Online social networks play a central role in information diffusion and opinion formation, making the identification of meaningful communities essential for understanding collective behavior. However, most community detection methods rely primarily on network topology and remain insensitive to the semantic and affective information embedded in user-generated content. To address this limitation, this study proposes a content-enriched community detection framework that integrates structural interactions (retweets and mentions) with topic- and sentiment-based attributes. The proposed approach includes two core methodological components: a learnable edge-weighting mechanism that infers interaction strengths from structural, topical, and sentiment features and a content-aware modularity function (Qattr) that explicitly rewards within-community semantic and affective alignment during Leiden-based optimization. Both components are model-agnostic and can be applied with either lightweight text representations or transformer-based models, allowing flexibility in trade-offs between accuracy and efficiency. The framework is evaluated through extensive experiments on Twitter data, including in-domain, leave-one-domain-out, and cross-platform settings. Performance is assessed using structural metrics (Modularity and Clique Percolation Method), semantic validation measures (Normalized Mutual Information and Adjusted Rand Index), and blinded human evaluation of topic coherence and sentiment grouping with inter-annotator agreement. Statistical significance testing across multiple runs confirms the robustness of the observed improvements. Results demonstrate that jointly incorporating structural, semantic, and sentiment information yields more coherent, semantically aligned communities than structure-only baselines while maintaining practical scalability. Overall, this work provides a principled and empirically validated foundation for semantic-aware community detection in social media networks.
