International Journal of Information Technology (IJIT) – Volume 7 Issue 2, Mar - Apr 2021

RESEARCH ARTICLE OPEN ACCESS

Automatic Answer Evaluation Using Machine Learning

Nandita Bharambe [1], Pooja Barhate [2], Prachi Dhannawat [3]
[1][2][3]
Information Technology Usha Mittal Institute of Technology
SNDT Women’s University Santacruz, Maharashtra
India

ABSTRACT
Nowadays, as we are moving towards automation there is a need for an automatic descriptive answer evaluation system.
Manual evaluation is a time and energy-consuming task. Presently, we have automated systems for objective type, single
sentence answers and descriptive answers with less accuracy level. In this paper, our goal is to develop an automated answer
evaluation system using machine learning. The system will evaluate the answer based upon the matched keywords and the
minimum length of the answer provided by the moderator. The scanned handwritten answer sheet will be given as input to
the proposed system. The system will use an artificial neural network with back-propagation algorithm.
Keywords: - OCR, NLP, ANN, Back-propagation algorithm.

I. INTRODUCTION
Normally educational or non-educational In our proposed system we are going to use ML
organizations evaluate a student's performance on the basis algorithm with ANN, OCR, back-propagation algorithm and
of examination either in the form of the subjective or cosine similarity.
objective pattern. Nowadays during this pandemic, everyone The system will convert the handwritten scanned
is working in a virtual manner. In the present scenario, document into machine-readable text using OCR. The
manual evaluation of subjective answers is a hectic task. handwritten character recognition process has two
Now, we have many systems which can evaluate objective categories –
type or MCQ questions. These techniques are evaluated in A. Online character recognition\newline
machines itself after providing a data-set of answers. But it B. Offline character recognition
helps only in competitive or objective type exam evaluation. Online character recognition[2][5] uses an optical pen for
The descriptive examination is the backbone of all the writing the characters. Then the characters are recognized at
universities. On the basis of the descriptive answer, the that time when we write with the optical pen and displayed
moderator will know how much knowledge the student has on the computer. Offline character recognition[2][5] extracts
gained. Manual evaluation of subjective answers is a very the text from a scanned handwritten or printed image.
tedious, time-consuming task and requires lots of manpower. In this system, we will use offline character recognition.
It also requires a high level of concentration. Answer We will train the model using the standard MNIST 0-9
evaluation varies from person to person according to their dataset by LeCun et al and the Kaggle A-Z dataset by
way of evaluation and interrelation between student and Sachin Patel datasets. The system will convert the
moderator. This affects the result of the student. handwritten scanned answer sheet into machine-readable
The project aims to automatically evaluate descriptive text using OCR and then it will extract keywords of answer
scanned answer sheet using ML and NLP. The system uses from the answer sheet. The cosine similarity is used in the
OCR to extract text and cosine similairity to measure the system to match these extracted keywords with keywords
similarity between extracted and provided keywords by the provided by the moderator. The system will use an artificial
moderator. neural network and a back-propagation algorithm to
After studying research papers[1] we have found that the calculate the error.
professors keep some important keywords in mind to
evaluate answers. So we have decided to propose a system II. LITERATURE REVIEW
in which we will evaluate descriptive types of questions. In
the proposed system, one has to scan an answer sheet in Optical Character Recognition Using Artificial Neural
PDF format. The moderator will provide the scanned answer Network, Sameeksha Barve[3] says that OCR is a process of
sheet, keywords for the answer and the minimum length of converting text obtained in the image which is scanned into
the answer. Our proposed algorithm will match these the machine- editable format. This paper trains the ANN
provided keywords with keywords extracted from the using a back-propagation algorithm which calculates the
answer using cosine similarity and it will also consider the error and modifies the weight. It represents each English
length of the answer as a parameter for evaluation of the letter by binary numbers that are used as input to a simple
answer. On the basis of the common keywords, the marks feature extraction system whose output, in addition to the
will be calculated. And the results are displayed in the range input, is fed to an ANN. Then, the Feed Forward Algorithm
1 to 10. gives insight into the inner workings of a neural network.

ISSN: 2454-5414 www.ijitjournal.org Page 1

 Evaluation of descriptive answer sheet using artificial b. Pre-processing - It is basically the operations that are
intelligence,Sk Asif Akram, Mousumi Saha, Tamasree applied to a digital image. The output of the pre-processing
Biswas[5] proposed a system based on Machine learning stage will be a normalized bitmap image. Pre-processing
and Natural language processing. The proposed system involves a number of following steps –
used Stemming algorithm, Stop word removal algorithm, • Binarization - It refers to the conversion of a gray-
Synonym generator and Gradient descent algorithm. scale image into a binary image.
• Noise reduction - It improves the quality of the
High accuracy Optical Character Recognition document.
algorithms using learning array of ANN, B.Vani, M. • Skew Correction - These method is used to align
Shyni Beaulah[2] OCR refers to recognition of handwritten the paper document with the coordinate system of
text to the machine-readable text which can be used for the scanner. Main methods for skew detection
searching, editing and indexing. This paper is using an involve correlation, projection profiles, Hough
artificial neural network with backpropagation algorithm transform.
for achieving high accuracy. The proposed system is • Slant removal - These method is used to normalize
implemented on a character database consisting of English all characters to a standard form.
characters, digits and special characters. The recognition c. Segmentation - The process of dividing pre-processing
rate is about 99.86. images into meaningful sub-images, units, such as words,
sentences, or topics. The techniques applied for image
Evaluating students descriptive answer using natural segmentation is the thresholding method, edge detection-
language processing and artificial neural networks, V. based techniques etc.
Lakshmi, Dr V. Ramesh[6] proposed a system which can d. Feature Extraction - It is the process of extracting the
evaluate descriptive answers using ANN and NLP. This important characteristics of an image. In the feature
system takes the dataset of answer and keywords from the extraction stage, each character is represented as a feature
moderators. The answer is first analyzed using NLP and the vector, which becomes its identity.
marks are stored in the database and then the same answer e. Classification and Recognition Using ANN - A feed-
is analyzed using ANN. PoS tagger, WordNet and forward back-propagation neural network is used in the
backpropagation were used. Then they are compared and process to classify and recognize characters.ANN
final marks are provided. The accuracy rate is about 96.75 classification is an example of a supervised learning
to 98.7%. algorithm.
f. Post-processing - The module will provide the
III. METHODOLOGY recognized character in a structured text form. It involves
The aim of the project is to propose a system which will data cleaning steps for documents that were digitized. It
evaluate student performance on the basis of the descriptive reduces the number of errors.
answers. In order to accomplish this, we will take B. Back-propagation Algorithm
handwritten scanned answer sheet of descriptive answers It is a method of finding the minimum value of the error
from students. The scanned answer sheet, keywords, function using weights. It is called back-propagation
minimum length of the answer will be provided by the because the error at layer n is dependent on the errors at the
moderator as input to the system. Some applications are next n+1 layer from the backward side. Weights of the
available for descriptive examination but they are not that network are updated with the gradient descent method or
accurate and they use different approaches to develop a partial differentiation of error with respect to weighted input
system. to neurons. The activation function of the algorithm
A. Text Extraction using OCR introduces non-linearity into the output of neurons. The
OCR, or Optical Character Recognition, is a process of activation function used for the model is ReLU.
recognizing text inside images and converting it into an C. Flowchart of the System
electronic form. These images could be of handwritten text, Figure 3.1 shows a flowchart of the Answer Evaluation
printed text like documents, receipts, name cards, etc., or using Machine learning. The system will accept a
even a natural scene photograph. Handwritten scanned image of the descriptive answer as an
input. OCR will extract text from the scanned answer sheet.
Steps involved for text extraction using OCR: The answer sheet goes through image preprocessing,
a. Image Acquisition - In this module, the system will segmentation which will give the bitmap image as output.
accept the scanned image as an input and image digitization Then features are extracted from the images and finally
takes place. Initially, some characters are written on a sheet ANN classifies and recognizes the characters. We get
of paper. Now the paper is scanned using a scanner. After machine-readable text using OCR. . Then the system will
the image is scanned, a bitmap image will be obtained. The calculate the length of the extracted text from the answer.
conversion of the handwritten document into an electronic Using NLP will extract keywords from the answer. After
format is termed Digitization. After digitization, an output extracting the keywords system will calculate the similarity
image i.e. bitmap image is given to the pre-processing stage percentage between provided and extracted keywords. If the
as an input. length of the answer is greater than or equal to the minimum
length provided by the moderator then the system will
display marks in the range of 0 to 10. And if the length of

ISSN: 2454-5414 www.ijitjournal.org Page 2

the answer is less than the minimum length provided by the V. CONCLUSION
moderator then the system will display marks in the range of This report gave the detailed information about the project
0 to 4. "Answer Evaluation Using Machine Learning". The
objective is to propose system which will automatically
evaluate the descriptive answer. The proposed system will
use OCR having artificial neural network with back-
propagation algorithm. The system will evaluate the answer
based on the scanned answer sheet, the keywords and length
provided by the moderator. The marks are assigned on the
basis of:
a. Number of keywords matched
b. The length of the answer.

VI. FUTURE WORK

In future, this model can be trained for many different
languages. We can provide data-sets of different handwritten
languages. Therefore, answer can be evaluated for languages
other than English. We can also train the system in such a
way that it can evaluate mathematical subjects.

REFERENCES
[1] Prince Sinha, Sharad Bharadia and Ayush Kaul,
“Answer Evaluation Using Machine Learning”, March
2018. Available: https://www.researchgate.net/
publication/333856264\_Answer\_Evaluation\_Using\_
Machine\_Learning [Accessed: September 19, 2020]
[2] Honey Mehta, Sanjay Singla and Aarti Mahajan,
"Optical Character Recognition (OCR) System for
Roman Script & English Language using Artificial
Fig. 3.1 Answer Evaluation System Neural Network (ANN) Classifier", in International
Conference on Research Advances in Integrated
Navigation System (RAINS - 2016), April 06 - 07,
D. Steps to evaluate the answer
2016.
• Provide a handwritten scanned answer sheet as an [3] B.Vani and M. Shyni Beaulah, "High accuracy Optical
input to the system. Character Recognition algorithms using learning array
• Provide keywords and the minimum length of the of ANN", in International Conference on Circuit, Power
answer. The system will extract keywords from the and Computing Technologies [ICCPCT] 2014, pp. 1474
given answer. The system will match extracted - 1479.
keywords with the provided keyword data-set. [4] Sameeksha Barve, "Optical Character Recognition
• The length of the answer will be calculated by Using Artificial Neural Network", in International
counting the words of the answer. Journal of Advanced Research in Computer
• Display the marks of the answer according to the Engineering \& Technology, Volume 1, Issue 4, June
following criteria. 2012, pp. 131-133.
[5] J. Pradeep, E. Srinivasan and S. Himavathi, "Neural
Network-based Handwritten Character Recognition
IV. RESULTS AND DISCUSSION system without feature extraction", in International
Worldwide in any educational system, every student's Conference on Computer, Communication and
performance is evaluated by conducting a series of test. Electrical Technology (ICCCET) 2011, 18th \& 19th
After research, we have found that there are many systems March, 2011, pp. 40-44.
available for objective type answer evaluation but only a [6] Sk Asif Akram, Mousumi Saha and Tamasree Biswas,
few are available for descriptive answer evaluation. To "Evaluation of descriptive answer sheet using machine
evaluate descriptive answer professors keep some important learning", in International Journal of Engineering
keywords in mind therefore in our system we are providing Sciences \& Research Technology (IJESRT), April
keywords of the answer to check whether the answer is 2019, pp. 184-186.
right. Our system takes an image of the answer sheet and [7] V. Lakshmi and Dr V. Ramesh, "Evaluating students
extracts keywords from the answer using OCR and cosine descriptive answer using natural language processing
similarity is used to compare extracted and provided and artificial neural networks", in International Journal
keywords from the answer sheet. On the basis of comparison, of Creative Research Thoughts (IJCRT), Volume 5,
our system will display the marks of the answer in the range Issue 4, December 2017, pp. 3168-3173.
of 1 - 10.
.

ISSN: 2454-5414 www.ijitjournal.org Page 3