针对双关语样本短缺问题,研究提出了基于伪标签和迁移学习的双关语识别模型(pun detection based on Pseudo-label and transfer learning)。该模型利用上下文语义、音素向量和注意力机制生成伪标签;然后,迁移学习和置信度结合挑选可用...针对双关语样本短缺问题,研究提出了基于伪标签和迁移学习的双关语识别模型(pun detection based on Pseudo-label and transfer learning)。该模型利用上下文语义、音素向量和注意力机制生成伪标签;然后,迁移学习和置信度结合挑选可用的伪标签;最后,将伪标签数据和真实数据混合到网络中进行训练,重复伪标签标记和混合训练过程。一定程度上解决了双关语样本量少且获取困难的问题。使用该模型在SemEval 2017 shared task 7以及Pun of the Day数据集上进行双关语检测实验,结果表明模型性能均优于现有主流双关语识别方法。展开更多
For many years, nerve transfer has been commonly used as a treatment option following peripheral nerve injury, although the precise mechanism underlying successful nerve transfer is not yet clear. We developed an anim...For many years, nerve transfer has been commonly used as a treatment option following peripheral nerve injury, although the precise mechanism underlying successful nerve transfer is not yet clear. We developed an animal model to investigate the mechanism underlying nerve transfer between branches of the spinal accessory nerve (Ac) and suprascapular nerve (Ss) in rats, so that we could observe changes in the number of motor neurons, investigate the 3-dimensional localization of neurons in the anterior horn of the spinal cord, and perform an electromyogram (EMG) of the supraspinatus muscle before and after nerve transfer treatment. The present experiment showed a clear reduction in the number of γ motor neurons. The distributional portion of motor neurons following nerve transfer was mainly within the neuron column innervating the trapezius. Some neurons innervating the supraspinatus muscle also survived post-transfer. Compared with the non-operated group, the EMG restoration rate of the supraspinatus muscle following nerve transfer was 60% in the experimental group and 80% in a surgical control group. Following nerve transfer, there was a distinct reduction in the number of γ motor neurons. Therefore, γ motor neurons may have important effects on the recovery of muscular strength following nerve transfer. Moreover, because the neurons located in regions innervating either the trapezius or supraspinatus muscle were labeled after Ac transfer to Ss, we also suggest that indistinct axon regeneration mechanisms exist in the spinal cord following peripheral nerve transfer.展开更多
Exogenous DNA expressing green fluorescent protein( GFP) and labeled with fluorescein isothiocyanate( FITC) was used to transform the Chinese oak silkmoth Antheraea pernyi( A. pernyi)via sperm-mediated gene transfer( ...Exogenous DNA expressing green fluorescent protein( GFP) and labeled with fluorescein isothiocyanate( FITC) was used to transform the Chinese oak silkmoth Antheraea pernyi( A. pernyi)via sperm-mediated gene transfer( SMGT). Sperms entry into the female reproductive system and eggs were observed using fluorescence microscopy. The ability of A. pernyi sperms to uptake exogenous DNA was confirmed,and transfer of the exogenous DNA was shown by GFP expression in the transgenic eggs. Our result suggested that SMGT could also be used to directly generate transgenic A. pernyi expressing functional genes of interest.展开更多
Ectomycorrhizal(EM)networks provide a variety of services to plants and ecosystems include nutrient uptake and transfer,seedling survival,internal cycling of nutrients,plant competition,and so on.To deeply their struc...Ectomycorrhizal(EM)networks provide a variety of services to plants and ecosystems include nutrient uptake and transfer,seedling survival,internal cycling of nutrients,plant competition,and so on.To deeply their structure and function in ecosystems,we investigated the spatial patterns and nitrogen(N)transfer of EM networks usingN labelling technique in a Mongolian scotch pine(Pinus sylvestris var.mongolica Litv.)plantation in Northeastern China.In August 2011,four plots(20 × 20 m)were set up in the plantation.125 ml 5 at.%0.15 mol/LNHNOsolution was injected into soil at the center of each plot.Before and 2,6,30 and 215 days after theN application,needles(current year)of each pine were sampled along four 12 m sampling lines.Needle total N andN concentrations were analyzed.We observed needle N andN concentrations increased significantly over time afterN application,up to 31 and0.42%,respectively.There was no correlation between needle N concentration andN/N ratio(R2=0.40,n=5,P=0.156),while excess needle N concentration and excess needleN/N ratio were positively correlated across different time intervals(R~2=0.89,n=4,P\0.05),but deceased with time interval lengthening.NeedleN/N ratio increased with time,but it was not correlated with distance.NeedleN/N ratio was negative with distance before and 6th day and 30th day,positive with distance at 2nd day,but the trend was considerably weaker,their slop were close to zero.These results demonstrated that EM networks were ubiquitous and uniformly distributed in the Mongolian scotch pine plantation and a random network.We found N transfer efficiency was very high,absorbed N by EM network was transferred as wide as possible,we observed N uptake of plant had strong bias forN andN,namely N fractionation.Understanding the structure and function of EM networks in ecosystems may lead to a deeper understanding of ecological stability and evolution,and thus provide new theoretical approaches to improve conservation practices for the management of the Earth’s ecosystems.展开更多
胸部疾病高发,且有些疾病种类的癌症转变率很高,因此基于卷积神经网络的胸部X光图像疾病自动检测分类方法是计算机辅助诊断的研究热点之一.然而,目前的自动分类方法仍面临胸部病灶的X光图像特异性特征表达不充分、不同胸部疾病发病率不...胸部疾病高发,且有些疾病种类的癌症转变率很高,因此基于卷积神经网络的胸部X光图像疾病自动检测分类方法是计算机辅助诊断的研究热点之一.然而,目前的自动分类方法仍面临胸部病灶的X光图像特异性特征表达不充分、不同胸部疾病发病率不平衡、卷积神经网络参数量过大等问题.针对上述问题,提出了一种端到端的基于八度卷积的ResNet(octave convolution based residual network,OC-ResNet)结构.首先,利用八度卷积改进ResNet中的普通卷积,将高低频特征分离,增强对高频信息的提取,以更好地表达胸部病灶的特异性特征,降低模型计算复杂度.其次,利用渐进式迁移学习,将OC-ResNet在ImageNet数据集进行预训练,获得网络的初始参数,然后固定网络浅层参数,在ChestX-Ray14数据集上微调网络深层参数.最后,为改善样本不平衡问题,网络训练时,采用了焦点损失函数,增加样本数较少类别的权重.在ChestX-Ray14数据集上的实验结果表明,OC-ResNet对14种胸部疾病分类的平均AUC值达到0.856,与目前先进的深度学习方法相比,其中13种疾病分类的AUC值达到最优,同时,计算复杂度相比基础网络降低了44.77%.展开更多
文摘针对双关语样本短缺问题,研究提出了基于伪标签和迁移学习的双关语识别模型(pun detection based on Pseudo-label and transfer learning)。该模型利用上下文语义、音素向量和注意力机制生成伪标签;然后,迁移学习和置信度结合挑选可用的伪标签;最后,将伪标签数据和真实数据混合到网络中进行训练,重复伪标签标记和混合训练过程。一定程度上解决了双关语样本量少且获取困难的问题。使用该模型在SemEval 2017 shared task 7以及Pun of the Day数据集上进行双关语检测实验,结果表明模型性能均优于现有主流双关语识别方法。
文摘For many years, nerve transfer has been commonly used as a treatment option following peripheral nerve injury, although the precise mechanism underlying successful nerve transfer is not yet clear. We developed an animal model to investigate the mechanism underlying nerve transfer between branches of the spinal accessory nerve (Ac) and suprascapular nerve (Ss) in rats, so that we could observe changes in the number of motor neurons, investigate the 3-dimensional localization of neurons in the anterior horn of the spinal cord, and perform an electromyogram (EMG) of the supraspinatus muscle before and after nerve transfer treatment. The present experiment showed a clear reduction in the number of γ motor neurons. The distributional portion of motor neurons following nerve transfer was mainly within the neuron column innervating the trapezius. Some neurons innervating the supraspinatus muscle also survived post-transfer. Compared with the non-operated group, the EMG restoration rate of the supraspinatus muscle following nerve transfer was 60% in the experimental group and 80% in a surgical control group. Following nerve transfer, there was a distinct reduction in the number of γ motor neurons. Therefore, γ motor neurons may have important effects on the recovery of muscular strength following nerve transfer. Moreover, because the neurons located in regions innervating either the trapezius or supraspinatus muscle were labeled after Ac transfer to Ss, we also suggest that indistinct axon regeneration mechanisms exist in the spinal cord following peripheral nerve transfer.
基金Scientific Research Project for High Schools of the Educational Department of Liaoning Province,China(No.2008643)
文摘Exogenous DNA expressing green fluorescent protein( GFP) and labeled with fluorescein isothiocyanate( FITC) was used to transform the Chinese oak silkmoth Antheraea pernyi( A. pernyi)via sperm-mediated gene transfer( SMGT). Sperms entry into the female reproductive system and eggs were observed using fluorescence microscopy. The ability of A. pernyi sperms to uptake exogenous DNA was confirmed,and transfer of the exogenous DNA was shown by GFP expression in the transgenic eggs. Our result suggested that SMGT could also be used to directly generate transgenic A. pernyi expressing functional genes of interest.
基金supported by National Natural Science Foundation of China(30830024)
文摘Ectomycorrhizal(EM)networks provide a variety of services to plants and ecosystems include nutrient uptake and transfer,seedling survival,internal cycling of nutrients,plant competition,and so on.To deeply their structure and function in ecosystems,we investigated the spatial patterns and nitrogen(N)transfer of EM networks usingN labelling technique in a Mongolian scotch pine(Pinus sylvestris var.mongolica Litv.)plantation in Northeastern China.In August 2011,four plots(20 × 20 m)were set up in the plantation.125 ml 5 at.%0.15 mol/LNHNOsolution was injected into soil at the center of each plot.Before and 2,6,30 and 215 days after theN application,needles(current year)of each pine were sampled along four 12 m sampling lines.Needle total N andN concentrations were analyzed.We observed needle N andN concentrations increased significantly over time afterN application,up to 31 and0.42%,respectively.There was no correlation between needle N concentration andN/N ratio(R2=0.40,n=5,P=0.156),while excess needle N concentration and excess needleN/N ratio were positively correlated across different time intervals(R~2=0.89,n=4,P\0.05),but deceased with time interval lengthening.NeedleN/N ratio increased with time,but it was not correlated with distance.NeedleN/N ratio was negative with distance before and 6th day and 30th day,positive with distance at 2nd day,but the trend was considerably weaker,their slop were close to zero.These results demonstrated that EM networks were ubiquitous and uniformly distributed in the Mongolian scotch pine plantation and a random network.We found N transfer efficiency was very high,absorbed N by EM network was transferred as wide as possible,we observed N uptake of plant had strong bias forN andN,namely N fractionation.Understanding the structure and function of EM networks in ecosystems may lead to a deeper understanding of ecological stability and evolution,and thus provide new theoretical approaches to improve conservation practices for the management of the Earth’s ecosystems.
文摘胸部疾病高发,且有些疾病种类的癌症转变率很高,因此基于卷积神经网络的胸部X光图像疾病自动检测分类方法是计算机辅助诊断的研究热点之一.然而,目前的自动分类方法仍面临胸部病灶的X光图像特异性特征表达不充分、不同胸部疾病发病率不平衡、卷积神经网络参数量过大等问题.针对上述问题,提出了一种端到端的基于八度卷积的ResNet(octave convolution based residual network,OC-ResNet)结构.首先,利用八度卷积改进ResNet中的普通卷积,将高低频特征分离,增强对高频信息的提取,以更好地表达胸部病灶的特异性特征,降低模型计算复杂度.其次,利用渐进式迁移学习,将OC-ResNet在ImageNet数据集进行预训练,获得网络的初始参数,然后固定网络浅层参数,在ChestX-Ray14数据集上微调网络深层参数.最后,为改善样本不平衡问题,网络训练时,采用了焦点损失函数,增加样本数较少类别的权重.在ChestX-Ray14数据集上的实验结果表明,OC-ResNet对14种胸部疾病分类的平均AUC值达到0.856,与目前先进的深度学习方法相比,其中13种疾病分类的AUC值达到最优,同时,计算复杂度相比基础网络降低了44.77%.
