covid19-vaccine-tweets-sentiment-analysis

Project in Natural Language Processing

About

In this project, we develop deep learning models for sentiment analysis classification of covid-19 vaccination tweets in to 3 categories:

• pro-vax
• neutral
• anti-vax

We develop and compare four different classifiers on the same datasets using real twitter data:

• Vaccine sentiment classifier using softmax regression
• Vaccine sentiment classifier using feed forward neural networks
• Vaccine sentiment classifier using recurrent neural networks
• Vaccine sentiment classifier by fine-tuning the pretrained BERT-base model

Softmax Regression Classifier

We experimented with :

• data cleaning / noise removal to prevent model from learning useless information / features than hinder the sentiment analysis from generalizing to unseen data
• different vectorizers (TF-IDF vectorizer, count vectorizer, Bag Of Words Vectorizing)
• max-df / min-df parameters to remove features with very low/high frequency
• regularization parameter to reduce overfitting
• Performance evaluated based on precision / recall / f1 score

Model / experiments in : /SoftmaxRegression_classifier/softmax_regression_model.ipynb

Feed Forward Neural Network Classifier

We experimented with :

• data cleaning / noise removal to prevent model from learning useless information / features than hinder the sentiment analysis from generalizing to unseen data
• pre trained GloVe word embeddings
• different number of layers, different learning rates
• linear activation functions, non-linear activation functions (ReLU)
• dropout layers to reduce overfitting
• different optimizers (Stochastic gradient descent, Adam, RMSprop)
• Cross entropy loss function since we have multiclass unbalanced classification
• Performance evaluated based on precision / recall / f1 score / ROC curves

Model / experiments in : /FFNN_classifier/feed_forward_nn_model.ipynb

Recurrent Neural Network Classifier

We experimented with :

• Bidirectional stacked RNN's with LSTM/GRU cells
• Gradient Clipping to tackle the exploding gradients problem of RNN's
• Early stopping regularization to avoid overfitting (LSTM/GRU models learn much faster than previous models)
• Skip connections to train deeper RNN's and tackle the vanishing gradients problem
• Added attention to our best model to further improve performance
• Performance evaluated based on precision / recall / f1 score / ROC curves

Model / experiments in : /RNN_classifier/recurrent_nn_model.ipynb

Classifier by fine-tuning pre-trained BERT-base model

• BertForSequenceClassification model was used
• Experimented with the hyperparameters suggested in BERT's paper

Model / experiments in : /BERT_classifier/bert_model.ipynb