针对基础深度学习模型特征提取能力不足,循环网络训练效率低等问题,将高校社交网络平台评论文本数据作为研究对象,提出了基于多尺度语义协同网络的高校网络舆论情感分类模型。预训练模型ALBERT(A Lite BERT)通过结合当前词的具体上下文...针对基础深度学习模型特征提取能力不足,循环网络训练效率低等问题,将高校社交网络平台评论文本数据作为研究对象,提出了基于多尺度语义协同网络的高校网络舆论情感分类模型。预训练模型ALBERT(A Lite BERT)通过结合当前词的具体上下文动态调整向量表示,提升词向量语义表征质量。多尺度语义协同网络捕捉评论文本不同尺度下的多通道融合情感特征,软注意力机制计算每个特征对情感分类结果的影响权重大小,加权求和后得到情感分类特征表示,线性层输出分布概率并得到具体情感倾向。在真实高校图书馆社交网络平台用户评论数据集进行实验,结果表明该模型F1分数达到了97.46%,优于近期表现优秀的实验对比模型,且通过消融实验证明了各个功能模块的有效性。展开更多
针对传统点云滤波算法参数多、端到端的深度学习点云滤波算法计算成本高等问题,设计了一种基于多尺度语义分割网络的点云滤波算法(multi-scale semantic segmenta⁃tion network for point cloud filtering,MSSF)。首先,对点云进行栅格...针对传统点云滤波算法参数多、端到端的深度学习点云滤波算法计算成本高等问题,设计了一种基于多尺度语义分割网络的点云滤波算法(multi-scale semantic segmenta⁃tion network for point cloud filtering,MSSF)。首先,对点云进行栅格化并提取高程、强度等特征,得到多通道特征图;然后,基于多尺度跨层连接模块构建语义分割网络,以特征图为输入提取地面像素,映射到三维点云获取初始高程基准点,经过插值拟合得到地面高程基准面;最后,设置点到该基准面的距离阈值,实现点云滤波。实验结果表明:该算法减少了参数设置,获得了更高的精度,能实现城市区域的Li⁃DAR点云稳健滤波。展开更多
Global semantic structures of two large semantic networks, HowNet and WordNet, are analyzed. It is found that they are both complex networks with features of small-world and scale-free, but with special properties. Ex...Global semantic structures of two large semantic networks, HowNet and WordNet, are analyzed. It is found that they are both complex networks with features of small-world and scale-free, but with special properties. Exponents of power law degree distribution of these two networks are between 1.0 and 2. 0, different from most scale-free networks which have exponents near 3.0. Coefficients of degree correlation are lower than 0, similar to biological networks. The BA （Barabasi-Albert） model and other similar models cannot explain their dynamics. Relations between clustering coefficient and node degree obey scaling law, which suggests that there exist self-similar hierarchical structures in networks. The results suggest that structures of semantic networks are influenced by the ways we learn semantic knowledge such as aggregation and metaphor.展开更多
文摘针对基础深度学习模型特征提取能力不足,循环网络训练效率低等问题,将高校社交网络平台评论文本数据作为研究对象,提出了基于多尺度语义协同网络的高校网络舆论情感分类模型。预训练模型ALBERT(A Lite BERT)通过结合当前词的具体上下文动态调整向量表示,提升词向量语义表征质量。多尺度语义协同网络捕捉评论文本不同尺度下的多通道融合情感特征,软注意力机制计算每个特征对情感分类结果的影响权重大小,加权求和后得到情感分类特征表示,线性层输出分布概率并得到具体情感倾向。在真实高校图书馆社交网络平台用户评论数据集进行实验,结果表明该模型F1分数达到了97.46%,优于近期表现优秀的实验对比模型,且通过消融实验证明了各个功能模块的有效性。
文摘针对传统点云滤波算法参数多、端到端的深度学习点云滤波算法计算成本高等问题,设计了一种基于多尺度语义分割网络的点云滤波算法(multi-scale semantic segmenta⁃tion network for point cloud filtering,MSSF)。首先,对点云进行栅格化并提取高程、强度等特征,得到多通道特征图;然后,基于多尺度跨层连接模块构建语义分割网络,以特征图为输入提取地面像素,映射到三维点云获取初始高程基准点,经过插值拟合得到地面高程基准面;最后,设置点到该基准面的距离阈值,实现点云滤波。实验结果表明:该算法减少了参数设置,获得了更高的精度,能实现城市区域的Li⁃DAR点云稳健滤波。
基金The National Natural Science Foundation of China(No.60275016).
文摘Global semantic structures of two large semantic networks, HowNet and WordNet, are analyzed. It is found that they are both complex networks with features of small-world and scale-free, but with special properties. Exponents of power law degree distribution of these two networks are between 1.0 and 2. 0, different from most scale-free networks which have exponents near 3.0. Coefficients of degree correlation are lower than 0, similar to biological networks. The BA （Barabasi-Albert） model and other similar models cannot explain their dynamics. Relations between clustering coefficient and node degree obey scaling law, which suggests that there exist self-similar hierarchical structures in networks. The results suggest that structures of semantic networks are influenced by the ways we learn semantic knowledge such as aggregation and metaphor.
