Research On Emotion And Stance Detection Of Weibo Text

With the rapid development of the Internet,the opinion of Weibo can represent the public opinion of the whole society in some way,while users and information of social network has shown explosive growth,so how to mine and analyze these big data has become the key issue.Automatic web opinion analysis with machine learning has achieved more attention from many companies and researchers.The overall public opinion of the Weibo consists of stance of each text,the aim of our work is to detect the emotion and stance of each text which is helpful for monitoring public opinion timely.Emotion detection of Weibo text is to predict many types of emotion such as happiness,sadness expressed by the user.We firstly utilize Chinese sentiment lexicons,English sentiment lexicons,Chinese-English dictionary to construct a Chinese-English sentiment dictionary,then we can translate the English sentiment words in the Weibo into Chinese sentiment words.Then we combine LSTM and CNN to capture the crucial emotional parts of the sentence.Unlike most work who train independent model for each emotion,we design a unified architecture for training all the emotion label in the same time,we utilize self-attention mechanism to construct relation between each emotion representation.And we utilize the dynamic loss function to handle the data-imbalance problem.Experimental results on the code-switching emotion detection dataset shows our method achieve better performance compared with some other deep learning method and ensemble method,our method is also parameter efficient and improve the ability in application.Single-Target stance detection aims to classify the author of the text is whether in favor of,against or neutral towards a given target.Because of language characteristics of each target,most method train separate model for each target.However the neural model can't generalize well because of less labelled training data for each target.We introduce pretrained language model BERT and sentiment classification dataset which is more easily to obtain.We build a share architecture across targets by share representation and private representation.Experimental results on target stance detection dataset demonstrates the better performance of our method.And further ablation study shows the effectiveness of proposed method.Multi-Target stance detection aims to jointly determine the stances expressed in text towards multiple correlated targets.We propose the convolutional attentive RNN model(CARNN)and adopt a multi-task learning method.The model utilizes convolutional attention to focus on important sentiment and semantic parts with regard to given target.Specifically,we firstly pretrain a general CARNN model on different targets,then finetune a joint network for each target-pair to improve adaption.Experimental result shows that our proposed method achieves the state-of-the-art performance on a benchmark dataset.Further analysis demonstrates the attention mechanism can efficiently capture crucial information,multi-task learning method can advantage of common linguistic features among different targets.