Content uploaded by Siddharth Shrotriya
Author content
All content in this area was uploaded by Siddharth Shrotriya on Aug 09, 2018
Content may be subject to copyright.
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 6.887
Volume 6 Issue VII, July 2018- Available at www.ijraset.com
807
©IJRASET: All Rights are Reserved
Smart Education System Developed by Sentiment
Analysis of Students Using PMM Neural Networks
Siddharth Shrotriya1, Nidhi Kumari2
1, 2Jazari Research Institute of Artificial Intelligence, India
Abstract: Deep learning can provide many improvements to the existing systems which can also help the society in having
reliable education management structures around them. Nowadays, students in SAARC nations follow a primitive education
curriculum which neither challenges nor completely utilise their abilities. In this research, we are proposing a unique and
manageable education system which can make India grow rapidly in near future as students are the backbone of any nation. We
have accomplished this proposal with the help of sentiment analysis techniques implemented through a new type of neural
network architecture, which we have named as PMM Networks. This system can be implemented not just by SAARC nations but
by many other developing nations to provide a desirable curriculum to their school and college level students. We have tested this
system on a group of students from several states of India and found great results and the students were also happy and eager to
attend colleges which will provide such curriculums. We have also developed a feasibility model for the Ministry of Education of
the nations to implement such a preferential system at the national level.
Keywords: Deep Learning, Machine Learning, Artificial Intelligence, Neural Networks, Activation Function, Education,
SAARC, Curriculum, Sentiment Analysis.
I. INTRODUCTION
Education is the key to growth of any developing nation. It is the responsibility of governing bodies to provide a fruitful and
challenging curriculum both at school and college level for the residential students. A preferential curriculum is made by
understanding the sentiments and abilities of the students and then deep learning helps in finding groups of such students from a
geographical area who can group and study together. Sentiment Analysis can help in understanding the behavioural pattern of any
individual by their online and offline habits. It is mainly used by online retailers to list suitable products for their customers or news
aggregators to provide useful news to their audience. Neural networks are widely used in finding patterns and helps in such type of
sentiment analysis. Most popular neural network in this field is LSTM (Long Short-Term Memory) networks [1]. But, these models
also have some drawbacks such as slowness in computation and not been able to remember high amount of sequences which are in
the order of 10,000’s. So, we are proposing a modified LSTM cell network which can remember only the desired cell states and
ignore the things which are not required for future predications. We have named this network structure as PMM or Preferential
Memory Model. These neural networks are solving the problem of Sentiment analysis [2] based on some personal and skill tests
taken on all the concerned students. Along with this, these networks are also learning the pattern among all the students in the
survey to group them in certain categories which can be further used for making custom batches in the respective institutions.
II. DATA COLLECTION
Input data is collected via a survey conducted online for the people from major cities of India. The aim of these survey is to perform
2 tests on the students. First test is a behaviour test which consist of 40 general questions and provides us the knowledge of real
desires of the students in relation to their career prospects. Second test is a skill test conducted after the first one which will analyse
if the student has a potential of pursuing the desired career. For example, a student who wants to be a pilot must have good eyesight
and student who wants to be an engineer must have the curiosity to learn old and new technologies. This data is then passed as input
to the first layer of the neural network. It’s used for both sentiment analysis and to build batches of students having likewise traits.
III. TECHNOLOGIES USED
Data clean-up and management is done using Hadoop Distributed File System and Resource Negotiator. Data Analysis and Deep
Learning part is done in Python programming language using third party libraries. The custom neural network is built in Keras
framework and TensorFlow libraries. Modified LSTM or PMM is made as a firmware script of Keras framework. I-python
Notebooks are used to perform Data visualisation and presentation.
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 6.887
Volume 6 Issue VII, July 2018- Available at www.ijraset.com
808
©IJRASET: All Rights are Reserved
IV. SENTIMENT ANALYSIS
Sentiment analysis is a type of data mining that measures the inclination of people’s opinions through natural language processing
(NLP), computational linguistics and text analysis, which are used to extract and analyse subjective information from the Web -
mostly social media and similar sources. The analysed data quantifies the general public's sentiments or reactions toward certain
products, people or ideas and reveal the contextual polarity of the information. That’s why it is also known as opinion mining. We
use this analysis to find the educational sentiments [3] of a person and what subjects he should learn to reach his desired level of
competence. A sample output of this technique looks like Fig. 1 and is used for finalizing the curriculum for like-minded students
having same set of subjects. A set of 7 subjects are chosen from a list of 100 possible students. Now a curriculum can have 5-7
chosen subjects for the secondary as well as the higher education of a student so that he can learn the preferred topics right from a
young age till graduation.
Fig. 1 A sample output from Sentiment Analysis.
V. SEMI SUPERVISED ENVIRONMENT
With more common supervised machine learning methods, you train a machine learning algorithm on a labelled dataset in which
each record includes the outcome information. This allows the algorithm to deduce patterns and identify relationships between your
target variable and the rest of the dataset based on information it already has. In contrast, unsupervised machine learning algorithms
learn from a dataset without the outcome variable. In semi-supervised learning, an algorithm learns from a dataset that includes both
labelled and unlabelled data, usually mostly unlabelled. In our model, we try to find the sentiments or actual educational
requirement of a student on uncategorised data which comes under unsupervised learning. Then we also find the common traits to
create batches of like-minded students which comes under supervised learning. Thus, our complete system follows semi-supervised
environment. [4].
VI. NEURAL NETWORK DESIGN
PMM (Preferential Memory Model) is built upon basic LSTM cell [5] while changing and adding some of the features. In a
traditional LSTM cell, the key is the cell state, the horizontal line running through the cell diagram. The cell state is kind of like a
conveyor belt. It runs straight down the entire chain, with only some minor linear interactions. It’s very easy for information to just
flow along it unchanged. The LSTM does have the ability to remove or add information to the cell state, carefully regulated by
structures called gates. Gates are a way to optionally let information through. They are composed out of a sigmoid neural net layer
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 6.887
Volume 6 Issue VII, July 2018- Available at www.ijraset.com
809
©IJRASET: All Rights are Reserved
and a pointwise multiplication operation. The first step in our PMM model is to decide what information we are going to throw
away from the cell state. This decision is made by a sigmoid layer called the forget gate layer. The next step is to decide what new
information we are going to store in the cell state. This has two parts. First, a sigmoid layer called the input gate layer decides which
values we will update. Next, a bipolar sigmoid layer creates a vector of new candidate values, that could be added to the state. In the
next step, we’ll combine these two to create an update to the state. While creating estimation vector, bipolar sigmoid is used in place
of usual Tanh activation to understand more information at a given timestep.
Fig. 2 PMM cell internal layer structure.
Preferential state line is added to keep the input gate in-check using an extra gate. If the input gate does not think that any
information [6] is worthy of storing because of its inability to remember long sequences, then preferential gate will add that
information by itself by accessing the longest possible sequences. It will remain idle if the default input gate is working fine and will
start triggering if some inputs are neglected which has preference while considering all the older contexts.
Finally, we need to decide what we are going to output. This output will be based on our cell state but will be a filtered version. First,
we run a sigmoid layer which decides what parts of the cell state we’re going to output. Then, we put the cell state through bipolar
sigmoid and multiply it by the output of the sigmoid gate, so that we only output the decided parts of the data. In addition to this, the
output is added to the Preferential state line so that it can keep track of all the information passing through the PMM cell shown in
Fig. 2.
Derivations of a PMM Cell:
ft =σ (Wf . [ht-1, xt] + bf)
it =σ (Wi . [ht-1, xt] + bi)
jt =σ (Wj . [ht-1, xt] + bj)
Ot =σ (Wo . [ht-1, xt] + bo)
ht=Ot+σσ (Ct)
mt= σ σ (Wm . [dt ,xt, Pt-1] + bo)
kt= ReLU (Wm . [dt ,xt, Pt-1] + bo)
dt = it * jt
Ct = ft * Ct-1 + dt + (mt * kt)
(σ) refers to sigmoidal activation
(σ σ) refers to bipolar sigmoidal activation
(ReLU) refers to Rectified Linear Unit Activation
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 6.887
Volume 6 Issue VII, July 2018- Available at www.ijraset.com
810
©IJRASET: All Rights are Reserved
VII. TEST CODE SNIPPETS
Firstly, the PMM cell is built as a custom layer in Keras. Then it is used to create our system with many such cells to form different
internal layers of the model. Default Class call in Keras syntax looks like Fig. 3 and sample usage of this cell to build a model is
done like Fig. 4.
Fig. 3. PMM Cell as a Keras Layer
Fig. 4. Code snippet for Neural Network model design
VIII. WORKING OF THE SYSTEM
Initial stage of our system starts with the collection of survey data from the students in consideration. As mentioned in Section II,
student information is collected via two dedicated tests. This information is used to build sentiment database of all the students using
Neural networks. These sentiments will give dedicated weights for 40 subjects currently taken into consideration. These subjects can
be increased as per the need of an institution. Then second phase of our model divides students in 20 types having 5-7 common
subjects which is a continuous learning process achieved by PMM Model. For Example, a sample output can be a group of 25
students has 7 subjects in common namely, Mathematics, Physics, Football, Dance, English, Physical training and Acting. Then
these students will learn the essential subjects as well as their preferred subjects and will become experts at an early age.
Such students will not only find easiness in studying a unique mix of subjects but will also have an upper hand on the students
having traditional curriculum in their schools and colleges. Third phase of the system is related to the implementation of such
complicated structure by the government at the national level. [7] As we have collected like-minded students, it will be easier for
governments to make their respective classes under an institution. Government can also have 20 different institutions for 20 types of
students as per the model. This can further help the students as now thousands of like-minded students selected from all over the
country can study together and prosper with great speed.
IX. OUTPUT USAGE
This model can be used by schools in secondary education and by college in graduation and post-graduation. At school level, it will
provide a curriculum as a combination of necessary and preferred subjects. In India, secondary education terminated at the 10th
standard Board examinations. So, our system will provide a custom Board Exam syllabus for the students having the preferred
curriculum. For example, a student who wants to be musician will have 3-4 core subjects such as mathematics, science/commerce, a
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 6.887
Volume 6 Issue VII, July 2018- Available at www.ijraset.com
811
©IJRASET: All Rights are Reserved
local/foreign language and the 2 subjects from music background as the syllabus to be taught in secondary level and in the
examination at 10th Board level.
At college level, we propose a new set of Degrees such as Bachelor of Preference, Master of Preference and Doctorate of Preference.
All these degrees will be tailored to the needs of a set of students. There will be 20 types of these degree and each type will have a
different syllabus. This syllabus will keep on improving monthly based on our deep learning outcomes. Each type will be the main
syllabus of a university, so only 20 only such institutes are required for teaching a wide range of subjects.
X. CONCLUSIONS
The Model is still in beta phase but promises huge benefits for people of both educational and political background. It invents the
use of PMM model for sentiment analysis. It also provides the needed change in primitive curriculums used in most of the
educational institutions in the developing nations. Training of the model takes about 2 days to get the final output when tested on the
dataset of 5 cities of India and ran through GeForce 1070 Ti processors.
In the continuation of this model, we will publish a stable release of the complete source code of this system as soon as possible. We
hope it will ignite a spark in the minds of government officials along with deep learning enthusiasts with some new and unique
combination of ideas proposed in the paper.
REFERENCES
[1] Felix A. Gers , Juan Antonio Pérez-Ortiz , Douglas Eck , Jürgen Schmidhuber, Learning Context Sensitive Languages with LSTM Trained with Kalman
Filters, Proceedings of the International Conference on Artificial Neural Networks, p.655-660, August 28-30, 2002.
[2] ASamy Bengio , Oriol Vinyals , Navdeep Jaitly , Noam Shazeer, Scheduled sampling for sequence prediction with recurrent Neural networks, Proceedings of
the 28th International Conference on Neural Information Processing Systems, p.1171-1179, December 07-12, 2015, Montreal, Canada.
[3] Yi Tay , Luu Anh Tuan , Siu Cheung Hui, Dyadic Memory Networks for Aspect-based Sentiment Analysis, Proceedings of the 2017 ACM on Conference on
Information and Knowledge Management, November 06-10, 2017, Singapore, Singapore.
[4] Rie Johnson , Tong Zhang, Supervised and semi-supervised text categorization using LSTM for region embeddings, Proceedings of the 33rd International
Conference on International Conference on Machine Learning, June 19-24, 2016, New York, NY, USA.
[5] Martin Wöllmer , Björn Schuller, Enhancing spontaneous speech recognition with BLSTM features, Proceedings of the 5th international conference on
Advances in nonlinear speech processing, November 07-09, 2011, Las Palmas de Gran Canaria, Spain.
[6] Hang Li , Zhengdong Lu, Deep Learning for Information Retrieval, Proceedings of the 39th International ACM SIGIR conference on Research and
Development in Information Retrieval, July 17-21, 2016, Pisa, Italy.
[7] Gautam, Mohan & Singh, Sunny & Fartyal, Gopal & Tiwari, Ankit & Singh Arya, Kuldeep. (2016). Education System in Modern India. International Journal
of Scientific Research And Education. 10.18535/ijsre/v4i01.16.
