基于弹幕与评论情感倾向的食品安全舆情预警研究

Public Opinion Early Warning on Food Security Based on Sentiment of Video Bullet Screen and Comments

[目的/意义]采用弹幕与评论的情感倾向建立负面舆情自动预警系统,并用于食品安全舆情预警,有助于增强食品企业的舆情应对能力。[方法/过程]通过ResNet思想连接BERT模型,BiGRU模型与前馈神经网络构建BERT-RGRU模型,用于弹幕情感分析。在情感标签划分中提出“假中性”的概念以增强模型的情感判别能力,再使用“注意力加权”的方式处理情感分析模型判别结果,计算预警指标取值,并以此作为当日是否存在负面舆情的判定依据。[结果/结论]BERT-RGRU情感分析模型在弹幕测试集上的F1值,比BERT模型提升2%,比传统BiGRU等模型提升10%以上,舆情预警系统在麦当劳实证分析中,准确地对食品安全负面舆情做出预警。

[Purpose/significance]The negative public opinion warning system is constructed by sentiment of comments and bullet screen,which can analyze public opinion of food security and enhance the ability of food enterprises to respond to public opinion.[Method/process]ResNet method is used to link BERT model,BiGRU model,and feedforward neural network to construct the BERT-RGRU model,which can analyze the sentiment of bullet screen.The concept of"fake-neutral"is proposed to enhance sentiment analysis ability of the model.Method of"attention-weight"is used to process results of sentiment analysis model and calculate the value of early warning index,which can be used to determine whether there is a negative public opinion.[Result/conclusion]BERT-RGRU model performs well on the bullet screen testing data,F1-score is 2%higher than BERT model and at least 10%higher than traditional model like BiGRU.The public opinion early warning system also raises the alarm accurately in McDonald's real example.