大规模多标签文本分类(XMTC)是从一个庞大且复杂的标签集合中查找与文本样本最相关标签的一项具有挑战性的任务。目前,基于Transformer模型的深度学习方法在XMTC上取得了巨大的成功。然而,现有方法都没能充分利用Transformer模型的优势...大规模多标签文本分类(XMTC)是从一个庞大且复杂的标签集合中查找与文本样本最相关标签的一项具有挑战性的任务。目前,基于Transformer模型的深度学习方法在XMTC上取得了巨大的成功。然而,现有方法都没能充分利用Transformer模型的优势,忽略了文本不同粒度下细微的局部语义信息,同时标签与文本之间的潜在关联尚未得到稳健的建立与利用。对此,提出了一种基于语义特征与关联注意力的大规模多标签文本分类模型SemFA(An Extreme Multi-Label Text Classification Model Based on Semantic Features and Association-Attention)。在SemFA中,首先拼接多层编码器顶层输出作为全局特征。其次,结合卷积神经网络从多层编码器浅层向量中获取局部特征。综合丰富的全局信息和不同粒度下细微的局部信息获得更丰富、更准确的语义特征。最后,通过关联注意力机制建立标签特征与文本特征之间的潜在关联,引入关联损失作为潜在关联不断优化模型。在Eurlex-4K和Wiki10-31K两个公开数据集上的实验结果表明,SemFA优于大多数现有的XMTC模型,能有效地融合语义特征与关联注意力,提升整体的分类性能。展开更多
Presents a new parallel image matching algorithm based on the concept of entropy feature vector and suitable to SIMD computer, which, in comparison with other algorithms, has the following advantages:(1)The spatial in...Presents a new parallel image matching algorithm based on the concept of entropy feature vector and suitable to SIMD computer, which, in comparison with other algorithms, has the following advantages:(1)The spatial information of an image is appropriately introduced into the definition of image entropy. (2) A large number of multiplication operations are eliminated, thus the algorithm is sped up. (3) The shortcoming of having to do global calculation in the first instance is overcome, and concludes the algorithm has very good locality and is suitable for parallel processing.展开更多
The traditional network simulator has function and performance limitation when simulating Internet worms,so we designed the grid-based Internet worm behavior simulator (IWBS Grid).IWBS Grid makes use of the real Inter...The traditional network simulator has function and performance limitation when simulating Internet worms,so we designed the grid-based Internet worm behavior simulator (IWBS Grid).IWBS Grid makes use of the real Internet topology,link and routing information,and simulates the worm behavior at the packet event-driven level;and proposes a high-performance Internet worms behavior simulation platform by right of the grid computing capability,resource and task management,and so on.The experimental results show that IWBS grid surpasses the traditional simulator in simulating capability,and the technology to track the worm propagation in packet level can propose the valuable information for the further study on worms.展开更多
文摘大规模多标签文本分类(XMTC)是从一个庞大且复杂的标签集合中查找与文本样本最相关标签的一项具有挑战性的任务。目前,基于Transformer模型的深度学习方法在XMTC上取得了巨大的成功。然而,现有方法都没能充分利用Transformer模型的优势,忽略了文本不同粒度下细微的局部语义信息,同时标签与文本之间的潜在关联尚未得到稳健的建立与利用。对此,提出了一种基于语义特征与关联注意力的大规模多标签文本分类模型SemFA(An Extreme Multi-Label Text Classification Model Based on Semantic Features and Association-Attention)。在SemFA中,首先拼接多层编码器顶层输出作为全局特征。其次,结合卷积神经网络从多层编码器浅层向量中获取局部特征。综合丰富的全局信息和不同粒度下细微的局部信息获得更丰富、更准确的语义特征。最后,通过关联注意力机制建立标签特征与文本特征之间的潜在关联,引入关联损失作为潜在关联不断优化模型。在Eurlex-4K和Wiki10-31K两个公开数据集上的实验结果表明,SemFA优于大多数现有的XMTC模型,能有效地融合语义特征与关联注意力,提升整体的分类性能。
文摘Presents a new parallel image matching algorithm based on the concept of entropy feature vector and suitable to SIMD computer, which, in comparison with other algorithms, has the following advantages:(1)The spatial information of an image is appropriately introduced into the definition of image entropy. (2) A large number of multiplication operations are eliminated, thus the algorithm is sped up. (3) The shortcoming of having to do global calculation in the first instance is overcome, and concludes the algorithm has very good locality and is suitable for parallel processing.
基金Sponsored by the National High Technology Research and Development Program of China (Grant No. 2007AA010503)the Science and Technology Development Program of Weihai (Grant No. 2007-96)the Science Foundation of HIT at Weihai (Grant No. HITWH 200702)
文摘The traditional network simulator has function and performance limitation when simulating Internet worms,so we designed the grid-based Internet worm behavior simulator (IWBS Grid).IWBS Grid makes use of the real Internet topology,link and routing information,and simulates the worm behavior at the packet event-driven level;and proposes a high-performance Internet worms behavior simulation platform by right of the grid computing capability,resource and task management,and so on.The experimental results show that IWBS grid surpasses the traditional simulator in simulating capability,and the technology to track the worm propagation in packet level can propose the valuable information for the further study on worms.
