As a kind of Graph Convolutional Neural Network （GCNN）， Topology Optimization based Graph Convolutional Network （TOGCN） model adopts auxiliary information in the network to optimize topological structure of the network， thereby helping to reflect the relational degrees between the nodes. However， TOGCN model only focuses on the association between local nodes， and not enough on the potential global structure information. Fusing global feature information， the model will help to improve performance as well as its robustness in dealing with incomplete information. A Global structure information Enhanced-TOGCN （GE-TOGCN） model was proposed， the attributes of neighboring nodes were utilized to optimize the topological graph， and the class information was regarded as the global structure information to maintain intra-class aggregation and inter-class separation. Firstly， the center vector of each class was calculated by the labeled nodes， then some unlabeled nodes were selected to update these class center vectors. Finally， all the nodes were assigned to the corresponding class according to their similarity to class center vectors， and a semi-supervised loss function was adopted to optimize the class center vector of each class and the final representation vectors of the nodes. On Cora and Citeseer datasets， node classification task and node visualization task were performed by using the obtained node representation vectors with the loss of label information. Experimental results show that compared with Graph Convolutional Network （GCN）， Graph Learning-Convolutional Network （GLCN） and other models， GE-TOGCN has the classification accuracy increased by 1.2-12.0 percentage points on Cora dataset， and the classification accuracy increased by 0.9-9.9 percentage points on Citeseer dataset. In node visualization task， the proposed model has higher degree of intra-class node aggregation and more obvious boundaries between class clusters. In summary， the fusion of class global information can reduce the negative influence of label information loss on learning effects of the model， and the node representations obtained by the proposed model have better performance in downstream tasks.

In order to improve the utilization efficiency of deep information in text context, based on Hierarchical Attention Network (HAN) and Independent Recurrent Neural Network (IndRNN), a regional bullying semantic recognition model called HACBI (HAN_CNN_BiLSTM_IndRNN) was proposed. Firstly, the manually annotated regional bullying texts were mapped into a low-dimensional vector space by means of word embedding technology. Secondly, the local and global semantic information of bullying texts was extracted by using Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), and internal structure information of text was captured by HAN. Finally, in order to avoid the loss of text hierarchy information and solve the gradient disappearance problem, IndRNN was introduced to enhance the description ability of model, which achieved the integration of information flow. Experimental results show that the Accuracy (Acc), Precision (P), Recall (R), F1 (F1-Measure) and AUC (Area Under Curve) values are 99.57%, 98.54%, 99.02%, 98.78% and 99.35% respectively of this model, which indicates that the effectiveness provided by HACBI is significantly improved compared to text classification models such as Support Vector Machine (SVM) and CNN.

To deal with the under-resourced labeled pronunciation data in mispronunciation detection, some other data were used to improve the discriminability of feature in the framework of Tandem system. Taking Chinese learning of English as object, unlabeled data, native Mandarin data and native English data which can be relatively easily accessed were selected as the assisted data. The experiments show that these types of data can effectively improve the performance of system, and the unlabeled data performs the best. And the effect to system performance was discussed with different length of frame context, the shallow and deep neural network typically represented by Multi-Layer Perception (MLP) and Deep Neural Network (DNN), and different structure of Tandem feature. Finally the strategy of merging multiple data streams was used to further improve the system performance, and the best system performance was achieved by combining the DNN based unlabeled data stream and native English stream. Compared with the baseline system, the recognition accuracy is increased by 7.96%, and the diagnostic accuracy of mispronunciation type is increased by 14.71%.