Sentiment Analysis

Text preprocessing and vocabulary building using Texthero

• default texthero preprocessing used for cleaning a dataset. Finally, we got 359985 words for negative sentiments, and 442559 words for positive sentiments.
• vocabulary lengths for different term frequency thresholds are(words with higher than term frequency thresholds took into consideration while making vocabulary):
  • ft:0, vocabulary lenght:802544
  • ft:5, vocabulary lenght:51758 (this was chosen for TF-IDF vocabulary)
  • ft:25, vocabulary lenght:15658
  • ft:50, vocabulary lenght:9896
  • ft:100, vocabulary lenght:6138
• WordClouds using Texthero

Positive

Negative

TFIDF Representation

TF-IDF Representation: TF-IDF works by determining the relative frequency of words in a specific document compared to the inverse proportion of that word over the entire document corpus. It's used with a vocabulary size of 51758 to take only words with term frequency of higher than 25 in positive and negative samples.

transformer=TfidfVectorizer(vocabulary=vocabs)

Experimenting with Perceptron and Multi-layer Perceptron classifiers:

Perceptron Classifier: A basic Neural Network used for sentiment analysis.

MLP Classifier: A Multi-layer Perceptron classifier that uses log-loss function using LBFGS or stochastic gradient descent. sklearn. MLP with a 2-hidden layer with 100 neurons on each layer respectively introduced, with a batch size of 512. Due to a large number of samples in the dataset, and time complexity of training, I set max_iter to 20 epoch.

MLPClassifier(hidden_layer_sizes=(100,100),batch_size=512, max_iter=20, verbose=True)

Experimenting with an Naive Bayes, Logistic Regression, Linear SVM, and RandomForest Classifiers:

• NB: Naive Bayes, BernoulliNB implements the naive Bayes training and classification algorithms for data that is distributed according to multivariate Bernoulli distributions
• LR:Logistic Regression (aka logit, MaxEnt) classifier.
• LSVM:Linear Support Vector Classification.
• RF : A random forest is a meta estimator that fits a number of decision tree classifiers on various sub-samples of the dataset and uses averaging to improve the predictive accuracy and control over-fitting.

Experimenting with an Fully Connected Neural Network (FCNN) Classifier:

• input layer consist of 512 neurons, with activation function of ReLU and droup out with probability of 40%
• 3-hidden layer deep neural network, with activation functions of ReLU, and layers with 256 neurons on each.
• output layer consist of 2 neurons, with activation function of sigmoid
• train have done in 3 iterations ( 30 minutes each iteration)
• I achieved much more reliable result with FCNN
• what else we can do?

Obtained Results for Models

Model Description	Precission	Recall	F1-Score	Accuracy
TFIDF+Perceptron	0.71	0.71	0.71	0.71
TFIDF+MLP	0.75	0.75	0.75	0.75
TFIDF+NB	0.77	0.77	0.77	0.77
TFIDF+LR	0.78	0.78	0.78	0.78
TFIDF+LSVM	0.78	0.78	0.78	0.78
TFIDF+RF	0.75	0.75	0.75	0.75
TFIDF+FCNN	0.79	0.79	0.79	0.79

Requirements

• keras
• tensorflow
• scikit-learn
• texthero
• nltk
• numpy
• tweet-preprocessor