High-pressure synthesis of lutetium hydrides from molecular hydrogen(H_(2))and lutetium(Lu)is systematically investigated using synchrotron X-ray diffraction,Raman spectroscopy,and visual observations.We demonstrate t...High-pressure synthesis of lutetium hydrides from molecular hydrogen(H_(2))and lutetium(Lu)is systematically investigated using synchrotron X-ray diffraction,Raman spectroscopy,and visual observations.We demonstrate that the reaction pathway between H_(2)and Lu invariably follows the sequence Lu→LuH_(2)→LuH_(3)and exhibits a notable time dependence.A comprehensive diagram representing the formation and synthesis of lutetium hydrides as a function of pressure and time is constructed.Our findings indicate that the synthesis can be accelerated by elevated temperature and decelerated by increased pressure.Notably,two critical pressure thresholds at ambient temperature are identified:the synthesis of LuH_(2)from Lu commences at a minimum pressure of~3 GPa,while~28 GPa is the minimum pressure at which LuH_(2)fails to transform into LuH_(3)within a time scale of months.This underscores the significant impact of temporal factors on synthesis,with the reaction completion time increasing sub-linearly with rising pressure.Furthermore,the cubic phase of LuH_(3)can be obtained exclusively through compressing the trigonal LuH_(3)phase at~11.5 GPa.We also demonstrate that the bandgap of LuH_(3)slowly closes under pressure and is noticeably lower than that of LuH_(2).展开更多
事件抽取是自然语言处理的重要任务,而事件检测是事件抽取的关键步骤之一,其目标是检测事件的发生并对其进行分类。目前基于触发器识别的中文事件检测方法存在一词多义、词与触发词不匹配的问题,影响了事件检测模型的精度。针对此问题,...事件抽取是自然语言处理的重要任务,而事件检测是事件抽取的关键步骤之一,其目标是检测事件的发生并对其进行分类。目前基于触发器识别的中文事件检测方法存在一词多义、词与触发词不匹配的问题,影响了事件检测模型的精度。针对此问题,提出基于双重注意力的无触发词事件检测模型(Event Detection Without Triggers based on Dual Attention,EDWTDA),该模型可跳过触发词识别过程,实现在无触发词标记情况下直接判断事件类型。EDWTDA利用ALBERT改善词嵌入向量的语义表示能力,缓解一词多义问题,提高模型预测能力;采用局部注意力融合事件类型捕捉句中关键语义信息并模拟隐藏的事件触发词,解决词与触发词不匹配的问题;借助全局注意力挖掘文档中的语境信息,解决一词多义问题;最后将事件检测转化成二分类任务,解决多标签问题。同时,采用Focal loss损失函数解决转化成二分类后产生的样本不均衡问题。在ACE2005中文语料库上的实验结果表明,所提模型相比最佳基线模型JMCEE在精确率、召回率和F1-score评价指标上分别提高了3.40%,3.90%,3.67%。展开更多
针对现有人脸检测深度学习算法计算量大,难以移植到嵌入式平台,无法满足移动设备实时性和便捷性需求的问题,提出一种基于YOLO(You Only Look Once)算法的适用于嵌入式平台的小型人脸检测网络E-YOLO(Enhance-YOLO)。借鉴YOLO算法的思想,...针对现有人脸检测深度学习算法计算量大,难以移植到嵌入式平台,无法满足移动设备实时性和便捷性需求的问题,提出一种基于YOLO(You Only Look Once)算法的适用于嵌入式平台的小型人脸检测网络E-YOLO(Enhance-YOLO)。借鉴YOLO算法的思想,将人脸检测问题转换为回归问题,将待检测的图像均分为S×S个单元格,每个单元格检测落在单元格内的目标。通过修改YOLO网络模型中的卷积神经网络结构,提高其检测的准确性,同时减少网络结构中卷积核的数目,降低模型的大小。实验结果表明,E-YOLO模型大小为43MB,视频的检测帧率为26FPS,在WIDERFACE和FDDB数据集上均有较高的准确率和检测速度,可以实现在嵌入式平台下的实时人脸检测。展开更多
基金supported by research grants of the Youth Innovation Promotion Association of CAS(Grant No.2021446)the National Science Foundation of China(Grant Nos.12204484,51672279,12174398 and 11874361)+1 种基金the Anhui Key Research and Development Program(Grant No.2022h11020007)the HFIPS Director’s Fund of the Chinese Academy of Sciences(Grant Nos.BJPY2022B02,YZJJ202102,YZJJ-GGZX-2022-01,and 2021YZGH03).
文摘High-pressure synthesis of lutetium hydrides from molecular hydrogen(H_(2))and lutetium(Lu)is systematically investigated using synchrotron X-ray diffraction,Raman spectroscopy,and visual observations.We demonstrate that the reaction pathway between H_(2)and Lu invariably follows the sequence Lu→LuH_(2)→LuH_(3)and exhibits a notable time dependence.A comprehensive diagram representing the formation and synthesis of lutetium hydrides as a function of pressure and time is constructed.Our findings indicate that the synthesis can be accelerated by elevated temperature and decelerated by increased pressure.Notably,two critical pressure thresholds at ambient temperature are identified:the synthesis of LuH_(2)from Lu commences at a minimum pressure of~3 GPa,while~28 GPa is the minimum pressure at which LuH_(2)fails to transform into LuH_(3)within a time scale of months.This underscores the significant impact of temporal factors on synthesis,with the reaction completion time increasing sub-linearly with rising pressure.Furthermore,the cubic phase of LuH_(3)can be obtained exclusively through compressing the trigonal LuH_(3)phase at~11.5 GPa.We also demonstrate that the bandgap of LuH_(3)slowly closes under pressure and is noticeably lower than that of LuH_(2).
文摘事件抽取是自然语言处理的重要任务,而事件检测是事件抽取的关键步骤之一,其目标是检测事件的发生并对其进行分类。目前基于触发器识别的中文事件检测方法存在一词多义、词与触发词不匹配的问题,影响了事件检测模型的精度。针对此问题,提出基于双重注意力的无触发词事件检测模型(Event Detection Without Triggers based on Dual Attention,EDWTDA),该模型可跳过触发词识别过程,实现在无触发词标记情况下直接判断事件类型。EDWTDA利用ALBERT改善词嵌入向量的语义表示能力,缓解一词多义问题,提高模型预测能力;采用局部注意力融合事件类型捕捉句中关键语义信息并模拟隐藏的事件触发词,解决词与触发词不匹配的问题;借助全局注意力挖掘文档中的语境信息,解决一词多义问题;最后将事件检测转化成二分类任务,解决多标签问题。同时,采用Focal loss损失函数解决转化成二分类后产生的样本不均衡问题。在ACE2005中文语料库上的实验结果表明,所提模型相比最佳基线模型JMCEE在精确率、召回率和F1-score评价指标上分别提高了3.40%,3.90%,3.67%。
文摘针对现有人脸检测深度学习算法计算量大,难以移植到嵌入式平台,无法满足移动设备实时性和便捷性需求的问题,提出一种基于YOLO(You Only Look Once)算法的适用于嵌入式平台的小型人脸检测网络E-YOLO(Enhance-YOLO)。借鉴YOLO算法的思想,将人脸检测问题转换为回归问题,将待检测的图像均分为S×S个单元格,每个单元格检测落在单元格内的目标。通过修改YOLO网络模型中的卷积神经网络结构,提高其检测的准确性,同时减少网络结构中卷积核的数目,降低模型的大小。实验结果表明,E-YOLO模型大小为43MB,视频的检测帧率为26FPS,在WIDERFACE和FDDB数据集上均有较高的准确率和检测速度,可以实现在嵌入式平台下的实时人脸检测。
