基于LSTM-BLS的突发气象灾害事件中公众情感倾向分析

Analysis of public sentiment tendency in sudden meteorological disasters based on LSTM-BLS

近年来长短期记忆网络(LSTM)在文本情感倾向分析方面显示出一定优势,但LSTM提取特征时存在语义不完整、精度不高等问题.研究者往往通过引入卷积神经网络(CNN)来弥补这一缺陷,但仍然未考虑到单词之间的句法依存问题.本文将以增量学习算法为核心的宽度学习(BLS)与LSTM相融合,提出了LSTM-BLS文本情感分析模型,并以2020断崖式降温事件为例,对突发气象灾害发生时公众情感倾向进行分析.结果表明:与基线模型K-means和支持向量机(SVM)相比,LSTM-BLS模型精度分别提高17.23和13.46个百分点;与已有深度模型LSTM、CNN-LSTM相比,本文模型精度分别提高7.13和4.17个百分点.

Meteorological disasters will not only bring huge economic losses,but also cause the outbreak of public opinion and even social panic.The Internet era brings opportunities as well as challenges to the governance of public opinion.In recent years,Long Short-Term Memory network(LSTM)has shown advantages in text sentiment analysis,yet it has problems such as incomplete semantics and low accuracy in feature extracting.Convolutional Neural Network(CNN)has been introduced to make up for this shortcoming,but it still cannot considere the syntactic dependence between words.Here,we combine LSTM and Broad Learning System(BLS)with incremental learning algorithm as its core,and thus propose an LSTM-BLS text sentiment analysis model.Then the 2020 cold wave event with gale process and great temperature drop in central and eastern China is taken as an example to analyze the public sentiment tendency when sudden meteorological disasters occur.The results show that the proposed LSTM-BLS reaches a high accuracy of 97.42%,which is 17.23 and 13.46 percentage points higher than the baseline models of K-means and Support Vector Machine(SVM),respectively,and 7.13 and 4.17 percentage points higher than the existing deep learning models of LSTM and CNN-LSTM,respectively.