Sentiment Analysis Using Bi-CARU with Recurrent CNN Models

Abstract

For many natural language processing tasks, sentiment analysis has become increasingly important for extracting meaningful information from social media data. With the out-performance of neural network technology, the task of sentiment analysis can be addressed by advanced deep learning models. In this work, a combination model of Bidirectional-CARU (Bi-CARU) and Recurrent CNN is introduced to the sentiment analysis tasks. The proposed Bi-CAUR consists of three layers designed to obtain the main features of the input sequence, which can alleviate the long-term dependency problem and perform kernel information filtering from concrete to abstract, effectively improving the performance of the intermediate network on this problem. Next, the recursive structure of the CNN is connected to Bi-CARU to determine the sentiment analysis. The proposed Recurrent CNN implementation accepts features produced by its own previous convolution and pooling, which incorporates the performance of a CNN and requires only fewer parameters. Experimental results show that we are slightly more accurate, achieve faster convergence, and require fewer training parameters.