Attention-based multi-level image and text sentiment analysis

Abstract:
People are keener and keener to express their opinions and share their lives on social media. Analyzing this information with an apparent emotional tendency has broad application prospects. Since the multi-dimensional characteristics of each modality and the interaction between modalities have not been considered in the previous graphic emotion analysis, this paper proposes an emotion analysis model (BRFANet model) based on various neural networks and attention mechanisms. Firstly, ALBERT and BiGRU are used to capture the deep semantic information of the text. Secondly, residual network and Faster RCNN combined with PFN are used to capture the image's global and local features, respectively. Then, to capture the target object that can better reflect emotions, the text features are used to guide attention interaction on local images. Finally, the self-attention network is used to automatically identify the key weights of each mode to complete the fusion. On the Flickr dataset, compared with the traditional multimodal sentiment analysis method, the classification accuracy and F1 value are improved, proving the model's effectiveness.