Aspect-based Capsule Neural Networks for Sentiment Analysis of Sensitive Information

To address the need for accurate sentiment masking of sensitive words in sensitive information, traditional sentiment analysis methods can only assess the overall polarity of a sentence, while existing Aspect-Based Sentiment Analysis (ABSA) mostly focuses on the field of commodity reviews, and insufficiently handles phenomena such as complex contexts and irony. To this end, this paper proposes the CapABSA model, which takes sensitive words as aspects, combines capsule neural networks to capture multidimensional semantic features, and migrates document-based sentiment analysis (DBSA) to ABSA through transfer learning to improve classification performance. Experiments show that the model performs effectively in sentence datasets containing sensitive words, providing a more fine-grained solution for sensitive information filtering.