In the cross-domain sentiment analysis， the labeled samples in the target domain are seriously insufficient， the distributions of features in different domains are very different， and the emotional polarities expressed by features in one domain differ a lot from the emotional polarities in another domain， all of these problems lead to low classification accuracy. To deal with the above problems， an aspect-level cross-domain sentiment analysis method based on capsule network was proposed. Firstly， the feature representations of text were obtained by BERT （Bidirectional Encoder Representation from Transformers） pre-training model. Secondly， for the fine-grained aspect-level sentiment features， Recurrent Neural Network （RNN） was used to fuse the context features and aspect features. Thirdly， capsule network and dynamic routing were used to distinguish overlapping features， and the sentiment classification model was constructed on the basis of capsule network. Finally， a small amount of data in the target domain was used to fine-tune the model to realize cross-domain transfer learning. The optimal F1 score of the proposed method is 95.7% on Chinese dataset and 91.8% on English dataset， which effectively solves the low accuracy problem of insufficient training samples.

Aiming at the problem that the model in the judicial field relation extraction task does not fully understand the context of sentence and has weak recognition ability of overlapping relations， based on Criminal-Efficiently learning an encoder that classi?es token replacements accurately （CriElectra）， an encoding-decoding relationship extraction model was proposed. Firstly， referred to the training method of Chinese Electra， CriElectra was trained on one million criminal dataset. Then， the word vectors of CriElectra were added to Bidirectional Long Short-Term Memory （BiLSTM） model for feature extraction of judicial texts. Finally， the vector clustering was performed to the features through Capsule Network （CapsNet）， so that the relationships between entities were extracted. Experimental results show that on the self-built relationship dataset of intentional injury crime， compared with the pre-trained language model based on Chinese Electra， CriElectra has retraining process on judicial texts to make the learned word vectors contain richer domain information， and the F1-score increased by 1.93 percentage points. Compared with the model based on pooling clustering， CapsNet can effectively prevent the loss of spatial information by vector operation and improve the recognition ability of overlapping relationships， which increases the F1-score by 3.53 percentage points.

Public opinions on major social security emergencies in the era of big data are mainly spread through the media. Most of the existing researches fail to consider the special group — news media and the influence of news media in a certain kind of specific events. In order to study the above problems， a method to evaluate the influence by integrating the network structure and behavioral relationship between users was proposed， and the Xinjiang and Paris violent and terrorist events were taken as examples to calculate the international influence of news media of different countries on such events on the Twitter platform. This evaluation method can better obtain the influence of various news media at the event level. By calculating the influence of news media in the violent and terrorist events in Xinjiang and Paris， the experimental results show that there are differences in the influence of news media of different countries in Xinjiang and Paris violent and terrorist events， which indicates that these two events of the same type have different influence scopes， and also reflects the differences of political positions of different countries.