The joint effect of semantic and syntactic word embeddings on sentiment analysis

Employing pre-trained word embeddings as preliminary features in convolutional neural networks (CNN) for natural language processing (NLP) tasks has been proved to be of benefit. We exploit this idea by taking advantage of different types of word embeddings at the same time. To be specific, we extend CNN models to coordinate two lookup tables, which exploit semantic word embeddings and syntactic word embeddings at the same time. We test our models on several review datasets and all results indicate the positive effect on sentiment analysis. To understand the reason behind, we explore the difference of the two word embeddings and how they influence the CNN models.