A study on word vector dimensions for sentence classifications using convolutional neural networks

Recently, convolutional neural networks (CNNs) have achieved remarkable results on sentence classification problems. In these approaches, each word in the sentences is transformed to real number vectors (called word vectors) and the sentences as input data to the CNN are represented by the sequences of the word vectors. A dataset for training and testing for the CNN includes the large number of words, therefore the word vectors are embedded so high-dimentional space. As a result of this, the input data space of the CNN becomes very high. When the input data have high dimension, much training data are required for enough training of the CNN. It is not always possible, however, to get enough number of data for training. If the enough data cannot prepare for learning, it is desirable to decrease the dimension of input data. This paper shows the results that the smaller dimensional word vectors are applied to sentence classifications by CNNs. The results have shown that some dimensionality reduction does not effect too much to the accuracy of the sentence classifications by CNNs. © 2019 The Institute of Electrical Engineers of Japan.