Sentiment Analysis with Bidirectional LSTM ,RNN and GRU

Binary text classification on IMDB movie reviews — positive or negative.

Results

Model	AVG Test Accuracy
BiLSTM	~87%
RNN	~68%
GRU	~89%

Model Architecture

Input Text
↓
Tokenization 
↓
Embedding Layer 
↓
Model
↓
Concatenate
↓
Dropout (0.4)
↓
Linear Layer
↓
Sigmoid → probability (>0.5 = Positive)

Concepts Covered

• Word Embeddings
• Padding & Packing sequences
• Vanishing Gradients
• LSTM Gates (forget, input, output)
• Bidirectional LSTM
• Gradient Clipping
• GRU
• RNN

Setup

pip install torch datasets
python sentiment_LSTM.py

pip install torch datasets
python sentiment_Model.py

Sample Output

• Epoch 1/10
• Train loss: 0.66|Train Acc: 0.59
• Test loss: 0.77| Test acc: 0.51

LSTM

• one scene is good but the movie is worst NEG (confidence: 0.01)
• this was an absolutely brilliant masterpiece POS (confidence: 0.97)
• terrible boring waste of my time NEG (confidence: 0.01)

RNN

• one scene is good but the movie is worst Positive (confidence: 0.92)
• this was an absolutely brilliant masterpiece Negative (confidence: 0.27)
• terrible boring waste of my time Negative (confidence: 0.01)

GRU

• one scene is good but the movie is worst Negative (confidence: 0.00)
• this was an absolutely brilliant masterpiece Positive (confidence: 1.00)
• terrible boring waste of my time Negative (confidence: 0.00)

Author

Vikas Reddy

• GitHub: @vikasreddy11
• LinkedIn: Vikas Reddy