Research On The Stance Detection In Social Network Text Based On Deep Learning

With the rapid development of social network,a large number of text data of social network have been produced.The stance analysis of social network text is to analyze the comments on a topic,whether they agree with the topic or not,and then show that the comment users' stance is support,opposition or neutrality.Stance analysis has an important practical application in the public opinion monitoring of hot events,and has become an important research tool to grasp hot topics,which can provide reliable data support and services for social management.In the traditional stance analysis task,researchers mainly mine the text semantic features by building feature engineering,emotion dictionary,etc.,and classify the stance by machine learning,which requires a lot of manpower in feature selection and design.This paper takes social network text stance analysis as the research content,and proposes a method of social network text stance analysis based on deep learning.In this paper,we use the strategy of "divide and rule" to train the sub vector space of five topics as word vector embedding,and study the influence of different word segmentation tools and different word vector space on the experimental results.This paper proposes a text stance analysis method based on attention convolution neural network,and makes a comparative study of attention pooling strategy,maximum pooling strategy and average pooling strategy.The experimental results show that compared with the traditional maximum pooling strategy and average pooling strategy,the macro average value of the proposed approach is increased by 5.97% and 7.3% respectively,which shows that the approach based on the attention convolution neural network is superior to the traditional pooling strategy.In this paper,the advantages of bi-directional short-term memory network,one-dimensional convolution neural network and attention mechanism are fully analyzed,and a deep fusion model based on attention fusion neural network CNN-ATBi LSTM is proposed,which solves the problem that single convolution neural network cannot obtain global semantic information and the gradient of traditional recurrent neural network disappears.In this model,the local feature and the whole text semantic information are considered,and the topic comment text information is extracted effectively.In the later stage,the attention mechanism is used to distribute the weight of the effective local information of the text,and finally the feature vector of the social network comment text is obtained.The experimental results on the NLPCC-2016 task 4 open data set show that the performance of the model in four of the five sub topics exceeds the optimal evaluation score,and the total macro average value exceeds the optimal value by 1.72%.The results verify the feasibility and effectiveness of the proposed method in social network text stance analysis.