Social media data multi-modal attitude analysis method

Abstract

The invention provides a social media data multi-modal attitude analysis method, which comprises the following steps of: selecting a picture from social media data, and performing text conversion on the picture by adopting a picture description method with an attention mechanism to generate a converted text; considering the integrating degree of the original text content of the picture and the content of the converted text; connecting the original text with high integrating degree with the converted text to realize the fusion of two types of modal data; constructing an image-text pair data set, and performing data fusion on the image-text pair data set by adopting the steps to generate an integrated text; according to the social media data multi-modal attitude analysis method, a pre-training BERT model is trained through an integrated text, a text attitude analysis model is generated through optimization, and attitude analysis is carried out on a tweet containing pictures and texts on the basis of the model. Compared with a traditional attitude analysis method, the method has the advantage that the accuracy of public attitude control is improved.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and in particular relates to a multi-modal attitude analysis method for social media data. Background technique [0002] Nowadays, all kinds of information released by the public on social media in response to a public event fully reflect public attitudes, so the analysis of such information is of great significance for controlling public attitudes. However, when users on social media express their opinions, they usually use not only single words, but also pictures and videos. The significance of using pictures and texts in tweets instead of just judging the attitude polarity of tweets based on the text part is: (1) Natural language, especially the language on social media, often has irony and irony. Using the information in the pictures can reduce the misleading of natural language rhetoric; (2) The pictures on social media often contain information that is not involved in the text, su...

AI Technical Summary

Problems solved by technology

It is easy to see that the features of each mode extracted by this method are the underlying features. The disadvantage of directly using the underlying features for mode fusion is that the internal features of each mode cannot be well combined.

However, the disadvantages of this method are: poor interpretability, the intermediate features from the two modalities are heterogeneous, and each dimension of the feature has different meanings, and the meaning of the interaction between the two during the fusion process is unclear; and due to poor interpretability, it is difficult to rely on rules Heuristically selects the best intermediate features, only blindly tried

Images