The rapid expansion of textual data across multilingual digital environments has intensified the need for robust, scalable, and linguistically adaptive text classification frameworks. Traditional monolingual models often struggle to generalize across heterogeneous linguistic contexts, particularly when handling semantic ambiguity, domain variation, and cross-lingual knowledge transfer. This study proposes a cross-lingual pretrained framework that integrates diverse feature aggregation mechanisms to enhance the classification performance of English documents. The framework leverages advances in transformer-based architectures, graph neural networks, and multi-view representation learning to capture syntactic, semantic, and contextual dependencies within textual data.

The research builds upon pretrained language models such as BERT and its variants, combining them with feature fusion strategies derived from convolutional neural networks, recurrent neural networks, and graph-based representations. By integrating multiple feature spaces—including lexical embeddings, contextual representations, and structural graph features—the proposed framework addresses limitations associated with single-representation learning. The model incorporates cross-lingual knowledge transfer through pretrained multilingual embeddings and heterogeneous graph attention mechanisms, enabling improved generalization across diverse datasets.

Methodologically, the study adopts a hybrid architecture that combines transformer encoders with feature aggregation layers and graph-based relational modeling. Experimental evaluation is conceptually structured using benchmark classification scenarios, focusing on accuracy, robustness, and scalability. The findings indicate that multi-feature aggregation significantly enhances classification performance, particularly in complex and noisy datasets where contextual dependencies are critical. Furthermore, the integration of cross-lingual pretrained models improves semantic consistency and reduces classification errors associated with linguistic variability.

The study contributes to the growing body of research on intelligent text classification by proposing a unified framework that bridges gaps between pretrained language models and multi-view feature integration. The implications extend to applications in information retrieval, sentiment analysis, content moderation, and enterprise document management. However, challenges related to computational complexity, data dependency, and interpretability remain critical considerations for future research. The paper concludes by outlining directions for optimizing cross-lingual architectures and enhancing feature fusion strategies for next-generation text classification systems.

Enterprise ERP systems still rely on large monoliths that bind finance, HR, reporting, and operations to a single codebase and database. That structure slows releases, widens regression risk, and spreads failures across unrelated functions. This review synthesizes recent research on migrating to a modular microservices architecture, with attention to migration sequence, service decomposition, data ownership, and post-migration control. The source base comprises 10 peer-reviewed publications from 2021 to 2025, including systematic reviews, literature reviews, case studies, and a peer-reviewed conference paper. Comparative source analysis, conceptual synthesis, typologization, and analytical generalization were used. The reviewed literature converges on four conditions of stable migration: phased transition, business-aligned bounded contexts, explicit persistence boundaries, and runtime control through observability and testing. On that basis, the article formulates a decision model for enterprise teams that need to modernize large business systems while preserving operational continuity, data reliability, and manageable coordination costs.

Insider threats remain one of the most critical and elusive challenges in enterprise cybersecurity due to their ability to exploit legitimate access while evading traditional detection mechanisms. In this study, a hybrid framework integrating Graph Neural Networks and Large Language Models is proposed to enhance insider threat detection through the fusion of behavioral graph modeling and semantic profiling. Using the CERT Insider Threat Dataset and the UEBA Dataset on Kaggle, the model captures both relational dependencies among users, devices, and resources, and contextual insights from unstructured textual data such as logs and communications. The experimental results demonstrate that the proposed hybrid model significantly outperforms traditional machine learning, sequence-based, and single-modality deep learning approaches, achieving an accuracy of 0.96, an F1-score of 0.92, and a ROC-AUC of 0.95. These improvements are primarily driven by the model’s ability to jointly learn structural anomalies and semantic deviations, enabling more accurate detection of multi-stage and stealthy insider attacks. Furthermore, the integration of explainable language-based outputs enhances interpretability and operational usability in enterprise security environments. The findings highlight the effectiveness of multi-modal learning in advancing insider threat detection and provide a scalable, practical solution for deployment in large-scale enterprise systems.