Towards Stress-Aware Sentence-Level Filipino G2P With Weakly-Supervised ByT5 Fine-Tuning
研究通过使用ByT5模型微调来解决菲律宾语句子级G2P转换中包含重音特征的问题,利用有限数据获得较好性能。
研究通过使用ByT5模型微调来解决菲律宾语句子级G2P转换中包含重音特征的问题,利用有限数据获得较好性能。
研究通过对比传统、AI优先和混合三种模式下用户在嵌入LLM的内容管理系统中的行为,发现AI辅助主要减少了交互努力而非提高速度,并且用户的委托行为更多取决于个人差异。
This study addresses the challenge of visualizing uncertainty in crowdsourced ordinal crowding data to support commuters in making better-informed travel decisions. Through an online user experiment, it presents the first systematic comparison of multiple visualization techniques—including cluster plots and bubble treemaps—in their ability to convey data variability and reliability. The study comprehensively evaluates how these visualizations influence users’ cognitive load, trust, and judgment accuracy. Results indicate that cluster plots significantly reduce cognitive load and enhance trust, whereas bubble treemaps yield higher accuracy in assessing crowding levels. These findings provide empirical evidence and actionable design guidance for visualizing uncertainty in crowding scenarios.
This study addresses the risk of novice programmers over-relying on generative AI, which may undermine critical thinking in programming education. To mitigate this, the authors propose a learner-centered, empathetic AI-assisted paradigm that integrates empathy mechanisms into the programming environment, emphasizing affective feedback during error correction rather than direct code generation. They designed and implemented an empathetic C-language IDE named Ceci and evaluated it through a controlled experiment (n=11) against VSCode augmented with ChatGPT, using the NASA-TLX cognitive workload scale and usability surveys. Although Ceci showed no significant differences in task completion, cognitive load, or overall usability compared to the baseline, it demonstrated significantly higher perceived helpfulness in error correction (p=0.022), underscoring the unique value of empathetic feedback in supporting learning.
This work addresses a critical limitation in existing physical adversarial attack research, which predominantly relies on single-frame image evaluations and fails to capture the robustness of real-world surveillance systems under temporal continuity, multi-sensor fusion, and practical deployment constraints. To bridge this gap, the authors propose a four-dimensional evaluation framework tailored for surveillance systems, encompassing temporal identity consistency, visible-infrared dual-modality evasion, controllable wearable carrier feasibility, and system-level target alignment. By integrating multi-object tracking, dual-modality adversarial sample generation, and real-world deployment tests, the study exposes the inadequacy of single-frame assessments and underscores the necessity of system-level robustness validation across time, modalities, and realistic operational conditions. It further highlights key challenges such as distance-dependent robustness and discrepancies in camera processing pipelines.
研究通过使用ByT5模型微调来解决菲律宾语句子级G2P转换中包含重音特征的问题,利用有限数据获得较好性能。
研究通过对比传统、AI优先和混合三种模式下用户在嵌入LLM的内容管理系统中的行为,发现AI辅助主要减少了交互努力而非提高速度,并且用户的委托行为更多取决于个人差异。
This study addresses the challenge of visualizing uncertainty in crowdsourced ordinal crowding data to support commuters in making better-informed travel decisions. Through an online user experiment, it presents the first systematic comparison of multiple visualization techniques—including cluster plots and bubble treemaps—in their ability to convey data variability and reliability. The study comprehensively evaluates how these visualizations influence users’ cognitive load, trust, and judgment accuracy. Results indicate that cluster plots significantly reduce cognitive load and enhance trust, whereas bubble treemaps yield higher accuracy in assessing crowding levels. These findings provide empirical evidence and actionable design guidance for visualizing uncertainty in crowding scenarios.
This study addresses the risk of novice programmers over-relying on generative AI, which may undermine critical thinking in programming education. To mitigate this, the authors propose a learner-centered, empathetic AI-assisted paradigm that integrates empathy mechanisms into the programming environment, emphasizing affective feedback during error correction rather than direct code generation. They designed and implemented an empathetic C-language IDE named Ceci and evaluated it through a controlled experiment (n=11) against VSCode augmented with ChatGPT, using the NASA-TLX cognitive workload scale and usability surveys. Although Ceci showed no significant differences in task completion, cognitive load, or overall usability compared to the baseline, it demonstrated significantly higher perceived helpfulness in error correction (p=0.022), underscoring the unique value of empathetic feedback in supporting learning.
This work addresses a critical limitation in existing physical adversarial attack research, which predominantly relies on single-frame image evaluations and fails to capture the robustness of real-world surveillance systems under temporal continuity, multi-sensor fusion, and practical deployment constraints. To bridge this gap, the authors propose a four-dimensional evaluation framework tailored for surveillance systems, encompassing temporal identity consistency, visible-infrared dual-modality evasion, controllable wearable carrier feasibility, and system-level target alignment. By integrating multi-object tracking, dual-modality adversarial sample generation, and real-world deployment tests, the study exposes the inadequacy of single-frame assessments and underscores the necessity of system-level robustness validation across time, modalities, and realistic operational conditions. It further highlights key challenges such as distance-dependent robustness and discrepancies in camera processing pipelines.