This repo is aiming for reproducing the sentence classifcation experiments in Mou et al. (EMNLP 2016). https://aclweb.org/anthology/D16-1046
https://drive.google.com/file/d/0B8yp1gOBCztyN0JaMDVoeXhHWm8/
Train+dev = 600,000 (imdb.neg+imdb.pos) 550,000 for train and 50,000 for dev
Test = 2,000 (rt_critics.test)
https://www.cs.cornell.edu/people/pabo/movie-review-data/
Train+dev+test = rt-polarity.neg + rt-polarity.pos
all = 5331*2 = 10662 = 8500(train) + 1100(dev) + 1062(test)
http://cogcomp.cs.illinois.edu/Data/QA/QC/
Train(train_5500.label, 5452 = 4,800(train) + 652(dev)) + test(TREC_10.label, 500)
"LSTM_sentence_classifier.py"
Remark:
This model is the simplest version of LSTM-Softmax Classifier.
It doesn't use mini-batch or pretrained word embedding.
Note that there is not fixed lenght of the sentences.
Its performance with Adam(lr = 1e-3) is 76.1 in terms of accuracy on MR dataset.
It is slower with Adam than with SGD, but the performance is much better.
However, it's unreasonable to use such code to train a very large dataset like IMDB.