Sentiment Analysis of Twitter tweets using supervised classification technique
0 likes85 views
The document discusses a sentiment analysis method for Twitter tweets, utilizing a supervised classification technique based on the Naive Bayes classifier. It involves collecting, preprocessing, and classifying tweets as positive, negative, or neutral to provide insights into public opinion on various topics. The results, displayed graphically, can assist consumers and corporations in understanding sentiment and making informed decisions.
![Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34
www.ijera.com 32 | P a g e
Sentiment Analysis of Twitter tweets using supervised
classification technique
Pranav Waykar, Kailash Wadhwani, Pooja More, Archana Kollu
Department Of Computer Engineering, DR.D.Y.Patil Institute Of Engineering And Technology,Pune.
ABSTRACT
Making use of social media for analyzing the perceptions of the masses over a product, event or a person has
gained momentum in recent times. Out of a wide array of social networks, we chose Twitter for our analysis as
the opinions expressed their, are concise and bear a distinctive polarity. Here, we collect the most recent tweets
on users' area of interest and analyze them. The extracted tweets are then segregated as positive, negative and
neutral. We do the classification in following manner: collect the tweets using Twitter API; then we process the
collected tweets to convert all letters to lowercase, eliminate special characters etc. which makes the
classification more efficient; the processed tweets are classified using a supervised classification technique. We
make use of Naive Bayes classifier to segregate the tweets as positive, negative and neutral. We use a set of
sample tweets to train the classifier. The percentage of the tweets in each category is then computed and the
result is represented graphically. The result can be used further to gain an insight into the views of the people
using Twitter about a particular topic that is being searched by the user. It can help corporate houses devise
strategies on the basis of the popularity of their product among the masses. It may help the consumers to make
informed choices based on the general sentiment expressed by the Twitter users on a product.
Keywords - Data Mining, Feature extraction Naïve Bayes Classifier, Natural language Processing, Twitter,
Unigram
I. INTRODUCTION
1
Twitter is a popular micro blogging
service where users create status messages (called
“tweets”). These tweets sometimes express
opinions about different topics. We propose a
method to automatically extract sentiment (positive
or neutral or negative) from a tweet. This is very
useful because it allows feedback to be aggregated
without manual intervention. Consumers can use
sentiment analysis to do a research on products or
services before making a purchase. Marketers can
use this to research public opinion of their
company and products, or to analyse customer
satisfaction. Organizations can also use this to
gather critical feedback about problems in newly
released products. There has been a large amount
of research in the area of sentiment classification.
Traditionally most of it has focused on classifying
larger pieces of text, like reviews. Tweets (and
micro blogs in general) are different from reviews
primarily because of their purpose: while reviews
represent summarized thoughts of authors, tweets
are more casual and limited to 140 characters of
text. Generally, tweets are not as thoughtfully
composed as reviews. Yet, they still offer
companies an additional avenue to gather feedback.
Previous research on analysing blog posts by Pang
et al. [3] have analysed the performance of
different classifiers on movie reviews. The work of
Pang et al. has served as a baseline and many
authors have used the techniques provided in their
work across different domains. In order to train a
classifier, supervised learning usually requires
hand-labelled training data.
With the large range of topics discussed
on Twitter, it would be very difficult to manually
collect enough data to train a sentiment classifier
for tweets. Hence, we have used publicly available
Twitter datasets. However, this dataset consist only
of positive and negative tweets. For neutral tweets,
we have used the publicly available neutral tweet
dataset provided. We run the machine learning
classifiers Naïve Bayes trained on the positive and
negative tweets dataset and the neutral tweets
against a test set of tweets.This can be used by
individuals and companies that may want to
research sentiment on any topic.
II. BACKGROUND
Defining the sentiment
For the purpose of this work, we define
sentiment as a positive or negative inclination of
the expression stated by the author. If the
expression doesn‟t bear any polarity, it is marked
as a neutral sentiment.
RESEARCH ARTICLE OPEN ACCESS](https://image.slidesharecdn.com/g060506032034-160811043440/85/Sentiment-Analysis-of-Twitter-tweets-using-supervised-classification-technique-1-320.jpg)
![Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34
www.ijera.com 33 | P a g e
Table 1: Example Tweets
Sentiment Keyword Tweet
Positive Weather The weather is pretty good
this morning!
Negative Work Dammnn…. I hate this
clerical work
Neutral Bus The bus arrives at 8 in the
evening.
Related Work
Topics related to the one discussed in this
work, have been researched before. Alec Go, Richa
Bhayani and et al [4] classify tweets using unigram
features and the classifiers are trained on data
obtained using distant supervision. Radha N and et
al [5] shows that using emoticons (distant
supervision) as labels for positive and sentiment is
effective for reducing dependencies in machine
learning techniques and this idea is heavily used in
[4]. Pang and Lee [3] researched the performance
of various machine learning techniques in the
specific domain of movie reviews.
III. METHODOLOGY
A. Pre-processing
The Twitter language model has many
unique properties. These properties can be used to
reduce the feature space:
1) Usernames
In order to direct their messages users
often include Twitter usernames in their tweets.
A de facto standard is to include @ symbol
before the username (e.g. @towardshumanity).
A class token (AT_USER) replaces all words
that begin with @ symbol.
2) Usages of links:
Users very often include links in their
tweets. To simplify our further work, we convert
a URL like “http://tinyurl.com/cmn99f” to the
token “URL”.
3) Stop words:
There are a lot of stop words or filler
words such as “a”, “is”, “the” used in a tweet
which does not indicate any sentiment and hence
all of these are filtered out.
4) Repeated letters:
Tweets contain very casual
language. For example, if you search “hello” with
an arbitrary number of „o‟s in the middle (e.g.
helloooo) on Twitter, there will most likely be a
nonempty result set. I use pre-processing so that
any letter occurring more than two times in a row
is replaced with two occurrences. In the samples
above, these words would be converted into the
token “hello".
B. Feature Vector
After pre-processing the tweets, we get
features which have equal weights.
Unigram
Features which are individually enough to
understand the sentiment of a tweet is called as
unigram. For example, words like „good‟, „happy‟
clearly express a positive sentiment.
C. Classification
For the purpose of classification of tweets,
we make use of Naïve Bayes classifier. Naïve
Bayes is a probabilistic classifier based on Bayes‟
theorem. It classifies the tweets based on the
probability that a given tweets belongs to a
particular class.
We consider three classes namely,
positive, negative and neutral. We assign class c*
to tweet d where,
In this formula, f represents a feature and
ni(d) represents the count of feature fi found in
tweet d. There are a total of m features. Parameters
P(c) and P(f|c) are obtained through maximum
likelihood estimates, and add-1 smoothing is
utilized for unseen features. We have used the
Python based Natural Language Toolkit library to
train and classify using the Naïve Bayes method.
IV. EVALUATION
A. Training data
There are publicly available data sets of
Twitter messages with sentiment indicated by [4].
We have used a combination of these two datasets
to train the machine learning classifiers. For the test
dataset, we used 20 tweets collected run-time
during the execution.
B. Experimental Setup
The Twitter API has a parameter that
specifies which language to retrieve tweets in. We
always set this parameter to English (en). Thus, our
classification will only work on tweets in English
because the training data is English-only.
We build a web interface which searches
the Twitter API for a given keyword for the past
one day or seven days and fetches those results
which is then subjected to pre-processing. These
filtered tweets are fed into the trained classifiers
and the resulting output is then shown as a graph in
the web interface.](https://image.slidesharecdn.com/g060506032034-160811043440/85/Sentiment-Analysis-of-Twitter-tweets-using-supervised-classification-technique-2-320.jpg)
![Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34
www.ijera.com 34 | P a g e
V. RESULTS
When a keyword was entered into the
search box, the tweets about the entered topic were
collected and classified. For the purpose of testing
the application, we searched tweets about “Donald
Trump”. 20 tweets were shown to the user and
classified using Naïve Bayes classifier. The result
of the classification was displayed in the form of a
pie-chart as follows:
Figure 1: Pie-chart
Once the results were displayed, the user
was asked if he wants to see the segregated tweets
for the sake of justification. The accuracy of the
results depend on the number of training tweets
being fed to the classifier. Higher the number of
tweets greater is the accuracy.
Now, this result about “Donald Trump”
can be used by voters and political analysts alike.
Voters can use the data to see the positive as well
as negative aspects of Mr.Trump whereas the
political analysts and psephologists can use it to
make their predictions.
VI. FUTURE WORK
Machine learning techniques perform well for
classifying sentiment in tweets. We believe the
accuracy of the system could be still improved.
Below is a list of ideas we think could help the
classification:-
A. Semantics
The polarity of a tweet may depend on the
perspective you are interpreting the tweet from. For
example, in the tweet “Federer beats Nadal :)”, the
sentiment is positive for Federer and negative for
Nadal. In this case, semantics may help. Using a
semantic role labeler may indicate which noun is
mainly associated with the verb and the
classification would take place accordingly. This
may allow “Nadal beats Federer :)” to be classified
differently from “Federer beats Nadal :)”.
B. Internationalization
Currently, we focus only on English
tweets but Twitter has a huge international
audience. It should be possible to use our approach
to classify sentiment in other languages with a
language specific positive/negative keyword list.
VII. CONCLUSION
A live Twitter feed is collected under the
keywords entered by the user. The feed is stored
locally in a json file. The data is pre-processed to
remove unnecessary spaces, symbols and useless
features. It still requires further work to remove as
much noise as possible. 20 tweets are then stored as
a csv file for analysis. A number of Lexicon based
methods are utilised on individual tweets from the
file to assess their usefulness. The chosen classifier
for this work is a Naive Bayes Classifier utilising
the text processing tools in NLTK and their
capacity to work with human language data. It is
trained on tagged tweets and then used to analyse
the sentiment in the tweets about the searched
topic. The result is represented in the form of a pie
diagram which shows the percentage of users who
have positive opinion on the searched topic as
compared to the ones have negative opinion or are
neutral.
REFERENCES
[1]. Adam Tsakalidis, Symeon Papadopoulos,
Alexandra Cristea, Yiannis Kompatsiaris,
“Predicting Elections for Multiple Countries
Using Twitter and Polls),” IEEE. 2015.
[2]. Gayo-Avello, Daniel, A meta-analysis of
state-of-the-art electoral prediction from
Twitter data, Social Science Computer
Review, 2013.
[3]. B. Pang, L. Lee, and S. Vaithyanathan.
Thumbs up? Sentiment classification using
machine learning techniques. In Proceedings
of the Conference on Empirical Methods in
Natural Language Processing (EMNLP),
pages 79–86, 2002.
[4]. Alec Go, Richa Bhayani, Lei Huang. Twitter
Sentiment Classification using Distant
Supervision. Technical report, Stanford
Digital Library Technologies Project, 2009.
[5]. Jiawei Han, Micheline Kamber, “Data
mining: concepts and techniques", Morgan
Kaufmann Publisher, second edition, pages
310-317.
[6]. Publicly available Twitter dataset -
http://www.sananalytics.com/lab/twitter-
sentiment/sanders-twitter-0.2.zip.
[7]. Steven Bird, Ewam Klein, Edward Loper,
“Natural Language Processing with Python”,
O‟Reilly, 2009.
[8]. Efthymios Kouloumpis, Theresa Wilson,
Johanna Moore, “Twitter Sentiment
Analysis: The Good the Bad and the
OMG!”, AAAI Conference on Weblogs and
Social Media.](https://image.slidesharecdn.com/g060506032034-160811043440/85/Sentiment-Analysis-of-Twitter-tweets-using-supervised-classification-technique-3-320.jpg)
Sentiment Analysis of Twitter tweets using supervised classification technique
- 1. Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34 www.ijera.com 32 | P a g e Sentiment Analysis of Twitter tweets using supervised classification technique Pranav Waykar, Kailash Wadhwani, Pooja More, Archana Kollu Department Of Computer Engineering, DR.D.Y.Patil Institute Of Engineering And Technology,Pune. ABSTRACT Making use of social media for analyzing the perceptions of the masses over a product, event or a person has gained momentum in recent times. Out of a wide array of social networks, we chose Twitter for our analysis as the opinions expressed their, are concise and bear a distinctive polarity. Here, we collect the most recent tweets on users' area of interest and analyze them. The extracted tweets are then segregated as positive, negative and neutral. We do the classification in following manner: collect the tweets using Twitter API; then we process the collected tweets to convert all letters to lowercase, eliminate special characters etc. which makes the classification more efficient; the processed tweets are classified using a supervised classification technique. We make use of Naive Bayes classifier to segregate the tweets as positive, negative and neutral. We use a set of sample tweets to train the classifier. The percentage of the tweets in each category is then computed and the result is represented graphically. The result can be used further to gain an insight into the views of the people using Twitter about a particular topic that is being searched by the user. It can help corporate houses devise strategies on the basis of the popularity of their product among the masses. It may help the consumers to make informed choices based on the general sentiment expressed by the Twitter users on a product. Keywords - Data Mining, Feature extraction Naïve Bayes Classifier, Natural language Processing, Twitter, Unigram I. INTRODUCTION 1 Twitter is a popular micro blogging service where users create status messages (called “tweets”). These tweets sometimes express opinions about different topics. We propose a method to automatically extract sentiment (positive or neutral or negative) from a tweet. This is very useful because it allows feedback to be aggregated without manual intervention. Consumers can use sentiment analysis to do a research on products or services before making a purchase. Marketers can use this to research public opinion of their company and products, or to analyse customer satisfaction. Organizations can also use this to gather critical feedback about problems in newly released products. There has been a large amount of research in the area of sentiment classification. Traditionally most of it has focused on classifying larger pieces of text, like reviews. Tweets (and micro blogs in general) are different from reviews primarily because of their purpose: while reviews represent summarized thoughts of authors, tweets are more casual and limited to 140 characters of text. Generally, tweets are not as thoughtfully composed as reviews. Yet, they still offer companies an additional avenue to gather feedback. Previous research on analysing blog posts by Pang et al. [3] have analysed the performance of different classifiers on movie reviews. The work of Pang et al. has served as a baseline and many authors have used the techniques provided in their work across different domains. In order to train a classifier, supervised learning usually requires hand-labelled training data. With the large range of topics discussed on Twitter, it would be very difficult to manually collect enough data to train a sentiment classifier for tweets. Hence, we have used publicly available Twitter datasets. However, this dataset consist only of positive and negative tweets. For neutral tweets, we have used the publicly available neutral tweet dataset provided. We run the machine learning classifiers Naïve Bayes trained on the positive and negative tweets dataset and the neutral tweets against a test set of tweets.This can be used by individuals and companies that may want to research sentiment on any topic. II. BACKGROUND Defining the sentiment For the purpose of this work, we define sentiment as a positive or negative inclination of the expression stated by the author. If the expression doesn‟t bear any polarity, it is marked as a neutral sentiment. RESEARCH ARTICLE OPEN ACCESS
- 2. Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34 www.ijera.com 33 | P a g e Table 1: Example Tweets Sentiment Keyword Tweet Positive Weather The weather is pretty good this morning! Negative Work Dammnn…. I hate this clerical work Neutral Bus The bus arrives at 8 in the evening. Related Work Topics related to the one discussed in this work, have been researched before. Alec Go, Richa Bhayani and et al [4] classify tweets using unigram features and the classifiers are trained on data obtained using distant supervision. Radha N and et al [5] shows that using emoticons (distant supervision) as labels for positive and sentiment is effective for reducing dependencies in machine learning techniques and this idea is heavily used in [4]. Pang and Lee [3] researched the performance of various machine learning techniques in the specific domain of movie reviews. III. METHODOLOGY A. Pre-processing The Twitter language model has many unique properties. These properties can be used to reduce the feature space: 1) Usernames In order to direct their messages users often include Twitter usernames in their tweets. A de facto standard is to include @ symbol before the username (e.g. @towardshumanity). A class token (AT_USER) replaces all words that begin with @ symbol. 2) Usages of links: Users very often include links in their tweets. To simplify our further work, we convert a URL like “http://tinyurl.com/cmn99f” to the token “URL”. 3) Stop words: There are a lot of stop words or filler words such as “a”, “is”, “the” used in a tweet which does not indicate any sentiment and hence all of these are filtered out. 4) Repeated letters: Tweets contain very casual language. For example, if you search “hello” with an arbitrary number of „o‟s in the middle (e.g. helloooo) on Twitter, there will most likely be a nonempty result set. I use pre-processing so that any letter occurring more than two times in a row is replaced with two occurrences. In the samples above, these words would be converted into the token “hello". B. Feature Vector After pre-processing the tweets, we get features which have equal weights. Unigram Features which are individually enough to understand the sentiment of a tweet is called as unigram. For example, words like „good‟, „happy‟ clearly express a positive sentiment. C. Classification For the purpose of classification of tweets, we make use of Naïve Bayes classifier. Naïve Bayes is a probabilistic classifier based on Bayes‟ theorem. It classifies the tweets based on the probability that a given tweets belongs to a particular class. We consider three classes namely, positive, negative and neutral. We assign class c* to tweet d where, In this formula, f represents a feature and ni(d) represents the count of feature fi found in tweet d. There are a total of m features. Parameters P(c) and P(f|c) are obtained through maximum likelihood estimates, and add-1 smoothing is utilized for unseen features. We have used the Python based Natural Language Toolkit library to train and classify using the Naïve Bayes method. IV. EVALUATION A. Training data There are publicly available data sets of Twitter messages with sentiment indicated by [4]. We have used a combination of these two datasets to train the machine learning classifiers. For the test dataset, we used 20 tweets collected run-time during the execution. B. Experimental Setup The Twitter API has a parameter that specifies which language to retrieve tweets in. We always set this parameter to English (en). Thus, our classification will only work on tweets in English because the training data is English-only. We build a web interface which searches the Twitter API for a given keyword for the past one day or seven days and fetches those results which is then subjected to pre-processing. These filtered tweets are fed into the trained classifiers and the resulting output is then shown as a graph in the web interface.
- 3. Pranav Waykar et al. Int. Journal of Engineering Research and Applications www.ijera.com ISSN: 2248-9622, Vol. 6, Issue 5, (Part - 6) May 2016, pp.32-34 www.ijera.com 34 | P a g e V. RESULTS When a keyword was entered into the search box, the tweets about the entered topic were collected and classified. For the purpose of testing the application, we searched tweets about “Donald Trump”. 20 tweets were shown to the user and classified using Naïve Bayes classifier. The result of the classification was displayed in the form of a pie-chart as follows: Figure 1: Pie-chart Once the results were displayed, the user was asked if he wants to see the segregated tweets for the sake of justification. The accuracy of the results depend on the number of training tweets being fed to the classifier. Higher the number of tweets greater is the accuracy. Now, this result about “Donald Trump” can be used by voters and political analysts alike. Voters can use the data to see the positive as well as negative aspects of Mr.Trump whereas the political analysts and psephologists can use it to make their predictions. VI. FUTURE WORK Machine learning techniques perform well for classifying sentiment in tweets. We believe the accuracy of the system could be still improved. Below is a list of ideas we think could help the classification:- A. Semantics The polarity of a tweet may depend on the perspective you are interpreting the tweet from. For example, in the tweet “Federer beats Nadal :)”, the sentiment is positive for Federer and negative for Nadal. In this case, semantics may help. Using a semantic role labeler may indicate which noun is mainly associated with the verb and the classification would take place accordingly. This may allow “Nadal beats Federer :)” to be classified differently from “Federer beats Nadal :)”. B. Internationalization Currently, we focus only on English tweets but Twitter has a huge international audience. It should be possible to use our approach to classify sentiment in other languages with a language specific positive/negative keyword list. VII. CONCLUSION A live Twitter feed is collected under the keywords entered by the user. The feed is stored locally in a json file. The data is pre-processed to remove unnecessary spaces, symbols and useless features. It still requires further work to remove as much noise as possible. 20 tweets are then stored as a csv file for analysis. A number of Lexicon based methods are utilised on individual tweets from the file to assess their usefulness. The chosen classifier for this work is a Naive Bayes Classifier utilising the text processing tools in NLTK and their capacity to work with human language data. It is trained on tagged tweets and then used to analyse the sentiment in the tweets about the searched topic. The result is represented in the form of a pie diagram which shows the percentage of users who have positive opinion on the searched topic as compared to the ones have negative opinion or are neutral. REFERENCES [1]. Adam Tsakalidis, Symeon Papadopoulos, Alexandra Cristea, Yiannis Kompatsiaris, “Predicting Elections for Multiple Countries Using Twitter and Polls),” IEEE. 2015. [2]. Gayo-Avello, Daniel, A meta-analysis of state-of-the-art electoral prediction from Twitter data, Social Science Computer Review, 2013. [3]. B. Pang, L. Lee, and S. Vaithyanathan. Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 79–86, 2002. [4]. Alec Go, Richa Bhayani, Lei Huang. Twitter Sentiment Classification using Distant Supervision. Technical report, Stanford Digital Library Technologies Project, 2009. [5]. Jiawei Han, Micheline Kamber, “Data mining: concepts and techniques", Morgan Kaufmann Publisher, second edition, pages 310-317. [6]. Publicly available Twitter dataset - http://www.sananalytics.com/lab/twitter- sentiment/sanders-twitter-0.2.zip. [7]. Steven Bird, Ewam Klein, Edward Loper, “Natural Language Processing with Python”, O‟Reilly, 2009. [8]. Efthymios Kouloumpis, Theresa Wilson, Johanna Moore, “Twitter Sentiment Analysis: The Good the Bad and the OMG!”, AAAI Conference on Weblogs and Social Media.