Binary classification

Binary classification on torchtext.datasets.SST

• https://pytorch.org/text/_modules/torchtext/datasets/sst.html
• https://nlp.stanford.edu/sentiment/

Test set accuracies on default splits, evaluated on default train-test split without cross validation:

Naive Bayes:

~83% acc - Wang and Manning (https://www.aclweb.org/anthology/P12-2018.pdf) have 83.45% acc on RT-2k with MNB-uni

Logistic regression with bag of words features:

~78% acc

Logisitc regression with features from pretrained glove.6B.100d:

~77% acc

Logisitc regression with features from CNNs as in Kim 2014 (without phrase filters):

• >80->81.5% acc with Adadelta, dropout_rate = 0.25, weight_decay = 0
• similarly with weigh_decay = 0.01
  • Kim 2014 has >81% acc (https://github.com/yoonkim/CNN_sentence/blob/master/README.md) with Adadelta, dropout = 0.5 and l2 regularization
  • more on dropout and l2: https://arxiv.org/pdf/1510.03820.pdf