框架语义角色标注(Frame Semantic Role Labeling,FSRL)是基于FrameNet标注体系的语义分析任务。语义角色标注通常对句法有很强的依赖性,目前的语义角色标注模型大多基于双向长短时记忆网络Bi-LSTM,虽然可以获取句子中的长距离依赖信息...框架语义角色标注(Frame Semantic Role Labeling,FSRL)是基于FrameNet标注体系的语义分析任务。语义角色标注通常对句法有很强的依赖性,目前的语义角色标注模型大多基于双向长短时记忆网络Bi-LSTM,虽然可以获取句子中的长距离依赖信息,但无法很好地获取句子中的句法信息。因此,引入Self-Attention机制来捕获句子中每个词的句法信息。实验结果表明,该模型在CFN(Chinese FrameNet,汉语框架网)数据集上的F_(1)值得到了提升,证明了融入self-attention机制可以改进汉语框架语义角色标注模型的性能。展开更多
在室内实例物体目标检测中,传统深度学习需要大量人工标注的训练样本进行网络训练,费时费力,为此提出并实现了一种采用奇异值分解(Singular Value Decomposition,SVD)和协同训练的半监督实例级目标检测网络SVD-RCNN。挑选关键样本进行...在室内实例物体目标检测中,传统深度学习需要大量人工标注的训练样本进行网络训练,费时费力,为此提出并实现了一种采用奇异值分解(Singular Value Decomposition,SVD)和协同训练的半监督实例级目标检测网络SVD-RCNN。挑选关键样本进行人工标注并预训练SVD-RCNN,以确保其获取更多先验知识,采用基于SVD的收敛-分解-微调策略,在SVD-RCNN中得到两个较强独立性的检测器以满足协同训练的要求,最后提出一种自适应的自标注策略,获得高质量的自标注及检测结果。在多个室内实例数据集上对该方法进行测试,在GMU数据集上只需人工标注199个样本,均值平均精度(mean Average Precision,mAP)达到了79.3%,相较于需标注3 851个样本的全监督Faster RCNN的81.3%mAP仅下降了2%。消融实验及系列实验证明了本文方法的有效性和普适性,本文提出的方法仅需人工标注5%的训练数据,即可达到与全监督学习相当的实例级目标检测精度,有利于智能机器人高效识别不同实例物体的实际应用。展开更多
Site-directed spin-labeling with continuous wave electron paramagnetic resonance spectroscopy was used to monitor autoproteolysis of HIV-1 protease, an enzyme essential for viral maturation. Two protein constructs wer...Site-directed spin-labeling with continuous wave electron paramagnetic resonance spectroscopy was used to monitor autoproteolysis of HIV-1 protease, an enzyme essential for viral maturation. Two protein constructs were examined, namely subtype F and the circulating recombinant form CRF01_A/E. As the protease undergoes self-cleavage, protein unfolds and small peptide fragments containing the spin label are generated, which collectively give rise to a sharp spectral component that is easily discernable in the high-field resonance line in the EPR spectrum. By monitoring the intensity of this spectral component over time, the autoproteolytic stability of each construct was characterized under various conditions. Data were collected for samples stored at 4 °C, 25 °C, and 37 °C, and on a subtype F HIV-1 protease sample stored at 25 °C and containing the FDA-approved protease inhibitor Tipranavir. As expected, the rate of autoproteolysis decreased as the storage temperature was lowered. Minimal autoproteolysis was seen for the sample that contained Tipranavir, providing direction for future spectroscopic studies of active protease samples. When compared to standard methods of monitoring protein degradation such as gel electrophoresis or chromatographic analyses, spin-labeling with CW EPR offers a facile, real-time, non-consuming way to monitor autoproteolysis or protein degradation. Additionally, mass spectrometry studies revealed that the N-termini of both constructs are sensitive to degradation and that the sites of specific autoproteolysis vary.展开更多
文摘框架语义角色标注(Frame Semantic Role Labeling,FSRL)是基于FrameNet标注体系的语义分析任务。语义角色标注通常对句法有很强的依赖性,目前的语义角色标注模型大多基于双向长短时记忆网络Bi-LSTM,虽然可以获取句子中的长距离依赖信息,但无法很好地获取句子中的句法信息。因此,引入Self-Attention机制来捕获句子中每个词的句法信息。实验结果表明,该模型在CFN(Chinese FrameNet,汉语框架网)数据集上的F_(1)值得到了提升,证明了融入self-attention机制可以改进汉语框架语义角色标注模型的性能。
文摘在室内实例物体目标检测中,传统深度学习需要大量人工标注的训练样本进行网络训练,费时费力,为此提出并实现了一种采用奇异值分解(Singular Value Decomposition,SVD)和协同训练的半监督实例级目标检测网络SVD-RCNN。挑选关键样本进行人工标注并预训练SVD-RCNN,以确保其获取更多先验知识,采用基于SVD的收敛-分解-微调策略,在SVD-RCNN中得到两个较强独立性的检测器以满足协同训练的要求,最后提出一种自适应的自标注策略,获得高质量的自标注及检测结果。在多个室内实例数据集上对该方法进行测试,在GMU数据集上只需人工标注199个样本,均值平均精度(mean Average Precision,mAP)达到了79.3%,相较于需标注3 851个样本的全监督Faster RCNN的81.3%mAP仅下降了2%。消融实验及系列实验证明了本文方法的有效性和普适性,本文提出的方法仅需人工标注5%的训练数据,即可达到与全监督学习相当的实例级目标检测精度,有利于智能机器人高效识别不同实例物体的实际应用。
文摘Site-directed spin-labeling with continuous wave electron paramagnetic resonance spectroscopy was used to monitor autoproteolysis of HIV-1 protease, an enzyme essential for viral maturation. Two protein constructs were examined, namely subtype F and the circulating recombinant form CRF01_A/E. As the protease undergoes self-cleavage, protein unfolds and small peptide fragments containing the spin label are generated, which collectively give rise to a sharp spectral component that is easily discernable in the high-field resonance line in the EPR spectrum. By monitoring the intensity of this spectral component over time, the autoproteolytic stability of each construct was characterized under various conditions. Data were collected for samples stored at 4 °C, 25 °C, and 37 °C, and on a subtype F HIV-1 protease sample stored at 25 °C and containing the FDA-approved protease inhibitor Tipranavir. As expected, the rate of autoproteolysis decreased as the storage temperature was lowered. Minimal autoproteolysis was seen for the sample that contained Tipranavir, providing direction for future spectroscopic studies of active protease samples. When compared to standard methods of monitoring protein degradation such as gel electrophoresis or chromatographic analyses, spin-labeling with CW EPR offers a facile, real-time, non-consuming way to monitor autoproteolysis or protein degradation. Additionally, mass spectrometry studies revealed that the N-termini of both constructs are sensitive to degradation and that the sites of specific autoproteolysis vary.
