Factors Influencing User Satisfaction with Artificial Intelligence Applications: An Integrated Approach of Semantic Network and BP Neural Network

Abstract

Generative AI has moved quickly from laboratory demos to everyday mobile apps in China, and user satisfaction has become the clearest test of whether these products actually deliver value. Yet most satisfaction studies still lean on questionnaires — useful, but too structured to capture what people really complain about or praise after prolonged use. This paper looks instead at what users themselves wrote. We collected 63,032 online reviews of ByteDance's DouBao APP (54,248 positive and 8,784 negative, November 2024 to October 2025, compiled via DianDian Data) and asked a simple question: what drives satisfaction, and how much does each driver matter? Rather than relying on a single method, we combined four steps. First, an IWE-AS keyword extraction model — which adds position encoding and a TDF adjustment to the classic TF-IDF — was used to surface terms that genuinely reflect concern, not just frequency. Semantic network mapping (ROSTCM6) and LDA topic modeling (with pyLDAvis for tuning) then revealed how those keywords cluster into themes. Finally, a BP neural network translated those themes into weights in a purely data-driven way, avoiding the subjectivity of expert scoring. The resulting evaluation system covers 6 primary dimensions and 24 secondary items. Interaction (0.347) and accuracy (0.238) dominate — together close to 60% of total weight — followed by intelligence (0.122), practicality (0.116), learning capability (0.101) and overall functionality (0.076). Within those, 'chat / picture generation' and 'failure to understand instructions' in the interaction dimension, and 'speechless / incorrect answers' in the accuracy dimension, carry the highest secondary weights. The implication is practical: for DouBao and similar domestic AI apps, improving how well the system understands colloquial instructions and how reliably it answers — especially for study-related queries — will move satisfaction far more than adding new peripheral features.