The temporal dynamics in brain evoked by the scale of visual attention with the cues of Chinese characters were studied by recording event-related potentials (ERPs). With the fixed orientation of visual attention, 14 ...The temporal dynamics in brain evoked by the scale of visual attention with the cues of Chinese characters were studied by recording event-related potentials (ERPs). With the fixed orientation of visual attention, 14 healthy young participants performed a search task in which the search array was preceded by Chinese characters cues, '大, 中, 小' (large, medium, small). 128 channels scalp ERPs were recorded to study the role of visual attention scale played in the visual spatial attention. The results showed that there was no significant difference in the ERP components evoked by the three Chinese characters cues except the in-feroposterior N2 latency. The targets evoked P2, N2 amplitudes and latency have significant differences with the different cues of large, middle and small, while P1 and N1 components had no significant difference. The results suggested that the processing of scale of visual attention was mainly concerned with P2, N2 components, while the P1, N1 components were mainly related with展开更多
目的手写汉字纠错(handwritten Chinese character error correction,HCCEC)任务具有两重性,即判断汉字正确性和对错字进行纠正,该任务在教育场景下应用广泛,可以帮助学生学习汉字、纠正书写错误。由于手写汉字具有复杂的空间结构、多...目的手写汉字纠错(handwritten Chinese character error correction,HCCEC)任务具有两重性,即判断汉字正确性和对错字进行纠正,该任务在教育场景下应用广泛,可以帮助学生学习汉字、纠正书写错误。由于手写汉字具有复杂的空间结构、多样的书写风格以及巨大的数量,且错字与正确字之间具有高度的相似性,因此,手写汉字纠错的关键是如何精确地建模一个汉字。为此,提出一种层级部首网络(hierarchical radical network,HRN)。方法从部首字形的角度出发,挖掘部首形状结构上的相似性,通过注意力模块捕获包含部首信息的细粒度图像特征,增大相似字之间的区分性。另外,结合汉字本身的层级结构特性,采用基于概率解码的思路,对部首的层级位置进行建模。结果在手写汉字数据集上进行实验,与现有方案相比,HRN在正确字测试集与错字测试集上,精确率分别提升了0.5%和9.8%,修正率在错字测试集上提升了15.3%。此外,通过注意力机制的可视化分析,验证了HRN可以捕捉包含部首信息的细粒度图像特征。部首表征之间的欧氏距离证明了HRN学习到的部首表征向量中包含了部首的字形结构信息。结论本文提出的HRN能够更好地对相似部首进行区分,进而精确地区分正确字与错字,具有很强的鲁棒性和泛化性。展开更多
基金This work was supported by theNational Natural Science Foundation of China (Grant No. 39970257)the Multidisciplinary Research Program of the Chinese Academy of Sciences (Grant No. KJCXl-07)by the Hundred Talents Program of the the Chinese Academy o
文摘The temporal dynamics in brain evoked by the scale of visual attention with the cues of Chinese characters were studied by recording event-related potentials (ERPs). With the fixed orientation of visual attention, 14 healthy young participants performed a search task in which the search array was preceded by Chinese characters cues, '大, 中, 小' (large, medium, small). 128 channels scalp ERPs were recorded to study the role of visual attention scale played in the visual spatial attention. The results showed that there was no significant difference in the ERP components evoked by the three Chinese characters cues except the in-feroposterior N2 latency. The targets evoked P2, N2 amplitudes and latency have significant differences with the different cues of large, middle and small, while P1 and N1 components had no significant difference. The results suggested that the processing of scale of visual attention was mainly concerned with P2, N2 components, while the P1, N1 components were mainly related with
文摘目的手写汉字纠错(handwritten Chinese character error correction,HCCEC)任务具有两重性,即判断汉字正确性和对错字进行纠正,该任务在教育场景下应用广泛,可以帮助学生学习汉字、纠正书写错误。由于手写汉字具有复杂的空间结构、多样的书写风格以及巨大的数量,且错字与正确字之间具有高度的相似性,因此,手写汉字纠错的关键是如何精确地建模一个汉字。为此,提出一种层级部首网络(hierarchical radical network,HRN)。方法从部首字形的角度出发,挖掘部首形状结构上的相似性,通过注意力模块捕获包含部首信息的细粒度图像特征,增大相似字之间的区分性。另外,结合汉字本身的层级结构特性,采用基于概率解码的思路,对部首的层级位置进行建模。结果在手写汉字数据集上进行实验,与现有方案相比,HRN在正确字测试集与错字测试集上,精确率分别提升了0.5%和9.8%,修正率在错字测试集上提升了15.3%。此外,通过注意力机制的可视化分析,验证了HRN可以捕捉包含部首信息的细粒度图像特征。部首表征之间的欧氏距离证明了HRN学习到的部首表征向量中包含了部首的字形结构信息。结论本文提出的HRN能够更好地对相似部首进行区分,进而精确地区分正确字与错字,具有很强的鲁棒性和泛化性。
