Affective computing and fine-grained feature mining technology of user feedback for generative AI mobile applications
Abstract
With the rapid popularization of Generative Artificial Intelligence (GenAI) on mobile terminals, mining multidimensional product features and sentiment tendencies from genuine user reviews is of profound guiding significance for the iterative optimization of application experiences. Traditional sentiment classification models often encounter bottlenecks in feature extraction when processing unstructured long texts that contain complex contexts and intersecting emotions. This paper proposes a fine-grained affective computing method that integrates Large Language Model (LLM) architectures with deep semantic networks. Focusing on the interactive characteristics of GenAI applications, this study constructs a multi-dimensional feature dataset based on real user feedback, including User Interface/User Experience (UI/UX), core generation functionality, and system stability. By leveraging the Multi-Head Self-Attention mechanism, aspect-level feature words are accurately extracted, and a dynamic masking strategy is combined to compute microscopic sentiment polarities. Experimental results demonstrate that the proposed binary classification model significantly outperforms traditional machine learning models, achieving an accuracy of 88.71%. Furthermore, it effectively reduces the false positive rate in complex interactive emotion recognition, providing a reliable quantitative paradigm for the experience measurement and functional iteration of intelligent software products.