SSK-DNN: Semantic and Sentiment Knowledge for Incremental Text Sentiment Classification

Abstract

Sentiment analysis is a crucial task that extracts useful insights from online unstructured user-generated content for incremental sentiment classification. Traditional feature construction methods for text sentiment analysis have their own benefits in interpretability, computation, and lowering the high dimensionality of textual input. Although this enhances the classification accuracy of traditional machine learning classifiers, still they struggle with sparsity due to a lack of appropriate data representation strategies. In contrast, DNN architectures for text sentiment analysis have overcome the sparsity issue and obtained better results than traditional feature-based machine learning methods. Yet, they suffer from high-dimensional feature space due to noisy and irrelevant features from the unstructured user-generated review text. Moreover, state-of-the-art sentiment analysis methods can not fully exploit semantic and sentiment knowledge in order to extract meaningful relevant contextual sentiment features. Our goal is to mitigate such problems by employing standard DNN-based methods (BERT, BiLSTM with Attention Mechanism) and traditional feature construction methods with PCA to extract relevant sentiment features and reduce the dimensionality of feature space. We propose an effective way of joining the traditional feature construction methods with the DNN-based methods to improve the performance of sentiment analysis. Additionally, we leverage incremental learning to adapt to new data effectively. Our main contributions are six folds (a) mining semantic and sentiment features in the review text for sentiment analysis, (b) creating dynamic word vector representations to enhance sentiment word embeddings, (c) filtering noisy and irrelevant features to reduce the high dimensional feature space, (d) analyzing the data's inherent relationships and optimizing the weighting of key features for accurate sentiment classification, (e) integrating essential new features to support incremental learning, (f) practicality, Our model undergoes evaluation on different real-world benchmark datasets, showcasing its capacity to enhance the performance of sentiment classification in comparison to several existing methods. © 2023 IEEE.