文章来源：Xu N, Mao W. Multisentinet: A deep semantic network for multimodal sentiment analysis[C]//Proceedings of the 2017 ACM on Conference on Information and Knowledge Management. 2017: 2399-2402.（CCF B）
1. Model structure
Looking at this article in 2017 from the current perspective, we can find that the idea is very simple. The idea is as follows:
In the image data, the target and scene information of the image are closely related to the final emotion classification. The target feature information and scene feature information in the image are extracted by using the VGG pre-trained on the ImageNet and Place-365 datasets respectively.
In the text data, after the Glove word vector representation, the LSTM network structure is used to integrate the contextual information of the text into the word vector, and finally the sum of all word vectors is used as the final representation of the tweet. The resulting tweet representation vectors imply that the contribution of all word vectors is the same, but this is clearly not true. In this paper, target feature vectors and scene feature vectors are used to carry out a cross-modal attention mechanism with word vectors, and high weights are given to word vectors related to emotions in the text, so as to obtain the representation vector of the final tweet .
Decision-making stage: The three features are sent to the multi-layer perceptron, and the softmax function is used for prediction. Use the cross-entropy function as the objective loss function.
The experimental results are as follows:
To sum up, the idea of this article is somewhat simple. Compared with another article written before, it is more like the predecessor of that article.
The link to the article is as follows: https://blog.csdn.net/qq_43775680/article/details/127717392 
https://blog.csdn.net/qq_43775680/article/details/127717392