This study examines how Cultural Relic AI Agents (CRAAs) foster users' cultural identification through interactive experiences with Sanxingdui artifacts. Addressing the limitations of superficial digital interactions with cultural heritage, we investigate the roles of anthropomorphism and interactivity in CRAA design, as well as their impact on users' perceived value and subsequent cultural identification formation. Through empirical analysis, we demonstrate that CRAA can facilitate deeper engagement with complex cultural content, offering theoretical insights and practical guidance for AI-driven strategies to enhance cultural dissemination and identity building.
Comics are a distinctive visual medium that has gained significant attention in recent years, but recognizing and analyzing them poses several challenges. These include the extraction and classification of comic elements (speech balloons, text, panels and characters) for different purposes. The systematic literature review (SLR) examines recent research on comic recognition (CR) approaches, datasets, and analyses of various recognition models. A thorough search of multiple databases (including Scopus, ACM, Google Scholar, Springer, and IEEE Xplore) was undertaken; 88 studies were identified based on inclusion/exclusion criteria, covering 2010 to 2025. The research identifies prominent parameters in comic images, including speech balloons (37.5%), panels and comic characters (20.5% each). The most common datasets, such as eBDtheque (36.8%) and Manga109 (34.7%), are used in the studies. The studies demonstrate that 57.2% used segmentation, and 42.8% used classification techniques to extract information from images. This study provides an overview of how comics can be analyzed and exploited for communication. The scientific study and exploitation of robust hybrid approaches combining classical image processing and deep learning is an unexplored area with significant future potential for improving accuracy.
Self-determination theory is a prominent theoretical framework within human-computer interaction (HCI) research on video games, with need satisfaction being one of the fundamental concepts used to investigate player experience. Although a handful of studies have tapped into need satisfaction in esports, they have generally focused on its role in explaining self-determined motivation and related outcomes. However, to date, research on psychological need satisfaction in esports participation remains scarce. Self-determination theory posits that the ways in which autonomy, competence, and relatedness are experienced and satisfied may vary across individuals and cultural contexts. Accordingly, this study explores the qualitatively different ways in which players experience need satisfaction in esports participation in the Chinese context. Findings indicate three hierarchically related categories of autonomy satisfaction, two of competence satisfaction, and three of relatedness satisfaction. These categories reflect increasingly inclusive and complex ways of experiencing need satisfaction in esports participation. The identified hierarchical structures may inform the development of measures of need satisfaction in esports participation. Importantly, the present study offers a novel perspective on the role of identity in esports participation, with the findings from the present phenomenographic outcome space suggesting that esports-related identity is more closely associated with relatedness satisfaction through social belonging than with autonomy satisfaction through self-expression.
Artificial Intelligence (AI) driven gamification has emerged as an effective approach for improving learner participation and personalized knowledge acquisition in digital financial learning ecosystems. The current financial learning platforms, however, have some shortcomings in personalization, reward structures are fixed, there is a lack of adaptive difficulty adjustment, and the understanding of corporate tax concepts is insufficient, which leads to a low level of engagement and retention of financial knowledge in learning activities. This research provide that the Knowledge-Guided Adaptive Tax Learning Network (KGATL-Net) to meet these challenges by providing intelligent and personalized learning of corporate tax compliance through maximizing the participation of users and learning effectiveness. The framework proposed combines TabTransformer, a learner profiling method, with Knowledge Graph (KG) reasoning to generate semantic concepts of a task, Proximal Policy Optimization (PPO) for adaptive recommendations, a Dynamic Reward Mechanism (DRM) for personalized rewards, and Adaptive Challenge Difficulty (ACD) for dynamically adapting task difficulty based on learner performance. They work together to create tailored learning paths, enhance engagement, and boost knowledge retention by adapting gamification elements. The proposed framework is experimentally evaluated and it is shown that the proposed framework outperforms other frameworks with 97.18% accuracy. Overall, KGATL-Net forms a powerful, scalable, and intelligent platform for next generation personalized digital financial learning environments.
Automated table tennis post-match analysis with large language models (LLMs) faces two challenges: representing match information for effective reasoning, and efficiently processing lengthy match narratives. We propose a framework with three contributions. First, an information representation combining a practical four-dimensional attribute system (Player, Technique, Landing, Quality) under annotation constraints with a hierarchical semantic encoding scheme; ablations confirm that removing individual attributes causes 9%–18% quality degradation (n=120), and hierarchical encoding outperforms flat formats by 7.9%–33.2%. Second, a hierarchical query framework that exploits the inherent Match → Game → Rally structure, improving over standard (+10.4%), structured (+6.9%), and multi-chain (+6.1%) retrieval baselines with 82% context reduction. Third, within-Qwen-family observations revealing that mid-capacity variants (7B–32B) gain the most from structured representation. Within structured racket-sport analysis and the evaluated Qwen family, a 7B model with structured representation closes the performance gap with a 72B model using raw representation (Comprehensive Score 8.71 vs. 6.92), reaches approximately 91.7% of the expert reference level on a single NVIDIA A80 GPU, and reduces the per-match tactical-reasoning turnaround from roughly 45 min under a manual expert workflow to 2.4 min under the automated pipeline, measured after TTML attributes are prepared and excluding the one-time offline annotation.
