Research On Stance Detection Method For Chinese Microblog Posts Within Targets

The continuous growth of social applications such as microblog has led to the emergence of popular information and online commentary,and has a potential impact on public perceptions and social development.From the perspective of public opinion analysis,the monitoring,analysis and management of online speech plays an important role in the fields of national security,public opinion investigation,and business decision-making.It is difficult to analyze and deal with the massive online public opinion.In recent years,researchers in the field of natural language processing have analyzed and guided the trend of network public opinion by studying the effective methods of sentiment analysis.As the target of microblog is updated,and online reviews continue to emerge in new forms and become more diverse,sentiment analysis research faces new challenges.Stance detection has emerged as a result of the fact that opinions in some microblog comments are difficult to be automatically identified due to their close relation to the target information and implicit expressing.Stance detection is an emerging orientation in the field of sentiment analysis.It is devoted to the automatic recognition of the attitude of microblog comments on the support of the target,and is essentially a text classification task.For the problem of difficulty analyzing the stance tendency of new form of microblog comments,difficulty fitting all targets via single analysis model,and the increasing complexity of the analysis model due to the emerge of new targets,the three main research work in this thesis are as follows:1)Stance classification based on deep learning: a deep learning model is proposed,which includes neural network layers such as bi-directional long short-term memory and attention mechanism.The text features are extracted from the training data,and the most noteworthy parts are selected and formed the final representation.Finally,the classification is performed.2)Ensemble learning strategy based on sub-classifier of targets: The classification model is used to train the data under each target in the data set,and then the sub-classifiers are ensembled to process the classification of the full-scale data.This section sets the three ensemble strategies on input,text representation,and output stage for comparison.3)New-target review prediction based on cross-domain transfer learning: based onthe ensemble model,extract the main text features in the original training data,and then use the new-target training data,respectively,with zero-shot learning and learning with small amount of sample,eventually large amount of test data under the target is predicted.The experimental results show that the proposed stance classification model has promising classification performance,and the evaluation metrics on four targets in the NLPCC 2016 stance detection dataset exceed the state-of-the-art metrics.In the comparison result of the ensemble strategy,concatenating the results of the sub-classifiers using ad-hoc strategy during output stage has achieved relatively better classification performance on the full-scale stance detection data.As the new-target data is fitted in the method of text cross-domain transfer learning,the prediction performance obtained by training the model using small amount of new-target training data is higher than the performance predicting using the new-target test data directly in all experimental scenarios.The problems and future work are explained at the end of this thesis.