ACDPSNet: Adaptive Cross Domain Polarity Aspect Level Learning Scalable Computing Model for Sentiment Classification and Quantification

Abstract

Automatic sentiment classification, identifying opinions as positive, negative, or neutral, is essential across diverse applications. However, applying a sentiment classifier trained on labeled data from one domain to a different domain often leads to degraded performance, as domain-specific language terms common in the source domain may not appear in the target domain. This research proposes an Adaptive Cross-Domain Polarity-Specific Network (ACDPSNet) for sentiment classification and quantification across domains. The model leverages labeled data from the source domain alongside labeled and unlabeled data from the target domain to build a robust, adaptable domain adaptation framework. Sensitivity to sentiment is enhanced by embedding polarity-specific sentiment annotations into semantic vectors, enabling accurate computation of distributional similarities between terms. The framework integrates a classifier that is both domain-specific and domain-invariant to ensure accurate analysis and classification. ACDPSNet achieves notable performance improvements, with an accuracy of 98.76%, recall of 97.85%, throughput of 96.94%, and a positive learning expression rate of 97.76%, demonstrating significant advancements over existing approaches. These metrics underscore ACDPSNet’s effectiveness in adapting to new domains, achieving high sentiment quantification accuracy, and enhancing cross-domain polarity detection.