Vous êtes sur la page 1sur 24

Emotion Analysis of Twitter

using Opinion Mining


Akshi Kumar

Prakhar Dogra Vikrant Dabas

Department of Computer Science


Delhi Technological University

Table of Contents
S. No.
1
2
3
4
5

Topic
Our Motivation
Our Goal
Background Work
System Architecture
Results

Our Motivation
A vital part of the information era has been to find out the

opinions of other people.


What other people think? -> one question whose answer
can help shape effective policies and make astute & wellformed decisions
A need for political information is an important factor for
seeking public opinion

Our Goal
The goal of our research was to do emotion analysis on

tweets using opinion mining.


The main idea of the research was to create a model that
is able to predict the type of emotion of a tweet using web
mining to initially extract tweets using the Twitter API and
then further using sentiment analysis to analyze the
emotions.
Instead of just classifying the emotion of the tweets as
happy or sad; we introduced a five emotion model
[Happiness, Sadness, Fear, Anger, Disgust]

Why select Twitter?


It is used by people of different domains to express their

emotions, views and opinions.


It contains sheer amount of text posts that grows at large
rates everyday
It is used by people from many countries.
It involves variety of people from ordinary users to
celebrities of various fields like entertainment, sports,
politics, education etc.

Background Work
Kumar and Sebastian introduced a hybrid approach which

combines the advantages of both dictionary and corpus


based methods along with the combination of NLP and ML
based techniques to classify tweets into positive, negative
and neutral categories.
EmpaTweet describes an emotion corpus that is created
from the micro-blogging service Twitter. The corpus
contains seven different emotions(Anger, Disgust, Fear, Joy,
Love, Sadness and Happiness) annotated across 14 topics.

Generic Sentiment Analysis Model

System Architecture
It mainly involves:-

Tweet Retrieval Module


Preprocessing Module
POS Tagging
Emotion Scoring Module

Whats new -> Proposed Five-Emotion Model


Fear
Anger

Happiness
FIVE
EMOTION
MODEL

Sadness

Disgust

Retrieving
Tweets

OUTPUT: set of
downloaded
tweets

Preprocessing
Module

INPUT: set
downloaded
tweets

of

OUTPUT: set of
cleaned tweets

Scoring
Module

INPUT: set of
cleaned tweets

OUTPUT:
average emotion
score for all
tweets

Scoring Module
1.) If an adverb or verb is encountered after another adverb
or verb then we add it. We repeat this process till another
adjective isnt encountered.
2.) Now if the added value of adverbs or verbs is less than 0
i.e., negative, then for the upcoming adjective, we subtract
its value from 5 instead of multiplying.
3.) And if added value of verbs or adverbs is positive and >=
0.5 then multiply it with the upcoming adjective else
multiply 0.5 with upcoming adjective.
4.) Later these multiplication results are added and the sum
is divided by 5 times the number of adjectives encountered.

Retrieved
Tweet

Example
#RememberingRajiv Let us build an India, Proud of being independent;
Powerful in defense of her freedom

Cleaned Tweet
Let us build an India Proud of being independent Powerful in defense
of her freedom
Cleaned Tweet with stop words
removed and the tweet POSTagged
Proud independent Powerful

[<Happiness>, <Anger>, <Sadness>, <Fear>, <Disgust>]


[3.46, 1.48, 1.25, 1.81, 1.22]

Emotion Scores of the Tweet

Example:
For example we have a tweet: She is not very beautiful
This tweet is first converted to not very beautiful using the POS Tagger.
In this string, words are processed one by one.
In this example we see that the first word (not) encountered is an adverb. The
score is assigned to the sum value. (sum = -0.8)
Then second word is encountered. We see that it (very) is also an adverb.
The new sum is an addition of the scores from the previous word and the present
word. In this case the sum value is (sum = -0.8 + 0.4 = -0.4)
Third word encountered is an adjective. So the sum produced is multiplied with the
emotion scores of that adjective. If there are no adverbs or verbs before the
adjective then no such sum is multiplied.
In this case, the scores produced are ([<Happiness>, <Anger>, <Sadness>, <Fear>,
<Disgust>] = [0.098, 0.756, 0.736, 0.738, 0.764].
In case of more than one adjective-adverb/verb set, the scores of these sets are
averaged.

Average Score for 1000 Tweets #AAPHelpLine

Average Score for 10000 Tweets #AAPHelpLine

Results Obtained

Applications
This model has several applications in:-

Review related websites like movie or product reviews


Business Intelligence like knowing customer opinions or

trends
In knowing public opinions for political leaders and work
being done by the government

Scope for Future Work


Following is the list of functionalities that can be added to
this model:-

Sarcasm Detection Module


Spell Correction Module
Apriori Association
Dictionary Expansion Module

Thank You

Vous aimerez peut-être aussi