Real Time Sentiment Analysis of

Student’s Feedback
Abstract

Educational Data Mining (EDM) plays integral role in the improvement of education by keeping a check on the performance
of the student in his/her studies and by trying to understand learning of the students. Conducting feedbacks from students
toward the year's end/semester has an issue and disadvantage, as the students who previously taken the course are not getting
the benefit of the course as their issues were not solved. In order to benefit the students, those who are currently taking the
course, their feedback should be taken in real time and their issues must be resolved in real time. This is a beneficial
approach as this allows the teacher and students to resolve the teaching and learning issues. Analyzing the student’s
comments and feedbacks by using the sentiment analysis techniques helps to identify student’s positive, neutral or negative
review towards the current teaching methodology adopted by teacher. In this article, we have proposed a system that will
analyze the student’s feedback in real time. Furthermore, based on the proposed solution it will also help the teacher to
benefit the students by managing the conflicts in an efficient way. Lastly this technique will help to depicts more accurate
results then previous solutions.

INDEX TERMS Sentiment analysis, Data Mining, Educational Data Mining, Student’s Feedback,
teaching methodology, Education, positive, negative, neutral, polarity.

I. Introduction analysis is to classify the polarity of the feedback or

Sentiment Analysis in education is developed to opinion given in the form of textual data as positive,
address the problems in education sector. Taking input negative or neutral [3]. Feedbacks from students are
from students to instructor permits the students to usually conducted at the end of semester or at the end
direct the educator into showing the course in a best of the term [1], [3], [5]. Gathering feedbacks from
possible that the can understand the lecture [1], [2], [3]. students toward the finish of semester has a weakness
Sentiment analysis is a field that works on the basis of of not helping the students those who are taking the
to make sense of a textual comment/feedback/opinion. course or that have already taken the course.
Comments or opinions can be positive or can be
negative[1], [4]. Analyzing student’s given feedback This paper aims to address the problem, creation of a
using the techniques of Sentiment Analysis will be system is proposed that will analyze the student’s
helpful to identify positive, negative or on the other feedback in real time and present those feedbacks to
hand significantly more refined practices of students, the teacher. In order to benefit the students who are
they have towards the current instructing [1]. The currently taking the course the feedback from students
feedbacks provided by students can highlight different should be taken in real time and their issues must be
issues students are facing with the lecture or with the addressed in real time. This will be help the teacher to
teacher. Most of the times students don't comprehend benefit students by resolving their issues in a beneficial
the subject teacher is attempting to clarify, they are not way. Analyzing the feedback provided by students by
getting the topic because the teaching methodology using sentiment analysis techniques will help to
adopted by teacher is very difficult to understand, by identify student’s positive, negative and significantly
giving feedbacks students can deliver their issue to the more refined sentiments that students have towards the
teacher [4]. current instructing methodology adopted by teacher.
Educational institute in order to improve the quality of This technique will also help to depicts more accurate
education takes the feedback from students [1]. The results then previous solutions.
comments of students then analyzed to read the
sentiments of the students to enhance and improve the The next section of this paper is formulated: Section II
teacher’s teaching performance [1]. The main focus of which described the existing literature work that is
sentiment analysis is to determine the feelings of done and what this paper aims to provide. In the follow
students given in the form of text. That feeling might up of Section III the description of the method is given
be student’s opinion or attitude. The vital role of the and how each step would perform its part in order to

VOLUME XX, 2017 1

 obtain the desired result that is required. Section IV either they are positive or negative provided by
provides the construction of sentiment word database students as feedback. Sentiment analysis helping the
and how it’s done. After wards the paper conclusion is students’ those who have issues to forward their
done. concerns to the teacher so that he can improve his/her
teaching methodology as a result students’ get quality
education and understand the lecture clearly [6].
II. Related Work
According to Chauhan, Agrawal [2] almost all of the In the light of above mentioned studies, it is stated that
educational institutes, most commonly universities the educational institutes used to get the feedback
showing a lot of interest to work on the nature of toward the finish of semester or year. Taking the
training by keeping check and balance or by observing feedback toward the finish of semester has an
teacher’s instructing, student’s learning by analyzing impediment that it isn't helping the understudies the
the feedback provided by students at the end of students who are taking the course or that have
semester. However, the feedback got at the end of effectively taken the course. To help the students who
semester is not enough. To improve the teaching- are currently taking the course the feedback from
learning process the greatest contribution of students, students should be taken in real time and their issues
instructors and different partners are needed to must be addressed in real time. This will be help the
recognize various parts of schooling. Another class of teacher to benefit students by resolving their issues in a
study by Ullah [1] which states the efforts and attempts beneficial way. Analyzing the feedback provided by
of the educational institutes to gather the student’s students by using sentiment analysis techniques will
feedbacks towards the teacher’s teaching in order to help to identify student’s positive, negative and
enhance the teaching performance of a teacher or significantly more refined feelings that students have
instructor. but they didn't get any accurate results. towards the current instructing methodology adopted
According to another by teacher. In brief this study is expected to contribute
study by Rajput, Haider [3] the feedback is an essential to benefit the students who are currently studying a
part of education to improve the quality of teacher and course and facing issues to understanding a lecture by
learning. Students express their issues by giving getting the feedback in real time and solving the issues
different comments/feedbacks these comments are in real-time.
analyzed by using TF-IDF and lexicon-based features
to analyze the sentiments expressed by students in their III. Material and Methods
textual feedback but the students don’t like long In this study, Long Short-Term Memory model
questionnaires to be filled and give their comments in (LSTM) model is implemented for the sentiment
form of writing analysis and text classification. Long Short-Term
Memory (LSTM) networks are a form of recurrent
The feedback system became very useful in the neural community able to getting to know order
educational domain, where the teacher delivers a dependence in collection prediction problems. This is a
lecture and student’s share their opinions towards that behavior required in complicated hassle domain names
lecture. The opinions are analyzed to address the like gadget translation, speech recognition, and more
difficulties of students by using sentiment analysis, an algorithm is designed for this purpose, explaining
where sentiment analysis is a computational study of the architecture of the system shown in Figure 1.
people’s like and their dislike about an entity or about System is divided into five phases 1) First phases is
an object also revealing the polarity of the natural Data preprocessing. 2) Second is word embedding, 3)
sentences/word e.g. positive, negative or neutral [1]. LSTM model for the hypothesis testing for the
Also analyzing students’ comments using sentiment prediction accuracy, 4) Dense layer in order to increase
analysis approaches can help to classify that the model complexity, 5) Softmax function for multi-class
students’ feelings are positive or negative. Students’ classification problems.
can mention or highlight the issues they faced during
Data Preprocessing
lecture, through their feedbacks they can indicate their
issues to the lecturer. In the domain of education
sentiment analysis is used to extract the hidden Embedding Layer
knowledge from students through a feedback The
evaluation of teacher and class by students’ now is
considered a usual practice in the educational Embedding Layer
institutes. This process helps to provide quality
education and helps the instructor to distill their
teaching methodology. The sentiment analysis Dense
techniques are used to identify the textual contents that
Softmax
VOLUME XX, 2017 1

.8 .7 .1 …………... .9
 The hyper parameters are, LSTM layer has 196 nodes,
it is the result aspect of word vector. In order to
train model varies parameters are used. The dropout
rate is 0.2 and softmax is used as activation function.
Adam optimization function with batch size 64 is used
Figure 1 Sentiment Analysis Model Structure. for model training, on the thick layer softmax
quickening work for multi-arrangement is utilized. In
order to avoid overfitting dropout regularization is
LSTM is A type of recurrent neural network which is used. The text feedbacks that are inputs entered to the
capable of learning order dependence in a sequence embedding layer, converting individual word to 300-
prediction problem is call Long Short-Term Memory dimensional vector.
(LSTM) network. The representation of a sentence is
done by LSTM model in a sequential manner. Word The word embedding layer parameters are
vector works as an input for one stage that fed the maximum features, embedding features, input
LSTM layer and in order to compute the next hidden length and vector is entered to LSTM model.
state, the previous closed state is taken to LSTM. The LSTM layer forwards the final output to the dense
fundamental benefits for the execution of LSTM for layer for the predictions of output. Categorical
sentence vector is that to out the decent length sentence cross-entropy for multiclass sentiment
vector for any irregular variable length sentences. It classification is used as the loss function.
additionally stores the request for word and it has no
reliance on other etymological highlights in order to
compute semantics. The predictions are in RNN, It Ot
sequentially, which assign a memory to the network Input Output
The 3 elements of an LSTM cell are called gates. The LSTM block
first element is referred to as Forget gate, the second
Ct tanh
one element is called the Input gate and the closing one
is the Output gate.The results got from the previous
predictions can help to improve the future predictions. tanh
LSTM provides an additional feature to RNN which
Ft
gives RNN a fine-grained control over the memory.
This feature is responsible to control that how much Forget
the current input concerns in the creation of new
memory, and which sections of memory plays
important role in output generation. The performance Figure 2 LSTM Model Block Diagram
of the model is improved by word2vec without any
large supervised preparing set. The following equation
shows the course of LSTM model where sigmoid
strategic function is used and there are some basic IV. Experiments and Results
gates, input gate controls how much new inputs are Nowadays, internet has become one of the major
added into model, forget gate shows the amount of old source for individuals to communicate their
information sent by past hidden state and the output sentiments. Clients are currently more ready to
gate is used to determine the amount of influence of impart and communicate their insights or inputs
current node to the external network. Following are the online [8]. There is a benefit as an ever increasing
equations involved in order to compute the values of number of suppositions can be separated from a
LSTM: more extensive scope of source. In data mining,
interpersonal organizations have been utilized for
quite a long time [9]. This gave such countless
great benefits in utilizing online media, for
example, twitter as twitter is forward-thinking and
offers data about current news and occasions
happening all around the globe [10]. In this project
the information must be continuously and for this
reason twitter will be utilized The tweets of the
clients are gathered from the twitter based on that
info indicated by client is as Hashtags [10]. The
course of tweet order is started first by the
collections of the tweets. It is feasible to gather the
information of twitter by utilizing a twitter API.

VOLUME XX, 2017 1

 RAuth library is utilized in performing validation model. Also supports the structure of bag of
by giving in keys. A consumer key, access token, words.
consumer secret and access token mystery for
twitter application and perform Handshake C. Long Short Term Memory (LSTM)
protocol. After this, a certificate downloaded and We examined the effectiveness of this technique
to get to the tweets a PIN is produced for the using the results extracted from twitter. The end
application. result of all feedback is assessed as strongly bad,
or fairly bad, or weakly bad, or strongly high-
A. Data Preprocessing quality, or fairly high-quality, or weakly high-
Once the data is obtained from the source, the data quality or impartial. This phase gives the
has to be pre-processed in order to extract useful development of sentiment phrase database. The
information. This can be done by removing the lexicon may be created both manually or
spelling mistakes, grammar mistakes. In this increasing mechanically from a seed of phrases. In
research the following steps for the pre-processing our LSTM Model the synonym is taken into
are considered. The process of filtration, in this consideration as a high-quality phrase and
process the punctuations, numbers and other antonym is taken into consideration as a bad
special characters or symbols are removed because phrase. There are 745 phrases in sentiment phrase
they are not giving some useful information, but database. Total intensifier phrases are 24, overall
on the other hand creates uncertainty and doubts high-quality phrases are 448 phrases and overall
which can mislead from the context. The second bad phrases are 263 phrases, the final phrases are
process, is the process of tokenization which is impartial phrases. The sentiment rating levels from
used word categorization from the sentence, -three to +three. The rating starting from 1 to a
breaking the sentence into words. The case few are taken into consideration as high-quality;
convention process which is used after the whereas any bad rating starting from -1 to -three.
filtration and tokenization process, the tokenized The sentiment rating equals to 0 is taken into
words are mutate into lower case. The articles in consideration as impartial category. Some example
the sentences are referred as stop words. After the phrases are proven in Table 1.
process of filtration, the tokenization process and
case conversion process are two processes which
Table 1: SAMPLE WORDS IN SENTIMENT WORD DATABASE
plays vital role in the removal of stop words.
These are used for the word connection and to Opinion Score Description
make understanding for sentences. Through stop word
words we are able to remove non-relevant care +2 verb
information by categorizing the article words into
useful +2 adjective
a meaningful context. As a result, the model
accuracy increases through the polarity assigned to fast -1 adjective
words. Daily 0 Adjective

B. Word Embedding Easily +25% Intensifier

Word Embedding is a form of word Most +100% Intensifier
representation, allowing words having similar
interpretation to have a same representation. In a Brilliant +3 adjective
document, the word embedding is the main Joyful +1 adjective
portrayal of words. The main focus is in
maintaining the relationship of words, capture the Confuse -3 Verb
context of word in a document in order to Sleepy -2 Adjective
distinguish the semantic and syntactic
Really +25% Intensifier
comparability. In our proposed model word
process is a pre-prepared word vector and it is Little -50% Intensifier
given as an input to the Long Short-Term Memory
(LSTM) network. First of all, these vectors are
given to Word2Vec model of Google and in the
second step they are pre-trained with the given
model on the updated datasets of Google news of
100 billion. A 300-dimentioanal vectors for 3
million words, phases are produced by Word2vec

VOLUME XX, 2017 1

 messages due to the fact. In this framework, we’ve
used LSTM version that could identify hashtags
and after that the results will be filtered out.

Same phrase in distinct domain names will have

distinct meanings. In our LSTM ‘fast’ is negative
opinion and its polarity rating is -1. For example,
‘I sense IPAD2 is fast running pace and better
display resolution’. In this sentence ‘fast’ is the
nice opinion in phone review evaluate area
however poor opinion in coaching evaluate area.
For example, ‘I recognize her coaching however
pace of presentation could be very fast for
different college students’. In this
sentence ‘fast’ is poor opinion (-1) and ‘very’
depicts intensifier phrase that improved with the
aid of using 50%. The sentiment phrase database
includes the rankings of the polarity of each
phrase. Some phrases aren't protected in current
LSTM. The polarity of some phrases written with
the aid of using college students in coaching
evaluate wishes to feature in sentiment Analysis.

Conclusion
The work defined in this paper is a step forward in
direction of quick filtration of hashtags. Short
messages are more difficult to categories than
large corpus of textual content. This is usually due
to the fact there are few phrase occurrences and
therefore it's very hard to capture the semantics of
such messages. Hence, conventional tactics like
“Bag-Of-Words” which were implemented to
classify short texts or hashtags now no longer
carry out in addition to expected. Existing works
on class of quick textual content messages
combine messages with meta-facts from different
facts reasserts including Wikipedia and WordNet.
Automatic textual content class and hidden topic
extraction tactics carry out properly while there's
meta-facts or while the context of the fast textual
content is prolonged with know-how extracted the
usage of massive collections. But those tactics
require on-line querying which may be very time-
eating and undeserving for actual time
applications. We have proposed a framework to
categories Twitter messages which serves an
excellent candidate for quick textual content

VOLUME XX, 2017 1