationally expensive. Graves et al. [5]
techniques are the ensuing techniques were taken: arranged that the pillar search algorithmic program will be
1)Decreasing the size of the string, 2) Giving tokens to the acclimated create K recommendations of strings given the
string based on the words 3) for each word: 3.1) check if the likelihood matrix. In our execution, the K sentences are at that
word given is a English word or identify the language of the point sustained into the vocabulary spell checker to guarantee
word if not, suggested words similar to that 3.2)Replace the that words will be found in a very vocabulary ( else it'll appear
word with any other word similar to that with the shortest as ). Every K proposition is then tokenized and continuous
weighted distance 4) Contract texts back to the its nourished into a language model to gauge the psychological
original kind. Every string is checked and delineated to the disarray of each sentence proposition. The sentence
string and therefore the string is de contracted Here the proposition with the base mental perplexity was picked. The
contractions of the string means the similar version the word voracious, dictionary search, and shaft search yields blessing
but at a smaller and shorter word . For example, have not → comparable and reasonable expectations for the picked
haven’t, I have → I’ve, (For more words look here ). In the Way models. In Figure about six, eye catching models are given.
of removing the contractions the contractions were used. The the essential line of Figure about six show cases any place the
below library shows the contractions and decontarctions vocabulary search algorithmic program given fixes that
whether necessary or not For deconstructing ambiguous cases adjusted the words. The prime model inside it, "cotching" (as
(e.g., ain’t → am not/are not), the library utilizes a language it was written entirely then it can be written) was "catching"
model (google news) to assess the correct case. corrected and "woving" was "waving" corrected. Moreover, the
beam search output corrected "a" to "all," but an area between
2.3 Tokenization of the string in the sentence: "lovely" and "things" is incomprehensible. In the second
Tokenization is a process of separating a string containing a example, the lexicon search output "selt" was born-again to
sentence into distinct words. The tokenize package from the "salt." "Selt" was wrongAll the same, though. Born-
natural language toolkit was used to execute this. This process again to "self" in the search results of the beam. Therefore, be
is continued till every word is tokenized and done by iterations am search performed worse in this example. In the third exam
ple, none of the three ways offered significantly accessible res
2.4 Suggest words if it’s not a word
ults. Finally, the lexicon search algorithm in the fourth example
For all the words that are Tokenized, the pyechent module is program incorrectly born-again “for him” into “forum”, but the
used to check the word if it is a English word or not if it is a beam search algorithmic program properly known “for him”.
English word no changes are made if it is not an English word Quantitatively, the greedy algorithmic program had a mean
then norvig spellchecker checks the word and suggests the
character error rate (CER) = eighteen.936 whereas the lexicon
nearest word possible for the word search had CER = eighteen.897. while not the weighted edit
distance, the lexicon search had CER = nineteen.204, well
4024
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 01, JANUARY 2020 ISSN 2277-8616

reducing the performance of algorithmic program. As given in comparative problem zone system to be incorporated into the
Figure half dozen, the CER improvement between exploitation knowledge by the school towards the beginning. The
the greedy algorithmic program and therefore the lexicon researchers ' arrangement sheet is to be moved within the
search was token is. Needless to say, the beam search record location, and each corresponding answer within the
algorithmic program out performed the greedy and therefore information archive area unit tokenizes what's more, divided
the lexicon search results with CER = 18.840. into words. At that point the key sentences zone unit
recovered from the data and zone unit tokenized and split into
the words The token partition clarification is to isolate the
sentence into word by word all together that it is clear to
coordinate with the arrangement the phonetics which methods
for the key sentences. By exploiting the nltk (Natural
Language Tool Kit) in-manufactured work known as "nltk.word
tokenize(sentence)," the token division is upheld by any place
sentence that is the parameter that will be tokenized.
Information: "It might be a fake language . it's a system
acclimated language." Yield: ['It', 'is', 'a', 'programming',
'language „. „It', 'is', 'a', 'strategy', 'arranged', 'language', '.']
Parts of Speech tagging unremarkably English language
comprises of eight segments of discourse. The POS tag is
designated upheld the utilization of words inside the
sentences. Similarly to see the use of words inside the content
archive, everything about words inside the tokenized
sentences is distributed with the worthy segments of discourse
tag. The key sentences for the relating answer region unit to
be recovered from the data and assigned with the satisfactory
segments of discourse tag. The parts of discourse labeling is
authorized by exploitation the nltk (Natural Language Tool Kit)
in-assembled work known as "nltk.pos_tag(words)" any place
words is that the parameter that zone unit tokenized by the
Figure 2:evalution of paper by using key sheet penetrable operate. Input: ['It', 'is', 'a', 'programming',
'language „. „It', 'is', 'a', 'system', 'arranged', 'language', '.']
Tokenization The solution sheet tokenization can contain Output: [('It', 'PRP'), ('is', 'VBZ'), ('a', 'DT'), ('programming',
multiple paragraphs and each paragraph contains multiple 'NN'), ('language‟,‟NN‟),(„.‟,‟.‟),(„It', 'NN'), ('is', 'VBZ'), ('a',
sentences in succession. Sentences should be divided from 'DT'), ('system', 'NN'), ('arranged', 'VBN'),('language', 'NN'), ('.',
paragraphs and words from sentences for analysis. Such a '.')] Stop Words Removal The response for the specific inquiry
cacophonic mechanism between sentences or paragraphs is shifts from one understudy to another. English lexicon
referred to as a token of separation comprises of a significant number of the filler words like "an, of
Speech Tagging elements The NLTK module includes a powe an.etc".This stop word evacuation module is used to evacuate
rful side called Speech tagging elements dealing with the assi the filler words from the content record of the student. The key
gnment of a label such as Noun, Adjective. etc. For each of sentences are to be recovered from the database for the
the words tokenized. Avoid Words Removal Stop System specificinquiry and the words stops are to be expelled from the
words.That doesn't have vital importance to be used in Search key sentences. The stop words evacuation is executed by
Queries. These words space unit sometimes filtered out of utilizing the English word lexicon and the nltk (Natural
search queries as a result of Brobdingnag Ian quantity of Language Tool Kit) worked in work called "stop_words =
surplus information.Stemming words Stemming is the method set(stopwords.words('english'))" Input: [('It', 'PRP'), ('is', 'VBZ'),
used to reduce inflected (or generally derived) words to their ('a', 'DT'), ('programming', 'NN'), ('language‟,‟NN‟),(„.‟,‟.‟),(„It',
stem, base, or root type in linguistic morphology and data 'NN'), ('is', 'VBZ'), ('a', 'DT'), ('strategy', 'NN'), ('situated',
retrieval. Computing the similarity of the similarity of the 'VBN'),('language', 'NN'), ('.', '.')] Output: [('programming', 'NN'),
sentence linguistics can be a measure illustrated over a ('language‟,‟NN‟),(„.‟,‟.‟), ('strategy', 'NN'), ('situated',
number of records or words. The hole between the 'VBN'),('language', 'NN'), ('.', '.')] Stemming words This module
arrangement of archives or terms is based on their similarity is used to examine the methods used for tense words since
that implies content or etymology rather than representation of the methods used for words take on an essential job rather
the sentence structure (such as their string group).Report than tenses. To be tokenized the key sentences zone unit and
Generation This Module is utilized to send the report back to each tokenized word zone unit to be stemmed to provide the
the researchers concerning their presentation in each right tense word coordination. The stemming words is
question. the arrangement sheet should be assessed for each implemented by abuse the nltk (Natural Language Tool Kit)
understudy and furthermore the outcome's sent to all or any module known as "Watchman Stemmer" and in- fabricated
the researchers through mail. The report concerning the perform known as "stem(words)" any place words is that the
general execution of the researchers, singular execution of the parameter that E. Projection of extreme Scores: Last
researchers, subjects {in that during which within which} the determined scores apportioned to understudy Unit of reaction
researchers region unit powerless and which must be focused zone specified in the report. Current progress in evaluating the
on extra zone unit sent to the school individuals for extra illustrative response of the framework is as follows
activities. Tokenization of the main sentences for the Step 1: start.

4025
IJSTR©2020
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 01, JANUARY 2020 ISSN 2277-8616

Step 2: type and store the correct answer in the table. As more enterprises embrace the Internet of Things, a host of
Step 3: Using POS tagger, set up watchwords, label new security vulnerabilities will emerge. The increased risk
catchphrases, and assigned loads endless proximity in can be attributed to device limitations, and because of missed
sentence. opportunities to enhance security. Here are 12 leading IoT
Step 4: Store equivalent words and antonyms into another security challenges that enterprises must address.
table. • Secure constrained devices
Step 5: set understudy score to zero. Info understudy reaction • Authorize and authenticate devices
and store it in another table i.e.SR table. • Manage device updates
Step 6: presently check whether catchphrase blessing in SR • Secure communication
table if blessing apportioned score=score + effectively • Ensure data privacy and integrity
distributed weight. • Secure web, mobile, and cloud applications
Step 7: If watchword not discovered sign in equivalent words • Ensure high availability
table and on finding allocate score=score + effectively • Detect vulnerabilities and incidents
distributed weight. • Manage vulnerabilities
Step 8: Check on the off chance that antonyms blessing in SR • Predict and pre-empt security issue Attacking IoT
table in the event that blessing, at that point score=score*(-1). • Default, weak, and hardcoded credentials
Step 9: Check the position vector of noun and verb • Difficult to update firmware and OS
combination in input answer and compare it thereto of correct • Lack of vendor support for repairing vulnerabilities
answer to verify dependencies of noun and verb in answer. • Vulnerable web interfaces (SQL injection, XSS)
Step 10: if grammatical mistakes area unit gift deduce a pair of • Coding errors (buffer overflow)
marks from internet score. • Clear text protocols and unnecessary open ports
Step 11: currently build summation of allotted several Student • DoS / DDoS
responses. • Physical theft and tampering
Step 12: If scores calculated area unit negative then answer
entered by student is inaccurate, else if scores area unit in 5 CONCLUSION:
same vary with already allotted scores then student response Here by we conclude that, through this automatic answer
is marked to be CORRECT. script evaluator which decreases the correction stress for the
faculty and also if there might be any miscorrections also there
4 Results: so to reduce those we can use this automatic answer script
Answer sheet: evaluator.This answer script evaluator is a better performer
when compared to other because of using CNN this can also
be implemented very easily.

6 REFERENCES:
[1] Raghava Prasad C., Kishore P.V.V., Morphological
differential gradient active contours for rolling stock
segmentation in train bogies,2016, ARPN Journal of
Engineering and Applied Sciences, Vol: 11Issue: 5,
pp: 2799 - 2804, ISSN 18196608
[2] Hari Priya D., Sastry A.S.C.S., Rao K.S., FPGA based
design and implementation for detecting Cardiac
arrhythmias ,2016, ARPN Journal of Engineering and
Applied Sciences, Vol: 11, Issue: 5, pp:3513 - 3518,
ISSN 18196608
[3] Kishore P.V.V., Srinivasa Rao M., Raghava Prasad C.,
Anil Kumar D., Medical image watermarking: Run
through review ,2016, ARPN Journal of Engineering
and Applied Sciences, Vol: 11, Issue: 5, pp:2882 -
2899, ISSN 18196608
[4] Mannepalli K., Sastry P.N., Suman M., MFCC-GMM
based accentrecognition system for Telugu speech
signals ,2016, International Journal of Speech
Technology, Vol: 19, Issue: 1, pp: 87 - 93,
ISSN13812416
[5] Ur Rahman M.Z., Mirza S.S., Process techniques for
Figure3: anwer script human thoracic electrical bio-impedance signal in
remote healthcare systems ,2016, Healthcare
Marks: Technology Letters, Vol: 3, Issue: 2, pp: 124 - 128,
9/10 ISSN 20533713
Key: [6] Krishna A.R., Chakravarthy A.S.N., Sastry A.S.C.S., A
1) Identifying the security challenges existing in IOT system? hybrid cryptographic system for secured device to
device communication ,2016, International Journal of
Electrical and Computer Engineering, Vol: 6, Issue: 6,
4026
IJSTR©2020
www.ijstr.org
INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 01, JANUARY 2020 ISSN 2277-8616

pp: 2962 - 2970, ISSN 20888708

[7] Satyanarayana P., Sai Prajwal K., Chandra Naga
Varma T., Sri Manojna E., Sitara S., Advanced motion
tracking based mobility assistance for physically
disabled ,2016, ARPN Journal of Engineering and
Applied Sciences, Vol: 11, Issue: 15, pp: 9545 -9552,
ISSN 18196608
[8] Abburi R., Sastry A.S.C., A novel algorithm for the
arrhythmia diagnosis in fetal monitoring system ,2016,
IIOAB Journal, Vol: 7, Issue: 9Special Issue, pp: 746 -
753, ISSN 9763104
[9] M. V. Srinu, M. V. Rao and K. P. R. R. Raju, "A
novel method for salient image enhancement,"
2018 Conference on Signal Processing And
Communication Engineering Systems (SPACES),
Vijayawada,2018, pp. 189-193.doi:
10.1109/SPACES.2018.8316343
[10] Venkata Srinu, M, Venkateswara Rao Morla, Kali
Vara Prasad
[11] Baditha, Vara Kumari. S, & Srinivas
Maddimsetti. "License Plate Recognition System
Using Matlab GUI For
Surveillance."International Journal of
Engineering & Technology [Online], 7.2.7 (2018):
1008-1016. Web. 20 Nov. 2019
[12] Srinu, M. Venkata et al. “A Multi-Scale Image
Enhancement Model using Human Visual System
Characteristics.” (2015)
DOI:10.5815/ijigsp.2015.06.01
[13] Mannepalli K., Sastry P.N., Suman M., FDBN: Design
and development of Fractional Deep Belief Networks
for speaker emotion recognition ,2016, International
Journal of Speech Technology, Vol: 19, Issue: 4, pp:
779 - 790, ISSN 13812416
[14] Kishore P.V.V., Kishore Kumar D., Anil Kumar D., Sai
Pujitha G., Simarjeeth Singh G., Bala Ananth Sai K.,
Mohan Kalyan K., Sri Siva Ananta Sai B., Manikanta
M., Nanda Kishore M. , Edge and texture preserving
hybrid algorithm for denoising infield ultrasound
medical images ,2016 , Journal of Theoretical and
Applied Information Technology ,Vol: 86 ,Issue: 1 ,pp:
120 - 129 ,ISSN 19928645
[15] Gowri T., Kumar P.R., Rahman Md.Z.U., Reddy
D.V.R.K., Efficient electrocardiogram signal
enhancement techniques using blockprocessed noise
cancelers ,2016, Journal of Medical Imaging and
Health Informatics, Vol: 6, Issue: 3, pp: 739 - 745,
ISSN 21567018
[16] Prasad M.V.D., Kishore P.V.V., Kiran Kumar E., Anil
Kumar D., Indian sign language recognition system
using new fusion-based edge operator ,2016, Journal
of Theoretical and Applied Information Technology,
Vol: 88, Issue: 3, pp: 574 - 584, ISSN 19928645
[17] Anil Kumar D., Kishore P.V.V., Venkatram N., Leela
Rani B., Hasitha A., Sabarish T.V.N.S.H., Sai Chandra
R., Indian sign language recognition: A comparison
between ANN and FIS ,2016, Journal of Theoretical
and Applied Information Technology, Vol: 89, Issue: 2,
pp: 502 - 511, ISSN 19928645

4027
IJSTR©2020
www.ijstr.org