The timely detection of psycho-emotional risks has become increasingly important due to the rapid growth of social media platforms. This study examines user-generated text as a potential source of early indicators of psychological vulnerability. The proposed NLP-based framework incorporates behavioral features to improve the interpretation of users’ psycho-emotional states. In addition to text classification, the study considers structured behavioral indicators to support psycho-emotional risk analysis. Particular attention is given to interpretability. SHAP-based techniques are applied to reveal the contribution of individual features and to provide a clearer explanation of model predictions. The evaluation was conducted on publicly available datasets containing textual data and aggregated behavioral/physiological indicators. No raw physiological streams, wearable sensor data, or biometric recordings were used. The two datasets were employed in complementary experimental settings and were not aligned at the individual-sample level; accordingly, the broader analytical perspective explored in this study should not be interpreted as a single end-to-end or fully aligned multimodal learning framework. The proposed BERT-based model with SHAP interpretability achieved an accuracy of 96.3%, an F1-score of 0.96, and a ROC–AUC score of 0.98, showing consistent improvement over baseline models, including Random Forests and Support Vector Machines.

Learning meaningful representations of scientific documents is essential for information retrieval, knowledge discovery, and recommendation systems. Traditional methods such as TF-IDF rely on lexical matching and fail to capture deeper semantic relationships, while transformer-based approaches typically depend on limited supervision signals. In this work, we propose a Triple-Source automatic supervision framework for learning document embeddings from scientific corpora. The model integrates three types of supervision–title–abstract pairs, same-category document pairs, and document-level semantic relationships—within a unified contrastive learning framework based on a multilingual XLM-RoBERTa encoder. Unlike prior approaches that rely on citation graphs or manual annotations, our method enables citation-free and annotation-free representation learning using only lightweight metadata. Experiments on a publicly available arXiv dataset consisting of 98,649 documents demonstrate improved semantic retrieval performance, achieving Recall@1 = 0.6181 for same-category retrieval and outperforming both TF-IDF and single-source transformer baselines. The learned embeddings also exhibit improved clustering of scientific domains, indicating more structured semantic representations.