Japanese language learning has long been constrained by acquisition difficulties arising from linguistic distance and by the limited availability of authentic opportunities for conversational interaction. Although mobile learning technologies offer considerable potential to alleviate these constraints, existing tools remain largely focused on vocabulary acquisition and grammar practice, while the dynamic progression of learner–artificial intelligence conversational interaction has received insufficient attention. In this study, a mobile Japanese conversational interaction system powered by a large language model was developed and integrated with automatic speech recognition and speech synthesis interfaces. Multi-topic conversational environments were constructed for learners at different proficiency levels. A four-week experiment involving 76 learners of Japanese was conducted, and interaction logs, JapaneseLanguage Proficiency Test–aligned mock assessments, and expert ratings were collected to characterize trajectories of conversation deepening and examine their associations with language learning outcomes. Pseudotime trajectory inference based on a minimum spanning tree demonstrated high internal consistency and external validity, and the composite conversation-depth score accounted for 43.8% of the variance in learner performance. Within individual conversational sessions, conversation depth increased significantly as the number of interactions turns increased. Intermediate learners exhibited greater mean pseudotime progression, more branch points, and faster rates of conversation deepening than beginner learners. These findings demonstrate that temporal modeling of interaction trajectories can be used to represent conversation deepening in generative artificial intelligence–supported mobile Japanese language learning as a quantifiable dynamic process. This approach provides a quantitative basis for investigating mobile conversation-based learning mechanisms and informing the design of adaptive conversational interactions.
New pathways for immersive presentation and community participation in the living transmission of intangible cultural heritage along the Grand Canal are provided by mobile augmented reality. However, severe challenges to spatial registration accuracy, robustness to network disconnection, and content update mechanisms are posed by weak network coverage in linear heritage environments, limited mobile-device computing power, and the need for intangible cultural heritage content to grow dynamically. A mobile augmented reality method centered on semantic anchors was proposed. Semantic anchors were generated through joint factor graph optimization of lightweight visual semantic features and global navigation satellite system/inertial measurement unit measurements, forming a geographically consistent and compact spatial index. Each anchor was stored compactly at the hundreds-of-bytes level, and decimeter-level spatial registration was achieved without prebuilt high-precision maps. On this basis, a device-edge-cloud collaborative architecture was constructed. Semantic anchors were simultaneously employed as persistent primitives for cross-session pose recovery and as spatial containers for user-generated intangible cultural heritage micro-narratives so that second-level relocalization and continuous display were supported under fully disconnected conditions. Through edge-side lightweight quality assessment and context-aware recommendation, community-based accumulation and targeted distribution of living content were achieved. It is demonstrated by the experimental results that, for mobile augmented reality-driven living transmission of large-scale linear cultural heritage, a lightweight, robust, and engineering-feasible technical pathway is provided by the semantic-anchor method.
Adult mobile learning is characterized by fragmentation, task orientation, and contextembedded features, which pose substantial challenges to conventional adaptive systems in generating real-time interaction strategies that simultaneously accommodate learners’ cognitive states and contextual constraints on mobile devices. To address these challenges, Mobile Adult Learning Adaptive Interaction Model (MAL-ADAPT), a context–cognition dual-channel adaptive interaction model driven by a lightweight on-device large language model agent, was proposed. Adaptive interaction was formulated as a constrained partially observable Markov decision process, in which learning gain, engagement, and learner autonomy were jointly optimized subject to latency, energy, memory, and privacy constraints. At the perception layer, behavioral, device, and environmental signals were integrated to construct a contextual state representation. At the cognitive layer, mastery probabilities and associated uncertainty were estimated through low-rank knowledge tracing. At the interaction layer, a unified state representation was constructed through a gated interaction mechanism. State-conditioned soft prompts were then injected into the on-device large language model, and executable adaptive actions were generated through low-rank adaptation-based parameter adaptation, action-space mapping, an adult-learning adaptation score, and feasibility masking. In addition, an edge–cloud hybrid inference gate was introduced to invoke cloud-enhanced inference under complex conditions when network connectivity permitted, while direct transmission of raw behavioral data was avoided. Six groups of experiments were conducted to evaluate component contributions, on-device performance, policy effectiveness, real-world use, robustness, and constraint sensitivity. The results demonstrate that MAL-ADAPT improved learner engagement, strategy acceptance, and self-regulated learning under mobile resource constraints while reducing response latency and dependence relative to cloud-based approaches. These findings provide a reusable technical pathway for on-device educational agents.
Human–artificial intelligence (AI) collaborative decision-making in mobile accounting systems is constrained by static division-of-labor boundaries and AI confidence bias, making automation efficiency and error control difficult to balance. A dynamic confidence threshold framework was proposed in this study. Confidence was calibrated through Platt scaling, reducing the expected calibration error to 0.031. An adaptive threshold function was constructed by integrating task complexity and historical error rates, and manual review task allocation was optimized by incorporating operator experience. Tasks were dynamically routed into three channels: automatic approval, human–AI collaborative correction, and in-depth manual review. Selective-risk statistical monitoring and drift detection mechanisms were also incorporated. Experimental results on a real-world accounting dataset and a mobile prototype system indicate that the auto-pass error rate was controlled at 4.2%, the automation rate reached 78.6%—an improvement of 13.3 percentage points over the optimal fixed threshold— accountant processing time was reduced by 23.0%, and cognitive load was reduced by 26.8%; favorable steady-state recovery was further exhibited under multiple perturbations. A systematic technical solution is thus provided for human–AI collaborative decision-making in mobile accounting scenarios.
Demand for digital forensics (DF) graduates is rising, yet undergraduate DF teaching still depends on costly desktop tools, fixed laboratory slots, and limited time for repeated practice— barriers that weigh most heavily on resource-constrained universities. This paper presents DFLearn-Mobile, a cross-platform mobile application for undergraduate DF education built with Flutter, Node.js, and MongoDB. The app delivers six short modules aligned with the NIST DF process model, each closing with an interactive task and quiz, plus a capstone branching case for end-to-end procedural reasoning. We evaluated the platform with N = 80 undergraduates at Rajamangala University of Technology Suvarnabhumi using a quasi-experimental design combining pre- and post-tests, the system usability scale (SUS), an adapted technology acceptance model (TAM) instrument, and server-side analytics. DF knowledge improved significantly (Wilcoxon V = 3,240, p < .001; all 80 learners improved; mean normalized gain = 0.586, medium band). The SUS score (M = 61.03) fell significantly below the 68 benchmark, identifying usability priorities for the next iteration. TAM construct means (3.85–3.95) indicated mild positive acceptance, with PU ↔ BI the strongest correlation (rho = 0.458, p < .001). The framework, instruments, and analytic toolkit offer a replicable basis for mobile DF teaching in resource-constrained settings. These are preliminary pilot findings; causal claims require a randomized controlled comparison.
