Sentiment analysis in legal documents presents significant challenges due to the intricate structure, domain-specific terminology, and strong contextual dependencies inherent in legal texts. In this study, a novel hybrid framework is proposed, integrating Graph Attention Networks (GATs) with domain-specific embeddings, i.e., Legal Bidirectional Encoder Representations from Transformers (LegalBERT) and an aspect-oriented sentiment classification approach to improve both predictive accuracy and interpretability. Unlike conventional deep learning models, the proposed method explicitly captures hierarchical relationships within legal texts through GATs while leveraging LegalBERT to enhance domain-specific semantic representation. Additionally, auxiliary features, including positional information and topic relevance, were incorporated to refine sentiment predictions. A comprehensive evaluation conducted on diverse legal datasets demonstrates that the proposed model achieves state-of-the-art performance, attaining an accuracy of 93.1% and surpassing existing benchmarks by a significant margin. Model interpretability was further enhanced through SHapley Additive exPlanations (SHAP) and Legal Context Attribution Score (LCAS) techniques, which provide transparency into decision-making processes. An ablation study confirms the critical contribution of each model component, while scalability experiments validate the model’s efficiency across datasets ranging from 10,000 to 200,000 sentences. Despite increased computational demands, strong robustness and scalability are exhibited, making this framework suitable for large-scale legal applications. Future research will focus on multilingual adaptation, computational optimization, and broader applications within the field of legal analytics.

The selection of optimal text embedding models remains a critical challenge in semantic textual similarity (STS) tasks, particularly when performance varies substantially across datasets. In this study, the comparative effectiveness of multiple state-of-the-art embedding models was systematically evaluated using a benchmarking framework based on established machine learning techniques. A range of embedding architectures was examined across diverse STS datasets, with similarity computations performed using Euclidean distance, cosine similarity, and Manhattan distance metrics. Performance evaluation was conducted through Pearson and Spearman correlation coefficients to ensure robust and interpretable assessments. The results revealed that GIST-Embedding-v0 consistently achieved the highest average correlation scores across all datasets, indicating strong generalizability. Nevertheless, MUG-B-1.6 demonstrated superior performance on datasets 2, 6, and 7, while UAE-Large-V1 outperformed other models on datasets 3 and 5, thereby underscoring the influence of dataset-specific characteristics on embedding model efficacy. These findings highlight the importance of adopting a dataset-aware approach in embedding model selection for STS tasks, rather than relying on a single universal model. Moreover, the observed performance divergence suggests that embedding architectures may encode semantic relationships differently depending on domain-specific linguistic features. By providing a detailed evaluation of model behavior across varied datasets, this study offers a methodological foundation for embedding selection in downstream NLP applications. The implications of this research extend to the development of more reliable, scalable, and context-sensitive STS systems, where model performance can be optimized based on empirical evidence rather than heuristics. These insights are expected to inform future investigations on embedding adaptation, hybrid model integration, and meta-learning strategies for semantic similarity tasks.

Significant advancements in artificial intelligence (AI) have transformed clinical decision-making, particularly in disease detection and management. Endometriosis, a chronic and often debilitating gynecological disorder, affects a substantial proportion of reproductive-age women and is associated with pelvic pain, infertility, and a reduced quality of life. Despite its high prevalence, non-invasive and accurate diagnostic methods remain limited, frequently resulting in delayed or missed diagnoses. In this study, a novel diagnostic framework was developed by integrating deep learning (DL) with explainable artificial intelligence (XAI) to address existing limitations in the early and non-invasive detection of endometriosis. Abdominopelvic magnetic resonance imaging (MRI) data were obtained from the Crestview Radiology Center in Victoria Island, Lagos State. Preprocessing procedures, including Digital Imaging and Communications in Medicine (DICOM)-to-PNG conversion, image resizing, and intensity normalization, were applied to standardize the imaging data. A U-Net architecture enhanced with a dual attention mechanism was employed for lesion segmentation, while Gradient-weighted Class Activation Mapping (Grad-CAM) was incorporated to visualize and interpret the model’s decision-making process. Ethical considerations, including informed patient consent, fairness in algorithmic decision-making, and mitigation of data bias, were rigorously addressed throughout the model development pipeline. The proposed system demonstrated the potential to improve diagnostic accuracy, reduce diagnostic latency, and enhance clinician trust by offering transparent and interpretable predictions. Furthermore, the integration of XAI is anticipated to promote greater clinical adoption and reliability of AI-assisted diagnostic systems in gynecology. This work contributes to the advancement of non-invasive diagnostic tools and reinforces the role of interpretable DL in the broader context of precision medicine and women's health.

As market saturation and competitive pressure intensify within the banking sector, the mitigation of customer churn has emerged as a critical concern. Given that the cost of acquiring new clients substantially exceeds that of retaining existing ones, the development of highly accurate churn prediction models has become imperative. In this study, a hybrid customer churn prediction model was developed by integrating Sentence Transformers with a stacking ensemble learning architecture. Customer behavioral data containing textual content was transformed into dense vector representations through the use of Sentence Transformers, thereby capturing contextual and semantic nuances. These embeddings were combined with normalized structured features. To enhance predictive performance, a stacking ensemble method was employed to integrate the outputs of multiple base models, including random forest, Gradient Boosting Tree (GBT), and Support Vector Machine (SVM). Experimental evaluation was conducted on real-world banking data, and the proposed model demonstrated superior performance relative to conventional baseline approaches, achieving notable improvements in both accuracy and the area under the curve (AUC). Furthermore, the analysis of model outputs revealed several salient predictors of customer attrition, such as anomalous transaction behavior, prolonged inactivity, and indicators of dissatisfaction with customer service. These insights are expected to inform the development of targeted intervention strategies aimed at strengthening customer retention, improving satisfaction, and fostering long-term institutional growth and stability.