基于生成对抗网络中的循环一致性原则,提出了一个基于循环一致性损失的知识图谱嵌入模型。该模型首先使用ConvE模型利用头实体和关系构造的“图片”对尾实体进行预测,再利用尾实体和关系构造的“逆图片”对头实体进行预测。同时根据循...基于生成对抗网络中的循环一致性原则,提出了一个基于循环一致性损失的知识图谱嵌入模型。该模型首先使用ConvE模型利用头实体和关系构造的“图片”对尾实体进行预测,再利用尾实体和关系构造的“逆图片”对头实体进行预测。同时根据循环一致性原理,构造了ConvE模型的一个新的损失函数,解决了网络的可逆性。在WN18、FB15k以及YAGO3-10三个数据集上设计实验,证明了模型有效地缩短了头实体和原头实体的语义空间距离。Based on the cyclic consistency principle in generative adversarial networks, a knowledge graph embedding model based on cyclic consistency loss is proposed. Firstly, the ConvE model is used to predict the tail entity by using the “picture” constructed by the head entity and the relationship, and then the “inverse picture” constructed by the tail entity and the relationship is used to predict the head entity. According to the principle of cyclic consistency, a new loss function of ConvE model is constructed to solve the reversibility of the network. Experiments are designed on WN18, FB15k and YAGO3-10 data sets, and it is proved that the model can effectively shorten the semantic space distance between the header entity and the original header entity.展开更多
Though numerical wave models have been applied widely to significant wave height prediction,they consume massive computing memory and their accuracy needs to be further improved.In this paper,a two-dimensional(2D)sign...Though numerical wave models have been applied widely to significant wave height prediction,they consume massive computing memory and their accuracy needs to be further improved.In this paper,a two-dimensional(2D)significant wave height(SWH)prediction model is established for the South and East China Seas.The proposed model is trained by Wave Watch III(WW3)reanalysis data based on a convolutional neural network,the bidirectional long short-term memory and the attention mechanism(CNNBiLSTM-Attention).It adopts the convolutional neural network to extract spatial features of original wave height to reduce the redundant information input into the BiLSTM network.Meanwhile,the BiLSTM model is applied to fully extract the features of the associated information of time series data.Besides,the attention mechanism is used to assign probability weight to the output information of the BiLSTM layer units,and finally,a training model is constructed.Up to 24-h prediction experiments are conducted under normal and extreme conditions,respectively.Under the normal wave condition,for 3-,6-,12-and 24-h forecasting,the mean values of the correlation coefficients on the test set are 0.996,0.991,0.980,and 0.945,respectively.The corresponding mean values of the root mean square errors are measured at 0.063 m,0.105 m,0.172 m,and 0.281 m,respectively.Under the typhoon-forced extreme condition,the model based on CNN-BiLSTM-Attention is trained by typhooninduced SWH extracted from the WW3 reanalysis data.For 3-,6-,12-and 24-h forecasting,the mean values of correlation coefficients on the test set are respectively 0.993,0.983,0.958,and 0.921,and the averaged root mean square errors are 0.159 m,0.257 m,0.437 m,and 0.555 m,respectively.The model performs better than that trained by all the WW3 reanalysis data.The result suggests that the proposed algorithm can be applied to the 2D wave forecast with higher accuracy and efficiency.展开更多
文摘基于生成对抗网络中的循环一致性原则,提出了一个基于循环一致性损失的知识图谱嵌入模型。该模型首先使用ConvE模型利用头实体和关系构造的“图片”对尾实体进行预测,再利用尾实体和关系构造的“逆图片”对头实体进行预测。同时根据循环一致性原理,构造了ConvE模型的一个新的损失函数,解决了网络的可逆性。在WN18、FB15k以及YAGO3-10三个数据集上设计实验,证明了模型有效地缩短了头实体和原头实体的语义空间距离。Based on the cyclic consistency principle in generative adversarial networks, a knowledge graph embedding model based on cyclic consistency loss is proposed. Firstly, the ConvE model is used to predict the tail entity by using the “picture” constructed by the head entity and the relationship, and then the “inverse picture” constructed by the tail entity and the relationship is used to predict the head entity. According to the principle of cyclic consistency, a new loss function of ConvE model is constructed to solve the reversibility of the network. Experiments are designed on WN18, FB15k and YAGO3-10 data sets, and it is proved that the model can effectively shorten the semantic space distance between the header entity and the original header entity.
基金This study is supported by the project supported by the Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai)(SML2020SP007)the National Natural Science Foundation of China(Nos.61772280 and 62072249).
文摘Though numerical wave models have been applied widely to significant wave height prediction,they consume massive computing memory and their accuracy needs to be further improved.In this paper,a two-dimensional(2D)significant wave height(SWH)prediction model is established for the South and East China Seas.The proposed model is trained by Wave Watch III(WW3)reanalysis data based on a convolutional neural network,the bidirectional long short-term memory and the attention mechanism(CNNBiLSTM-Attention).It adopts the convolutional neural network to extract spatial features of original wave height to reduce the redundant information input into the BiLSTM network.Meanwhile,the BiLSTM model is applied to fully extract the features of the associated information of time series data.Besides,the attention mechanism is used to assign probability weight to the output information of the BiLSTM layer units,and finally,a training model is constructed.Up to 24-h prediction experiments are conducted under normal and extreme conditions,respectively.Under the normal wave condition,for 3-,6-,12-and 24-h forecasting,the mean values of the correlation coefficients on the test set are 0.996,0.991,0.980,and 0.945,respectively.The corresponding mean values of the root mean square errors are measured at 0.063 m,0.105 m,0.172 m,and 0.281 m,respectively.Under the typhoon-forced extreme condition,the model based on CNN-BiLSTM-Attention is trained by typhooninduced SWH extracted from the WW3 reanalysis data.For 3-,6-,12-and 24-h forecasting,the mean values of correlation coefficients on the test set are respectively 0.993,0.983,0.958,and 0.921,and the averaged root mean square errors are 0.159 m,0.257 m,0.437 m,and 0.555 m,respectively.The model performs better than that trained by all the WW3 reanalysis data.The result suggests that the proposed algorithm can be applied to the 2D wave forecast with higher accuracy and efficiency.
