Stance Detection Method Of Chinese Micro-blog Based On Deep Learning

In the past two decades,the world has entered the Internet era.People 's online social activities and offline life culture are completely intertwined.The way people accept information has changed from passive to active.Everyone has become a digger and disseminator of information.Various social media operations are born.Chinese Micro-Blog,as the mainstream domestic social media network,has accumulated hundreds of millions of users.Users can express their opinions on the newly released policies,industrial products,current affairs hot spots,etc.on the network.Stance Detection has become an important issue in the analysis of online public opinion.Chinese Micro-Blog text has the characteristics of networking,diversified noise,colloquialization,and diversified relationships,which greatly reduces the accuracy of position detection.When performing stance detection on Chinese Micro-Blog text,it is often judged according to some dependent word groups.Models based on convolutional neural networks perform convolution operations on word sequences to perceive multiple word features as continuous text for representation,but they cannot use the dependency relationships between words to guide position classification.Aiming at the language characteristics of Chinese language environment and Chinese Micro-Blog text,this paper proposes a position detection method based on deep learning,which not only combines Chinese language characteristics,but also effectively captures the dependency characteristics between words at different distances in text.The main works of the paper are as follows:(1)This article regards the words in the text and the dependencies between them as graph structure data,and proposes a method for constructing a Chinese Micro-Blog text graph structure.First,based on the CRF algorithm,construct a syntactic analyzer,and use the Chinese Micro-Blog syntactic relationship tree library for training.Then use the trained analyzer to construct the position text dependency syntax sequence,store it in the Trie tree and use the pruning strategy for pruning to obtain a graphical representation of the text.Finally,use the pointwise mutual information between the nodes in the figure to delete the weaker word dependency.(2)A network model for Chinese Micro-Blog stance detection based on deep learning is proposed.The model includes long short-term memory network layer,graphconvolution network layer,target topic related attention layer and other structures.In the graph convolution network layer,the text graph structure information is used to guide the network to convolve,and the interdependence between words is strengthened by pointwise mutual information to obtain a richer text feature representation.(3)In order to solve the problem of too few data sets for the current stance detection task,the classification model has an occasional effect in classification.A new position detection data set was established,and for each piece of data,the different positions of the target topic were manually marked according to user text.The new target topic set contains a total of two topics with a total of 3000 Chinese Micro-Blog texts.In this paper,experiments are conducted on the Chinese Micro-Blog stance detection task dataset published by the Natural Language Processing and Chinese Computing(NLPCC)and the newly constructed topic stance dataset.The experimental results show that the comparison between the deep learning model proposed in this paper and the existing research methods in terms of position detection has been improved.