Font Size: a A A

Research On Positive And Negative News Recognition Algorithm Based On Pre Training

Posted on:2022-09-16Degree:MasterType:Thesis
Country:ChinaCandidate:X P MengFull Text:PDF
GTID:2518306341453924Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
With the rapid economic development,our country has become more and more closely connected with the international community.The domestic cultural industry has been fully developed,and a large number of highly active online news media have been produced.The amount of online news text data has also been increased year by year,and gradually showing a trend of subjective diversification,which makes the sentiment analysis of news texts,especially the identification of positive and negative news more and more popular.In order to help people better understand and analyze news texts,this article uses the latest technology of natural language processing and proposes a new model for analyzing positive and negative sentiment tendencies of news texts.The main research contents are as follows:(1)Deeply understand the process and methods of text information processing,study the traditional text information feature construction technology,analyze the characteristics and shortcomings of the classic model of text sentiment classification-the long and short-term memory neural network model(LSTM),and deeply analyzes the necessity and advancement of pre-training model as a feature extraction technology.(2)Based on the use of the pre-training model to extract text feature information,this paper proposes a new feature construction method based on the characteristics of the text data set:put the output of the BERT model into the two-way GRU,and further the contextual semantics Extract and use the output of the bidirectional GRU as the first part of the feature vector,use the CLS bits of the last four-layer encoder of the pre-training model as the semantic representation of the entire text,and concatenate the four CLS bits to form the second part of the feature vector.The feature vectors of the two parts are spliced together as the final feature vector.(3)Three sets of experiments were set up,and three models were used for training,namely:traditional feature construction+LSTM classic model,standard pre-training model,and model built using new feature construction methods.After a reasonable control experiment,it was found that the accuracy of the three models,the Macro-F1 value,etc.,have been significantly improved,which proves the effectiveness of strengthening the feature construction of the new model.The feature of this study is that it strengthens the feature extraction of the pre-training model,which can effectively improve the classification effect of the news text sentiment classification model.At the same time,it also found that the method is insufficient for long text processing and other shortcomings,which provides a powerful reference for subsequent research.
Keywords/Search Tags:Sentiment analysis, Deep learning, Positive and negative news, Pre training model, Characteristic structure
PDF Full Text Request
Related items