Sentiment-Analysis-Model-Experiments

Dataset Used:

Twitter US Airline Sentiment - Kaggle

Experiments:

Experiment-1. Using Machine Learning Algorithms and Vectorization Techniques For Sentiment Analysis

Text-Vectorization Techniques used:

• CountVectorizer
• TfidfVectorizer
• OneHotEncoding
  
  
  

ML Algorithms used:

• Logistic Regression
• Naive Bayes
  
  
  

Result:

BEST MODEL: TFIDFvectorizer_LogisticRegression

Experiment-2. Multi-Layer Perceptron (MLP) Models with different Model Architectures and Optimizers For Sentiment Analysis

Model Architectures:

• Model-1
  
  Layer (type) - Output Shape
  
  layer_1 (Dense) - (None, 64)
  
  layer_2 (Dense) - (None, 64)
  
  layer_3 (Dense) - (None, 3)

• Model-2
  
  Layer (type) - Output Shape
  
  layer_1 (Dense) - (None, 32)
  
  layer_2 (Dense) - (None, 3)

• Model-3
  
  Layer (type) - Output Shape
  
  layer_1 (Dense) - (None, 10)
  
  layer_2 (Dense) - (None, 3)

Optimizers:

• adam
• rmsprop
• sgd
  
  
  

Result:

Experiment-3. Recurrent Neural Network (RNN) Models For Sentiment Analysis

Model Architectures:

• Simple RNN Model
  
  Layer (type) - Output Shape
  
  embedding_12 (Embedding) (None, 22, 100)
  
  layer_1 (SimpleRNN) (None, 128)
  
  layer_2 (Dense) (None, 10)
  
  output_layer (Dense) (None, 3)
  

• LSTM Model
  
  Layer (type) - Output Shape
  
  embedding_12 (Embedding) (None, 22, 100)
  
  layer_1 (LSTM) (None, 128)
  
  output_layer (Dense) (None, 3)
  

• GRU Model
  
  Layer (type) - Output Shape
  
  embedding_12 (Embedding) (None, 22, 100)
  
  layer_1 (GRU) (None, 128)
  
  output_layer (Dense) (None, 3)
  

• Bidirectional LSTM Model
  
  Layer (type) - Output Shape
  
  embedding_12 (Embedding) (None, 22, 100)
  
  bidirectional_6 (Bidirectional) (None, 128)
  
  output_layer (Dense) (None, 3)
  
  
  
  

Result:

Experiment-3. Pretrained and Finetuned BERT Model For Sentiment Analysis

Result:

Overall Best Model: BERT