twitter-sentiment-analysis

Comparing various Neural Models in Sentiment Analysis tasks on Twitter datasets

Project for the course Natural Language Processing 1 @ University of Amsterdam

Davide Belli, Gabriele Cesa, Linda Petrini

Paper

python main.py --cuda --epochs 20 

See experiment_final.sh for more examples

Command line arguments description for 'main.py':

• '--data'       default='./data/dataset/preprocessed/' 'location of the data corpus'
• '--model'       default='LSTM'
  1. RNN_TANH
  2. RNN_RELU
  3. LSTM
  4. GRU
  5. LSTM_BIDIR
  6. LSTM_REV
  7. RAN
  8. RAN_BIDIR
  9. CNN
• '--emsize'       default=300,       'size of word embeddings'
• '--nhid'       default=200,       'number of hidden units per layer'
• '--nlayers'       default=2,       'number of layers'
• '--nreduced'       default=30,       'number of units in the reduced layer'
• '--lr',       default=0.01,       'initial learning rate'
• '--lamb',       default=0.1,       'lambda for L2 regularization (weight decay)'
• '--lrdecay',       default=0.0,       'learning rate decay parameter for Adagrad'
• '--clip',       default=0.25,       'gradient clipping'
• '--epochs'       default=40,       'upper epoch limit'
• '--batch_size'       default=20,       'batch size'
• '--bptt'       default=35,       'sequence length'
• '--dropout',       default=0.5,       'dropout applied to layers (0 = no dropout)'
• '--seed'       default=1111,       'random seed'
• '--log-interval'       default=50,       'report interval'
• '--save'       default='model.pt'       'path to save the final model'
• '--recallsave'       default='model_recall.pt'       'path to save the final model'
• '--pause_value'       default=0,       'not optimise embeddings for the first 5 epochs'
• '--initial'       default=None,       'path to embedding file. If not set they are initialized randomly'

Booleans

• '--tied' = 'tie the word embedding and softmax weights'
• '--plot' = 'plot confusion matrix'
• '--last' = 'backpropagate at the end of a tweet'
• '--pre' = 'use preprocessed data'
• '--pause' = 'not optimise embeddings for the first 5 epochs'
• '--cuda' = 'use CUDA'
• '--shuffle' = 'shuffle train data every epoch'