By Aman Kharwal

Twitter Sentiment Analysis

Twitter Sentiment Analysis is┬áthe process of computationally identifying and categorizing tweets expressed in a piece of text, especially in order to determine whether the writer’s attitude towards a particular topic, product, etc. is positive, negative, or neutral.

In this Article I will do twitter sentiment analysis with Natural Language Processing using the nltk library with python.

Also, read – 10 Machine Learning Projects to Boost your Portfolio

Twitter Sentiment Analysis

Lets start with importing the libraries

import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
from sklearn.model_selection import train_test_split # function for splitting data to train and test sets

import nltk
from nltk.corpus import stopwords
from nltk.classify import SklearnClassifier

from wordcloud import WordCloud,STOPWORDS
import matplotlib.pyplot as plt

Download the data set

data = pd.read_csv('Sentiment.csv')
# Keeping only the neccessary columns
data = data[['text','sentiment']]

First of all, splitting the data set into a training and a testing set. The test set is the 10% of the original data set.

For this particular analysis I dropped the neutral tweets, as my goal was to only differentiate positive and negative tweets.

# Splitting the dataset into train and test set
train, test = train_test_split(data,test_size = 0.1)
# Removing neutral sentiments
train = train[train.sentiment != "Neutral"]

As a next step I separated the Positive and Negative tweets of the training set in order to easily visualize their contained words.

After that I cleaned the text from hashtags, mentions and links. Now they were ready for a WordCloud visualization which shows only the most emphatic words of the Positive and Negative tweets.

train_pos = train[ train['sentiment'] == 'Positive']
train_pos = train_pos['text']
train_neg = train[ train['sentiment'] == 'Negative']
train_neg = train_neg['text']

def wordcloud_draw(data, color = 'black'):
    words = ' '.join(data)
    cleaned_word = " ".join([word for word in words.split()
                            if 'http' not in word
                                and not word.startswith('@')
                                and not word.startswith('#')
                                and word != 'RT'
    wordcloud = WordCloud(stopwords=STOPWORDS,
    plt.figure(1,figsize=(13, 13))
print("Positive words")
print("Negative words")
twitter wordcloud

Interesting to notice the following words and expressions in the positive word set: truthstronglegitimatetogetherlovejob

In my interpretation, people tend to believe that their ideal candidate is truthful, legitimate, above good and bad.

At the same time, negative tweets contains words like: influencenewselevatormusicdisappointingsoftballmakeupcherry pickingtrying

In my understanding people missed the decisively acting and considered the scolded candidates too soft and cherry picking.

After the vizualization, I removed the hashtags, mentions, links and stopwords from the training set.

Stop Word: Stop Words are words which do not contain important significance to be used in Search Queries.

Usually these words are filtered out from search queries because they return vast amount of unnecessary information. ( the, for, this etc. )

tweets = []
stopwords_set = set(stopwords.words("english"))

for index, row in train.iterrows():
    words_filtered = [e.lower() for e in row.text.split() if len(e) >= 3]
    words_cleaned = [word for word in words_filtered
        if 'http' not in word
        and not word.startswith('@')
        and not word.startswith('#')
        and word != 'RT']
    words_without_stopwords = [word for word in words_cleaned if not word in stopwords_set]
    tweets.append((words_without_stopwords, row.sentiment))

test_pos = test[ test['sentiment'] == 'Positive']
test_pos = test_pos['text']
test_neg = test[ test['sentiment'] == 'Negative']
test_neg = test_neg['text']

As a next step I extracted the so called features with nltk lib, first by measuring a frequent distribution and by selecting the resulting keys.

# Extracting word features
def get_words_in_tweets(tweets):
    all = []
    for (words, sentiment) in tweets:
    return all

def get_word_features(wordlist):
    wordlist = nltk.FreqDist(wordlist)
    features = wordlist.keys()
    return features
w_features = get_word_features(get_words_in_tweets(tweets))

def extract_features(document):
    document_words = set(document)
    features = {}
    for word in w_features:
        features['contains(%s)' % word] = (word in document_words)
    return features

Hereby I plotted the most frequently distributed words. The most words are centered around debate nights.


Using the nltk NaiveBayes Classifier I classified the extracted tweet word features.

# Training the Naive Bayes classifier
training_set = nltk.classify.apply_features(extract_features,tweets)
classifier = nltk.NaiveBayesClassifier.train(training_set)

Finally, with not-so-intelligent metrics, I tried to measure how the classifier algorithm scored.

neg_cnt = 0
pos_cnt = 0
for obj in test_neg: 
    res =  classifier.classify(extract_features(obj.split()))
    if(res == 'Negative'): 
        neg_cnt = neg_cnt + 1
for obj in test_pos: 
    res =  classifier.classify(extract_features(obj.split()))
    if(res == 'Positive'): 
        pos_cnt = pos_cnt + 1
print('[Negative]: %s/%s '  % (len(test_neg),neg_cnt))        
print('[Positive]: %s/%s '  % (len(test_pos),pos_cnt))    
[Negative]: 842/795 
[Positive]: 220/74 

Leave a Reply