基于BERT多特征融合的网络舆情情感识别

Sentiment Analysis of Internet Public Opinion Based on BERT Multi-Feature Fusion Model

社交网络文本含有丰富的情感信息,通过社交网络文本情感识别掌握网络舆情动态具有重要的意义。由于文本数据的高维稀疏性,情感分析任务面临着巨大的挑战。为了解决上述问题,提出了一种基于BERT多特征融合的网络舆情情感识别新模型。首先,使用BERT预训练语言模型对输入文本进行编码;然后,根据BERT编码层输出的特点,从三个通道分别对其生成的特征向量进行进一步的处理,形成三个特征向量;最后,对这三个特征向量进行拼接,构建网络舆情情感识别模型。以新冠疫情期间网民的微博评论为数据集验证模型的可行性和优越性,模型的精确率、召回率和F1值分别达到92.7%、93.9%以及93.2%。实验结果表明,基于BERT多特征融合的特征向量包含更加丰富文本的语义信息,能够有效提升网络舆情情感识别的性能。

Social network texts contain abundant emotional information that provide crucial insights into the dynamics of online public opinion via the recognition of inherent sentiments.However,sentiment analysis faces enormous challenges due to the high-dimensional sparsity of text data.Hence,a new sentiment recognition model for online public opinion was proposed based on BERT multi-feature fusion.A model for sentiment analysis from Internet public opinions was constructed in three steps.First,the pre-trained BERT language model was used to encode input text.Second,according to the characteristics of the output of the BERT encoding layer,the vectors generated by the three channels were further processed to form three embeddings.Third,the three embeddings were concatenated to construct the online public opinion sentiment recognition model.The model's feasibility and superiority were verified using netizens'Weibo comments during the COVID-19 pandemic as the dataset.The model's accuracy,recall,and F1-score reached 92.7%,93.9%,and 93.2%,respectively.The experimental results showed that the feature vector based on BERT multi-feature fusion contained richer semantic information from text,which could effectively improve the performance of sentiment recognition.