跨模LDA融合的多模态数据主题分析方法

Multimodal data topic analysis method based on cross-modal LDA fusion

随着互联网的高速发展,社会大众可以通过网络对医疗事件以及医患关系自由地发表个人意见和观点言论,这对于引导公众正确的价值导向有着重大研究意义.然而,仅考虑单模态数据的主题分析算法不能精准地把握整个舆情事件的真相,存在主题提取不准确、个人情感先入为主等问题.提出一种基于LDA的多模态数据主题分析算法MD_LDA(multimodal data topic analysis based on LDA).通过对各模态主题分析结果进行决策级融合来计算多模态的主题分析结果,进而解决传统方法对多模态数据考虑不全面的缺陷.实验结果表明,针对多模态舆情事件,在主题词的提取效果上,所提出的MD_LDA算法优于单一模态数据进行主题分析的算法.而相对于传统的关键词提取算法TF_IDF与TextRank和MD_LDA算法的准确率以及主题词提取效率均有所提高,验证了结合多模态数据进行主题分析的MD_LDA算法的有效性.

With the rapid development of the Internet,the public can freely express personal opinions on medical events and doctor-patient relationships through the Internet,which are of the correct value for guiding the public.Orientation has great research significance.However,the topic analysis algorithm that only considers single-modal data cannot accurately grasp the truth of the entire public opinion event,and there are problems such as inaccurate topic extraction and preconceived personal emotions.To solve this problem,this paper proposes a LDA-based multimodal data topic analysis algorithm,named MD_LDA(multimodal data topic analysis Based on LDA).The multimodal topic analysis is calculated by the decision-level fusion of the results of each modal topic analysis.As a result,it further solves the defect that traditional methods do not fully consider multimodal data.The experimental results show that for multimodal public opinion events,the proposed MD_LDA algorithm is better than the algorithm for topic analysis of single-modal data in terms of the extraction effect of topic words.Compared with the traditional keyword extraction algorithms TF_IDF and TextRank,the accuracy of the MD_LDA algorithm and the extraction efficiency of subject words are improved,which proves the effectiveness of the MD_LDA algorithm for subject analysis combined with multimodal data.