Sentiment Analysis Of Social Short Texts Based On Deep Learning

With the rapid development of the Internet,various social platforms provide Internet users with channels for expressing personal preference views on a certain field or topic.These social platforms have certain restrictions on the length of the text,which generated a lot of emotion short text message.Through the collection,processing,and analysis of these short texts,the emotional polarity can be predicted and used to guide public opinion analysis and business decision-making.Therefore,sentiment analysis of these short social texts has special research value.The research in this thesis is based on deep learning-based social short text sentiment analysis method.The main work includes the following aspects:(1)The thesis studies the contribution of emotional part of speech to the task of text sentiment classification.The part of speech contribution factors are introduced and the contribution coefficients are determined through experiments,and then based on this,an improved TF-IDF algorithm TF-IDF-EW for sentiment classification tasks is proposed.The algorithm can integrate the advantages of part-of-speech contribution and traditional TF-IDF,and calculate the contribution weight of the vocabulary in the text.And on the basis of this algorithm,the Word2 Vec word vector is weighted to form a short text vector representation method for sentiment analysis.(2)Aiming at the shortcomings of short text vector representation methods and with the help of the advantages of the Word2 Vec model,Combined with(1),this thesis proposes a BiLSTM sentiment classification model based on improved TF-IDF weighted Word2 Vec word vector: WTE-BiLSTM-ATT.This model combines the Word2 Vec model with the improved algorithm TF-IDF-EW.Then the context information is captured through the BiLSTM layer,and the attention mechanism is introduced to obtain emotional features.The results of comparative experiments show that the method has improved the effect of emotional binary classification,and also verifies the effectiveness of the model.(3)Due to the fact that social short texts are rich in network words,and it has the characteristics of colloquial text and strong emotional tendency.Its requirements for fine-grained emotion classification are more complex and performance requirements are higher.For this,BERT is introduced to express text.First,an emotional language set combining the shallow features of social short texts is constructed,and then combined with the deep features of syntax and semantics,a dual-channel sentiment classification model based on BERT(DC-BERT-ATT)is proposed.The model includes an input layer,a feature extraction layer,a feature fusion layer,and an output layer.Structurally,the model consists of two information processing channels: one side is the BERT-BiGRU-ATT channel,and the other side is the BERT-FT-CNN channel through the BERT fine-tuning output layer.The two channels capture deep and shallow features respectively,and get emotional features through concatenating in feature fusion layer.The model finds the optimal value through parameter experiments,effectively combining the emotional language set of shallow features with the syntactic and semantic information of deep features.The model comparison experiment achieves the best results in the fine-grained multi-classification experiment,which verifies the effectiveness of the model.