Analyzing Public Opinion Bias through Social Media Using a Hybrid RoBERTa-BiGRU-DPCNN Sentiment Analysis Framework

Abstract

During the development of public opinion, effective guidance through social media is conducive to the formation of a positive public opinion bias, while incorrect guidance may lead to a further loss of control over public opinion. This paper designed a sentiment analysis algorithm called robustly optimized bidirectional encoder representations from transformers pretraining approach-bidirectional gated recurrent unit-deep pyramid convolutional neural network (RoBERTa-BiGRU-DPCNN) to analyze the sentiment categories of texts. The algorithm first used RoBERTa-wwm-ext to obtain the vector representation of a text, then employed BiGRU and DPCNN to extract features, and finally classified sentiment through a softmax layer. Then, the guiding effect of social media was analyzed based on changes in sentiment categories in public opinion events. The sentiment classification performance of the algorithm was evaluated using the SMP2020 dataset, which contains 22,019 samples in the training set and 4,010 in the test set. The Micro F1 value was used to measure the classification performance. It was found that the RoBERTa-BiGRU-DPCNN algorithm achieved a micro F1 value of 73.58%, outperforming algorithms such as TextCNN. The algorithm was applied to analyze two public opinion events, and it was found that the social media guidance effect was poor and failed to form a positive public opinion bias. These results verify the effectiveness of the proposed algorithm in reflecting the guiding effect of social media indirectly.