AI-Enabled Digital Trust, Ethics, and Safety-Risk Signal Analysis in Contact-Based Sport Communities: ESG-Oriented Text Mining and Sentiment Classification of Judo and Brazilian Jiu-Jitsu Platform Discourse

Existing platform-monitoring methods for sport communities commonly rely on isolated descriptive text-mining outputs or general sentiment scores; they rarely integrate interpretable ESG issue coding with class-sensitive risk detection and provide limited support for auditable, privacy-conscious analysis of safety, ethics, and institutional trust. These limitations motivate a multi-stage framework that converts heterogeneous platform discourse into complementary structural and evaluative signals. Conceptually, digital trust is treated as the focal governance outcome; ethics and safety are substantive domains of concern; ESG provides the bounded classification and response ontology; and early warning denotes a prototype, human-reviewed weak-signal triage concept rather than incident prediction or a deployed security-monitoring system. Using 377,700 cleaned Korean-language comments on judo and Brazilian Jiu-Jitsu (BJJ) collected from Naver News and YouTube between 2010 and 2025, the framework combines n-gram analysis, LDA topic modeling, CONCOR network analysis, bounded ESG discourse classification, and three-class sentiment prediction. The individual analytical algorithms are established; the methodological contribution lies in their governance-oriented orchestration through a bounded ESG/non-ESG coding gate, a study-specific index layer, and a human-reviewed pathway from aggregate discourse signals to proportionate review. The analytical workflow identifies issue salience, relational topic structures, ESG dimensions, sentiment risk, legitimacy balance, and platform-specific risk concentration while excluding personally identifiable information. Empirically, social and governance concerns dominate the corpus, and governance-related negative sentiment consistently exceeds social-risk sentiment, highlighting rule transparency, coach ethics, misinformation, platform reputation, and institutional response as central trust-risk domains. Cell-weighted sensitivity checks preserved the governance-over-social and YouTube-over-Naver risk ordering, although the pooled salience estimate remained sensitive to the rapid expansion of BJJ discourse on YouTube. The fine-tuned KLUE-BERT model achieved a Macro-F1 of 0.838 and a negative-class F1 of 0.862, outperforming the strongest baseline, Text-CNN (Macro-F1 = 0.791), by 0.047 absolute Macro-F1 points (approximately 6.0% relative improvement). These findings support the feasibility of a batch-oriented, human-reviewed prototype for prioritizing aggregate discourse patterns. They do not establish the effectiveness of a real-time security-monitoring, incident-detection, or operational early-warning system.