针对词向量语义信息不完整以及文本特征抽取时的一词多义问题,提出基于BERT(Bidirectional Encoder Representation from Transformer)的两次注意力加权算法(TARE)。首先,在词向量编码阶段,通过构建Q、K、V矩阵使用自注意力机制动态编...针对词向量语义信息不完整以及文本特征抽取时的一词多义问题,提出基于BERT(Bidirectional Encoder Representation from Transformer)的两次注意力加权算法(TARE)。首先,在词向量编码阶段,通过构建Q、K、V矩阵使用自注意力机制动态编码算法,为当前词的词向量捕获文本前后词语义信息;其次,在模型输出句子级特征向量后,利用定位信息符提取全连接层对应参数,构建关系注意力矩阵;最后,运用句子级注意力机制算法为每个句子级特征向量添加不同的注意力分数,提高句子级特征的抗噪能力。实验结果表明:在NYT-10m数据集上,与基于对比学习框架的CIL(Contrastive Instance Learning)算法相比,TARE的F1值提升了4.0个百分点,按置信度降序排列后前100、200和300条数据精准率Precision@N的平均值(P@M)提升了11.3个百分点;在NYT-10d数据集上,与基于注意力机制的PCNN-ATT(Piecewise Convolutional Neural Network algorithm based on ATTention mechanism)算法相比,精准率与召回率曲线下的面积(AUC)提升了4.8个百分点,P@M值提升了2.1个百分点。在主流的远程监督关系抽取(DSER)任务中,TARE有效地提升了模型对数据特征的学习能力。展开更多
Long-term changes of phytoplankton community by water sampling method in Xiagu Sea waters of Xiamen,China,were investigated in this study.Species composition of the phytoplankton community in these waters changed grea...Long-term changes of phytoplankton community by water sampling method in Xiagu Sea waters of Xiamen,China,were investigated in this study.Species composition of the phytoplankton community in these waters changed greatly since the 1950s.The numbers of Dinophyta species increased significantly,although Bacillariophyta species are generally dominant.The succession of dominant species in phytoplankton community is obvious： large-size dominant species such as Biddulphia sinensis of the 1950s were gradually replaced by small-size ones such as Cyclotella striata and Nitzschia closterium,and species that still maintain dominant such as Skeletonema costatum are also small ones,leading the whole phytoplankton community of smaller size.Cell density of phytoplankton community increased greatly,among which cell density of the most dominant species Skeletonema costatum have been increasing in exponent function.Margalef index of phytoplankton community decreased,indicating decline of biodiversity of the community,and dominant character of Skeletonema costatum increased.Generally,the structure of the entire phytoplankton community is becoming more and more singular and unstable,which makes the occurrence of red tides more frequent.The succession in the phytoplankton community is related to the long-term changes in marine environment,influenced by human activities and global climate changes,especially the increases of nutrient content.展开更多
文摘针对词向量语义信息不完整以及文本特征抽取时的一词多义问题,提出基于BERT(Bidirectional Encoder Representation from Transformer)的两次注意力加权算法(TARE)。首先,在词向量编码阶段,通过构建Q、K、V矩阵使用自注意力机制动态编码算法,为当前词的词向量捕获文本前后词语义信息;其次,在模型输出句子级特征向量后,利用定位信息符提取全连接层对应参数,构建关系注意力矩阵;最后,运用句子级注意力机制算法为每个句子级特征向量添加不同的注意力分数,提高句子级特征的抗噪能力。实验结果表明:在NYT-10m数据集上,与基于对比学习框架的CIL(Contrastive Instance Learning)算法相比,TARE的F1值提升了4.0个百分点,按置信度降序排列后前100、200和300条数据精准率Precision@N的平均值(P@M)提升了11.3个百分点;在NYT-10d数据集上,与基于注意力机制的PCNN-ATT(Piecewise Convolutional Neural Network algorithm based on ATTention mechanism)算法相比,精准率与召回率曲线下的面积(AUC)提升了4.8个百分点,P@M值提升了2.1个百分点。在主流的远程监督关系抽取(DSER)任务中,TARE有效地提升了模型对数据特征的学习能力。
基金The Scientific Research Foundation of Third Institute of Oceanography,State Oceanic Administration under contract Nos TIO 2007009 and TIO 2009007the River basin-Estuary ecological security assessment and Management strategy under contract No.200805064+4 种基金the Natural Science Foundation of Fujian Province under contract No.2010J01260the "908" Project under contract No.908-02-02-01 special subjectthe Program of Chinese Marine Chemistry Investigation and Research under contract No.908-ZC-I-03the Special Fund of State Oceanic Administration under contract No.908-02-01-02the Major State Basic Research Development Program of China (973 Program) under contract Nos 2010CB428704 and 2005CB422305
文摘Long-term changes of phytoplankton community by water sampling method in Xiagu Sea waters of Xiamen,China,were investigated in this study.Species composition of the phytoplankton community in these waters changed greatly since the 1950s.The numbers of Dinophyta species increased significantly,although Bacillariophyta species are generally dominant.The succession of dominant species in phytoplankton community is obvious： large-size dominant species such as Biddulphia sinensis of the 1950s were gradually replaced by small-size ones such as Cyclotella striata and Nitzschia closterium,and species that still maintain dominant such as Skeletonema costatum are also small ones,leading the whole phytoplankton community of smaller size.Cell density of phytoplankton community increased greatly,among which cell density of the most dominant species Skeletonema costatum have been increasing in exponent function.Margalef index of phytoplankton community decreased,indicating decline of biodiversity of the community,and dominant character of Skeletonema costatum increased.Generally,the structure of the entire phytoplankton community is becoming more and more singular and unstable,which makes the occurrence of red tides more frequent.The succession in the phytoplankton community is related to the long-term changes in marine environment,influenced by human activities and global climate changes,especially the increases of nutrient content.
