Technical Answers for Real World

Problems (TARP)

Project title:

SENTIMENT ANALYSIS OF USERS USING RNN LSTM MODEL

Team Members:

Sno. Name Reg No

1 Tushar Krishnan 19BDS0158

2 Jogi Pavan Omkar 19BDS0101

3 Sathvick Sudarsan 19BDS0107

4 Kunal Samnani 19BCE2288

5 Vibhu Kumar Singh 19BCE0215

Given the current COVID status, the lack of socializing has affected most of us who live in
nucleated families. To many, it has proved to be a mental pressure. In a world driven by
technology, we slowly are losing personal connectivity. This, in turn, has been affecting our
mental health. This project aims to help better the same. The main aim is to understand the
emotions of society (users) during the pandemic and be able to predict their mindset. The
prospect lies in applying it to our life (Ex. on a social media platform, healthcare, etc.). In
this project, an RNN LSTM model will be applied to the data, and the respective mood
(positive or negative) is predicted.

(Keywords: RNN, LSTM, mental health, Covid-19 pandemic, technology, healthcare)

Introduction

From the end of the first quarter of the year 2020, we have been faced with a significant
threat globally; the spread of a pandemic-COVID. This is a virus that spreads when one
comes in contact with it and hence one must take all precautions necessary. The wearing of
masks, sanitising and cleaning our hands, and social distancing have been essential to
reduce the chance of being affected. While people are following these safety measures,
there is one thing that becomes difficult to follow; not going out of the house unless
necessary. Hence, a lot of people have moved digital. Living in such constrained
environments is becoming challenging for many. Consciously or subconsciously, people
are trying their best to adapt to the changes and keep up with them. Living constrained in
their house is one thing that has affected a lot of people, especially people that live alone or
in a nucleated family. This is because they always looked for entertainment outside the
house. Now that that is stopped, it becomes very difficult to enjoy. The lack of recreation
and socialization leads one to have a lot of time. One can use that effectively, or it can even
get the better of you. As very rightly said, "An empty mind is like that of the devil at
work." The availability of so much time has, in turn, led to more social media usage as well
as higher screen time. Under such circumstances, mental health has become a significant
concern. Mental health issues in our society are neither discussed very freely nor given
much importance. It is, in fact, considered they are insignificant and are at times even
treated as taboo. As mentioned in [1], Deep learning-based methods have recently
demonstrated a novel way to identify distinct prediction challenges. In order to forecast
COVID-19, two learning algorithms—deep learning and reinforcement learning—were
built in this study. In order to predict the number of newly affected people, losses, and
cures over the next several days, this article builds a model utilising recurrent neural
networks (RNN), specifically the modified long short-term memory (MLSTM) model. This
project aims to break that chain of thought by helping people understand that life is about
staying happy and that both physical as well as mental health are essential for all. The
project aims to monitor and help work on mental health based on peoples' posts on Social
Media which would not only help boost the happiness index but could also have a positive
impact on a person. There are many alternative applications to sentiment analysis in the
context of social media. With the increasingly extreme weather events and various
disasters, people are paying more attention to environmental issues than ever, particularly
global warming. Public debate on it has grown on various platforms, including newspapers
and social media. Sentiment analysis shows that [2] most people express positive emotions
about global warming, though the most evoked emotion tends to be found in fear, followed
by trust. With the current business environment and rapid changes in technology, the
amount of data produced is increasing as each day passes. This huge collection of data is
what can make or break such institutions, so it is vital for such a sector to efficiently utilise
the data generated. Effective tools and analyses are required to make sure that this data is
comprehended and organised in such a manner that it can be used for the tasks at hand [3] .
The challenge faced here is knowing how to extract and use the data to the benefit of the
business world. The objective of understanding the underlying emotion displayed in each
opinion that is voiced is a huge exercise. Through this paper, an attempt has been made to
understand how the gap between consumers and providers can be bridged by analysing
secondary data through the Sentiment Analysis tool. [4]

Literature Survey

The main approaches used for sentiment analysis are CNN [1-2,4,6-7,10,14-15,17-18],
RNN [3,6,9,10-12,16] , LSTM[1-4,6,8,11-12], Bi-LSTM[5,9,13] ,SVM[7]. The key
features of each are as follows. Lee et al. [5] is a paper on Predicting Emotions using RNN
LSTM. In this research, the Deep Learning model is applied to the movie review dataset.
This is used to predict emotions in the best way. It was also found that the predicted and
target emotions have higher levels of cosine similarity. Vikesh Kumar and Aditya
Timmaraju [6] have proposed an intelligent Recursive RNN model for classifying movie
reviews that classifies the sentiments present in the language (text). It is an extension of
sentence-level sentiment classification. Anand Kumar and Sachin Kumar [7] have
proposed a new system using CNN. Sentiment Analysis is done using NLP on a text that
contains information about emotions and opinions (feedback), which can then be
categorized into positive, negative, or neutral. Huang et al. [8] described a Sentiment
Classification model using RNN and LSTM, which converts part of speech to increase the
representation of sentence level achieving an accuracy of 76.6%. Hassan and Mahmood [9]
have shown that CNN can have the capability to learn higher-level features that cannot be
changed for local translation. They combined the CNN and RNN to capture the long-term
dependencies when the length of the input sequence was developed. Recurrent layers have
helped to reduce the loss of information and get the long-term dependencies more
effectively. Fenna Miedema [10] proposed an advanced system using LSTM that is used to
classify the tweets which are in text format. It was one of the most efficient models (until
that time). The authors of Akhtar et al. [11] proposed a CNN framework that uses a
sentiment word vector for practice and finally uses the softmax function to predict the
sentiment. It had produced better results than other baseline work for all datasets in various
other languages and domains. Chen and Zhang [12] proposed the CNNs-SVM model,
which analyses the sentiment based on data in the text. CNN's-SVM model outperformed
the traditional model. Xia and Zhang [13] have implemented a Text Emotion Analysis
using CNN, RNN, and LSTM NN and were successful in showing that the combination of
CNN and word2vec produced more accuracy than other models present for the analysis of
emotions from the text. Mark Ceilieba and JanDeriu [14] have implemented a system for
Twitter messages. It classifies the sentiments of Twitter messages using the deep learning
approach- using a 2-layer convolution neural network (CNN), dividing the entire task into
3 subtasks. Chatterjee et al. [15] a paper on Emotional Analysis applied CNN LSTM. The
proposed model was successful at detecting emotions like happy, angry, and sad from text
based on sentiment-related sentences and produced precise results. Alharbi and de Doncker
[16] have posed a model using CNN and LSTM, which gives higher accuracy (88.71%)
than other classification systems. Li Yiming et al.[17] has proposed a BiLSTM Sentiment
Classification model. It has successfully shown that an aspect layer is helpful for document
representation. It also shows that considering user information can boost sentiment
classification performance by a large margin. In the Meisheri et al.,[18] Sentiment
Extraction is done using Bi-LSTM. In this, the word, character level, and attention,
character level Bi-LSTM produced a higher performance on the SemEval dataset. For
WASSA, the word and dropout (avoid overfitting) with bias initializer were worked
together and gave the best result than the baseline model.Hassan et al.[19] shows sentence
classification based on sentiment analysis using CNN and RNN. It analyzed the sentence
classification of a short text by various models. He found that the combined model of CNN
and RNN produced more accuracy than the other model. He also concluded that to reduce
more number of CNN layers, LSTM has to place on top of the CNN model. Dragoni and
Giulio [20] have proposed a multi-domain sentiment analysis model using RNN and
LSTM. The model called NeuroSent for multi-domains was successful in obtaining better
results as compared to the baseline model on the Dranziera dataset. Lin et al. [21] cover a
CNN LSTM model applied for Sentiment Learning on cross-media content. It was found
that Explicit Emotion Signal (EES) for an image and text combination gave better results
than only on an image or only on text. Kratzwal d et al.[22] ran a paper on Emotional
recognition using RNN and Bi-LSTM. The proposed transfer learning method (sent2affect)
was used to affect computing. This, in turn, improved the performance and efficiency of
the sentiment analysis. Sentiment analysis on social media text gained a lot of popularity as
it contains recommendations and suggestions. Compared to high-resource languages such
as English, Chinese, French, etc., sentiment analysis task in low-resource language suffers
due to the absence of annotated corpus and tools to extract features [23]. Many proposed
models are not so successful in rightly classifying the long text data. On the other hand,
models incorporating LSTM networks are showing impressive results due to their
capability to deal with long text data. Hence, for this paper, we choose to do my
classification using Recurrent neural network (RNN), Long-short term memory (LSTM).
Through analysis in [24], we discover fresh possibilities for LSTM system enrichment and
merge these additions into the Vanilla LSTM network to create the most universal LSTM
version to date. The intended reader is receptive to a different educational strategy and has
already been introduced to RNNs and LSTM networks through a variety of available
resources. The conclusions and derivations in this treatise will also be helpful to a machine
learning practitioner looking for assistance on how to deploy our new enhanced LSTM
model in software for testing and study.

Proposed Methodology:
RNNs are powerful and robust neural networks, and belong to the most promising
algorithms in use because it is the only one with an internal memory.

Long short-term memory networks (LSTMs) are an extension for recurrent neural
networks, which basically extends the memory. Therefore it is well suited to learn from
important experiences that have very long time lags in between.

Sigmoid belongs to the family of non-linear activation functions. It is contained by the

There are three different gates in an LSTM cell: a forget gate, an input gate, and an output
gate.