A Tree-Structured Deep Learning Model for Improving Classification with Self-Adaption and Self-Learning.

Tree-structured deep learning classifier models are widely used in dimensional sentiment analysis for efficient feature representation and learning. From this perspective, an Adversarial Tree-structured Convolutional Neural Network with Long Short-Term Memory (A-T-CNN-LSTM) model was developed that adopts the Semantic-enabled Frequency-aware Generative Adversarial Network (SFGAN) to create more adversarial samples for predicting the Valence-Arousal (VA) of the texts or image classes. In contrast, an abrupt change in input data was not handled that impacts the model accuracy. Hence, this article proposes an Adversarial Attention T-CNN-LSTM (AA-T-CNN-LSTM) model to handle abrupt changes and uncertainties in the input data for dimensional sentiment analysis. This model aims to enhance self-adaptation and self-learning efficiency by integrating an attention strategy with the A-T-CNN-LSTM network. This model is constructed based on the SFGAN, CNN, LSTM and attention strategy layers. The CNN captures the spatial dependencies, whereas the LSTM captures the temporal dependencies of the given input data. The attention strategy layer is included after LSTM to adaptively control the proportion of spatial and temporal dependencies by emphasizing a few weights for final output vectors. Moreover, the prediction of VA ratings of the texts or image classes is achieved based on the final output vectors. Finally, the testing outcomes reveal that the AA-T-CNN-LSTM model on the Stanford Sentiment Treebank (SST) and CIFAR-10 datasets reaches an accuracy of 91.84% and 93.14%, respectively, contrasted with the state-of-the-art models. [ABSTRACT FROM AUTHOR]