Research On Emotional Analysis Method For Social Network Text

With the rapid development of Internet technology and social networking platforms,people publish opinions and share information through the Internet,which behavior generates a large amount of text data containing emotional information.These data contain rich social and commercial value,which attracts many scholars to pay attention to text emotional analysis research.However,most of the web texts are short and diverse,which makes the research of emotional analysis more difficult.In addition,with the increasing cultural exchanges,multi-language mixed expressions are also widely used in network platforms such as Zhihu,Micro-blog and Twitter.If such text is translated into the same language for emotional analysis,it may cause problems such as semantic changes and information loss.To this end,this thesis conducts an in-depth study on the emotional analysis methods of sentence-level texts and mixed Chinese and English texts.The main work includes the following two aspects:(1)Proposing a classification method of emotional sentences based on emotional cognition model(OCC model)and Bayesian network.This thesis analyzes the emotional generation process of OCC model from the perspective of cognitive psychology.It extracts a set of emotion evaluation variables,and proposes a emotion evaluation variables assignment algorithm based on semantic relations by syntactic and semantic analysis of the text.This thesis takes the emotion evaluation variable as the node of the Bayesian Network and combines the emoji features,and obtains emotional classification Bayesian network(ECBN)through structure learning and parameter learning.Experiments on the public evaluation data set show that after adding the expression features,the average precision under the loose standard and the strict standard conditions are increased by 5.5% and 4.4%,respectively.The average precision of ECBN under loose standard conditions is better than the best evaluation results,which verifies the validity of ECBN for emotional classification.(2)Proposing a code-switching text emotional analysis model(MF-CSEL,CostSensitive Ensemble Learning based on Multi-Feature fusion).According to the language characteristics of Chinese and English mixed texts,this thesis extracts a variety of text features for model input.Firstly,it obtains a text vector containing word order information by serializing the combined word vector.Secondly,it extracts corresponding text emotion features for different languages.Finally,the term frequency-inverse document frequency(TF-IDF)feature is extracted to represent the keyword weight.The imbalance of various emotional samples often leads to the classification of the model to the category with more samples.For this reason,this thesis proposes a sample space reconstruction algorithm based on semantic similarity,which preserves the diversity as much as possible while balancing the sample data.Then,according to the principle of cost and minimum,a cost-sensitive integration strategy is formulated to fuse the classification results of the base classifiers on each sample space.Experiments on NLPCC 2018 Code-switching text emotional analysis datasets verified the validity of MF-CSEL model.