Visual and textual sentiment analysis of daily news social media images by deep learning

In recent years, following the exploding spread of Social Networking platforms, more and more people have started to share large quantities of data on Internet where they express personal opinions, ideas or emotional states regarding any kind of topic. With the increasing amount of these type of data, finding a way to analyzing it has become a need for major companies, political parties or whatever organization based on its own customers’ feedback. This paper proposes a new case of study that has seen a limited assortment of similar proposals in the current state-of-art: in particular, we propose an innovative approach for analyzing both visual and textual features of Social Media images using Deep Convolutional Neural Networks (DCNNs), in order to collect more accurate results than the single analysis of both type can do alone. The deep learning approach estimates the overall sentiment of daily news-related pictures from social media based on both visual and textual clues. The proposed approach was applied and tested on a new public dataset with more than 9.000 annotated Instagram images. Experimental results confirmed the effectiveness of the approach, showing high values of accuracy.