Batch size for training convolutional neural networks for sentence classification

Abstract

Sentence classiication of shortened text such as single sentences of movie review is a hard subject because of the limited inite information that they normally contain.We present a Convolutional Neural Network (CNN) architecture and better hyper-parameter values for learning sentence classiication with no preprocessing on small sized data.The CNN used in this work have multiple stages.First the input layer consist of sentence concatenated word embedding.Then followed by convolutional layer with different ilter sizes for learning sentence level features, followed by max-pooling layer which concatenate features to form inal feature vector.Lastly a softmax classiier is used.In our work we allow network to handle arbitrarily batch size with different dropout ratios, which is gave us an excellent way to regularize our CNN and block neurons from co-adapting and impose them to learn useful features.By using CNN with multi ilter sizes we can detect speciic features such as existence of negations like "not amazing".Our approach achieves state-of-the-art result for sentence sentiment prediction in both binary positive/negative classiication.