In the digital age, the rapid access of children to online multimedia platforms presents significant risks related to exposure to violent and inappropriate content. This paper proposes a component-based ontological framework for the semantic modeling and detection of violent content in children’s digital media. By integrating domain knowledge and leveraging labeled datasets such as the Real-Life Violence Situations Dataset and Violence-Detect1 Dataset, the proposed Component-based Violence Detection Ontology (CBVD-Onto) enhances the ability to identify and classify violent behaviors and objects that may jeopardize child safety. The ontology is designed with core components including Event, Actor, Action, Context, and Media clusters, enabling scalable, extensible, and interoperable semantic representation. This approach addresses current challenges faced by automated content moderation systems, notably the difficulty in detecting subtle or context-dependent violent scenes. The model supports improved content filtering and proactive safeguarding of children in various digital environments. Experimental validation highlights the ontology’s effectiveness in integrating heterogeneous data sources and facilitating advanced reasoning to better protect children from harmful content.
The proliferation of code-mixing, especially Hinglish (Hindi-English), on online conversation platforms presents enormous challenges to state-of-the-art machine translation (MT) systems. Hinglish text, in its fluid intra-sentential language switching, Roman script usage over Hindi words, and colloquial grammatical structures, is likely to be handled badly by standard translation models fine-tuned on monolingual data. In this paper, we address the task of code-mixed Hinglish-to-standard English translation with the help of the strength of transformer models. We fine-tuned a T5-base model on a wisely curated dataset of 50,000 Hinglish-English parallel sentence pairs with a layer-freezing strategy where the first 6 layers were left frozen during training. To specifically target improvements in translation fluency, coherence, and contextual appropriateness, we introduce and add new loss function components in the fine-tuning process: Smoothness Loss, Multi-Task Gradient Alignment Loss (MTGAL, with the assumption of an auxiliary NER task being feasible), and Attention Entropy Regularization. These components are intended to regularize the internal representation of the model and attention. We compare the performance of our fine-tuned model with standard metrics, such as BLEU and ROUGE (ROUGE-1, ROUGE-2, ROUGE-L), computed on a test set of 20,000 samples, and present preliminary qualitative results. We also discuss the impact of limited computational resources, specifically the use of Kaggle P100 GPUs, on the scope of our experimentation. This paper contributes to the building of stronger and more effective MT systems for low-resource, code-mixed language pairs commonly found in multilingual populations.
Psychological problems, sometimes referred to as mental disorders or psychopathologies, are thought, feeling, or behavior patterns that considerably depart from cultural norms and frequently cause psychological suffering or functional difficulties. These conditions can affect many facets of daily living, including thinking, feeling, interacting with others, and general functioning. Numerous factors, including genetic, biochemical, environmental, and psychosocial factors, might contribute to psychological illnesses. This covers a wide range of illnesses, including personality disorders, psychotic disorders, depression, and anxiety. However, the extensive use of social media presents a special chance to examine user characteristics and behavior. With an emphasis on identifying psychological problems, this research investigates the use of deep learning for author personality trait detection. We introduce an approach for social media text analysis that incorporates cutting-edge deep neural algorithms. We build and assess deep learning models to identify markers of psychological diseases, including anxiety and depression, using annotated datasets. This paper suggests methods to increase the precision and dependability of predictive models while highlighting the moral and practical ramifications of automated psychological disease detection. This study attempts to apply a few deep learning architectures in order to compare them using a thorough analytic approach based on the accuracy outcome. The findings, which had an average accuracy of 93
This study explores the development of an English-to-Tunisian Arabic (Derja) neural machine translation (NMT) system designed for real-time interaction within medical simulation environments. The work primarily addresses the broader challenge of improving translation quality for low-resource dialects, where parallel corpora are extremely scarce. To overcome this limitation, a synthetic parallel corpus and a specialized medical lexicon were constructed to support domain adaptation. Several state-of-the-art multilingual models (mBART-50, M2M-100, NLLB-200, and its distilled variant) were fine-tuned and compared using BLEU, chrF++, COMET, and BERTScore metrics. Results show that the NLLB-200 (1.3B) model consistently achieves the best performance after fine-tuning, reaching a BLEU score of 15.31, a chrF++ of 47.18, a COMET score of 0.6980, and a BERTScore of 0.6976. These findings highlight the potential of domain-oriented adaptation and lexical enhancement to improve translation accuracy for under-resourced dialects like Derja, enabling more natural and effective communication in simulation contexts.
Intelligent systems, with their crucial rule-based components, are fundamental to modern digital infrastructure. As these systems grow, direct conflicts between rules inevitably arise, threatening consistency and the correctness of system outputs. Effectively resolving these rule conflicts is therefore a critical task for maintaining robust and dependable intelligent systems. This paper presents a method for detecting and resolving direct rule conflicts, representing a step toward fully automated conflict management. The proposed two-phase process involves (1) systematic grouping of rules with shared antecedent–consequent structures and (2) pattern-based conflict analysis supported by type-specific resolution strategies. Our method formalizes conflict handling through dependency analysis, conflict pattern recognition, and guided resolution. Experiments conducted on diverse and complex rule bases demonstrate the method’s effectiveness and efficiency in identifying and mitigating conflicts, thereby preserving the consistency of the rule base. This work contributes a structured and scalable foundation for advancing automation in conflict resolution within rule-based intelligent systems.
