基于图卷积网络的无监督跨模态哈希检索算法

针对无监督跨模态检索任务中不能充分利用单个模态内的语义关联信息的问题,提出了一种基于图卷积网络的无监督跨模态哈希检索方法。通过图像和文本编码器分别获得两个模态的特征,输入到图卷积网络中挖掘单个模态的内部语义信息,将结果通过哈希编码层进行二值化操作后,与模态间的深度语义关联相似度矩阵进行对比计算损失,不断重构优化生成的二进制编码,直到生成样本对应的健壮哈希表达。实验结果表明,与经典的浅层方法和深度学习方法对比,该方法在多个数据集上的跨模态检索准确率均有明显提升。证明通过图卷积网络能够进一步挖掘模态内的语义信息,所提模型具有更高的准确性和鲁棒性。     

Size effect of layer thickness on stress fields due to interface core-spreading dislocation arrays in multilayers

The elastic fields due to interface dislocation arrays with spreading core in multilayers are derived analytically by means of the superposition of Green function of individual compact dislocations. The numerical results for Cu-Nb multilayers with dislocation arrays demonstrate that:(1) There exists a critical layer thickness(CLT) in describing interface shear stress. The maximum interface shear stress(MISS) increases with the layer thickness as it exceeds CLT; while the MISS decreases rapidly with the layer thickness as it is thinner than CLT.(2) Both the density and core width of interface dislocations have salient effect on the stress fields. The overall stress field is confined by high density of dislocations, and the stress singularity is released by core spreading.(3) A new dislocation prefers to glide in the middle region between the two adjacent interface dislocations by means of Peach-Kohler(P-K) force acting on the dislocation and elastic energy of dislocations. These findings would provide foundations to deepen the understanding of microscopic plastic deformation mechanisms and their related macroscopic mechanical properties of metallic anisotropic multilayers.     

一种基于用户兴趣的微博实体链接方法*

本文针对微博内容较短、歧义较大的问题,利用概率主题模型对用户的兴趣进行建模,提出了一种基于用户兴趣的微博实体链接方法。具体地,本文首先利用现有的主题模型从知识库的大量数据中训练实体与上下文词汇的语义关联,然后提出用户兴趣主题模型来建模用户对实体的兴趣以及微博的语义,并完成实体链接的任务。此外,本文在真实数据集上进行了大量实验和分析,取得了87.6%的实体链接准确率,实验结果表明,与现有方法相比,该方法通过用户兴趣的建模更好地刻画了微博的语义,因而也取得了更高的实体链接准确率。