Deep Sentient network with multifarious features and inter-mutual attention mechanism for target-specific sentiment classification

Abstract

Target-based aspect level sentiment analysis (TBASA) seeks to discover the polarity of the text towards certain aspect terms in each text. Most of the recent studies utilize deep learning (DL) frameworks like Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to predict the influences of multiple contextual aspects on sentiment polarity. Both CNN and RNN are successfully used earlier to create complicated semantic representations. However, existing approaches fail to capture the sequence information due to the high dimensionality. In this paper, a typical approach called a Deep Sentient network with a novel inter-mutual attention mechanism is proposed to tackle this issue. The proposed model adds the sequence information identified with RNN into CNN to consistently anticipate the polarity. It also learns the contextual and target terms sequentially to understand the mutual impact between the features. Furthermore, both Part-of-Speech (POS) and position information are also included in the input layer as background knowledge. Finally, a series of experiments are performed on various benchmark datasets to verify the efficacy of our proposed approach.