Sentiment analysis using convolutional neural network via word embeddings

Convolutional neural networks are known for their excellent performance in computer vision, achieving results in the state of the art. Moreover, recent research has shown that these networks can also provide promising results for natural language processing. In this case, the basic idea is to concatenate the vector representations of words into a single block and use it as an image. However, despite the good results, the problem of using convolution networks is the large numbers of design decisions that need to be made á priori. These models require the definition of many hyper-parameters, including the type of word embeddings, which consists of the data vectorized representation, the activation function that prints the non-linearity characteristics to the model, the size of the filter that applies data convolution, the number of feature maps, which are responsible for identifying the attributes and the pooling method used for data reduction. In addition, one must also predefine the regularization constant and the dropout rate, which are responsible for avoiding any network over-fitting. In existing research works, convolutional neural network architectures capable of overcoming the performance of traditional machine learning models are presented. Even though these can compete with more complex models, the problem of how the different setting of the hyper-parameters may affect the performance of this type of network has not yet been explored. In this paper, we propose an efficient sentiment analysis classifier using convolutional neural networks by analyzing the impact of the hyper-parameters on the model performance. The main interest in analyzing sentiment comes from the advent of social media and the technological advances that flood the Internet with opinions. Nonetheless, mining the Internet for opinion and sentiment analysis is not an easy task and thus needs outstanding models with the best hyper-parameters setting to be able to get pertinent answers. The results achieved are obtained with the use of GPU and show that the different configurations exceed the reference models in the most of the cases with gains of up to 18% and have similar performance to the models of the state of the art with gains of up to 2% in some cases.