Research On Sentiment Analysis Of Chinese Short Text Combined With Emoticons

With the rapid development of applications such as social platforms,a huge amount of user social text data has been produced,which provides favorable support for sentiment analysis,but most research on sentiment analysis only focuses on text information,it ignores the significance of emoji.In a specific scene,emoji tend to highlight emotional colors more easily than text.Deep neural network has performed well in tasks such as sentiment analysis,but there are still some challenges and problems.When the text is maliciously disturbed or damaged,the deep neural network will misjudge the classification due to its poor robustness.This thesis conducts research on the above issues,the main work is as follows:(1)A dataset containing a large number of emojis was constructed.Due to the difficulty of finding open source data sets with many emojis.Based on the characteristics of large number of Weibo comments and short text length,this thesis collected and sorted out a large quantity of comments containing emojis,and manually selected them to mark emotional labels to provide data for subsequent research work.(2)A Skip-Gram model based on emoji and a negative sampling method are proposed to embed emoji vectors.The effectiveness of the algorithm is proved by experiments,which provides powerful support for sentiment analysis of weibo text containing emoticons.(3)A bidirectional gated recurrent unit network combined with the self-attention mechanism(SA-Bi GRU)sentiment analysis mode is proposed.The self-attention mechanism can calculate the dependencies between words,emojis,and between words and emojis,the bidirectional gated recurrent unit network can simultaneously capture the semantic dependencies of the forward and backward directions.The experimental results show that the combination of self-attention mechanism and bidirectional gated recurrent unit network can effectively improve the performance of sentiment classification.(4)To solve the problem of bad robustness of deep learning network,and in order to improve the performance of the classifier.In this thesis,we improve the adversarial attack method based on synonym replacement algorithm.At the same time,combined with the characteristics of Weibo text,a text adversarial sample generation algorithm combined with emoji is proposed.On this basis,the method is introduced into adversarial training to improve the robustness and classification performance of the model.