Sentiment Analysis Of Microblog Incorporating Dynamic Character And Word Features

As the world’s largest Chinese social platform,Microblog has become one of the most important channels for the public to express their emotions,opinions and attitudes.By analyzing the sentiment of microblogs and mining valuable information,we can obtain information about people’s preferences for specific products,and their concerns and emotional changes regarding government policies or social phenomena.This is important for enterprise product marketing,online opinion monitoring and analysis,national and social security,etc.However,microblog sentiment texts have a large number of “different meanings for the same word”;they are more colloquial and contain many Internet buzzwords with a higher degree of irregularity;and they are shorter and contain limited information,which requires a higher level of ability in semantic understanding within the model.All these present considerable challenges in microblog sentiment analysis.Therefore,this dissertation focuses on the improvement of microblog sentiment analysis effect,and designs and implements two more efficient microblog sentiment analysis models.The main research and contributions of this dissertation include:1.In response to the shortcomings of existing methods and the characteristics of microblog speech,the dual-channel microblog sentiment analysis model ALBERT-CCMHSAG with joint character and word features is proposed.Firstly,character embedding and word embedding are combined to achieve full extraction of microblog text features.Secondly,the cross-channel feature fusion operation is combined with multi-head self-attention mechanism to replace the pooling operation of Text CNN,so that the model can acquire both local key features and global semantic information of the text.Finally,the gating mechanism is used to establish the interaction between character-level and word-level semantic information and adaptively acquire the more critical information in both,which further improves the model’s microblog sentiment analysis capability.Experiments show that the sentiment analysis of the model outperforms existing mainstream deep learning models.2.In response to the shortcomings of ALBERT-CCMHSAG model,a microblog sentiment analysis model DMM-CNN based on multi-head self-attention pooling and multigranularity feature interaction fusion is proposed.Firstly,the dual-stream dynamic encoding of characters and words is proposed to enable the model to acquire more comprehensive and accurate text semantic features.Secondly,the multi-head self-attention pooling method is introduced to enable the model to obtain key semantic information and achieve feature dimensionality reduction at the same time.Finally,the multi-granularity feature cross-fusion mechanism is proposed to enable the model to output higher quality joint character and word feature information.Experiments show that the model further improves microblog sentiment analysis compared to the ALBERT-CCMHSAG model.