The code of our paper "Entity-related Unsupervised Pretraining with Visual Prompts for Multimodal Aspect-based Sentiment Analysis"
The MABSA dataset can be derived from the paper: Vision-Language Pre-Training for Multimodal Aspect-Based Sentiment Analysis (https://github.com/NUSTM/VLP-MABSA)
The pre-training dataset can download from the COCO2014: https://cocodataset.org/
The split_coco.py is used to split COCO2014 for pre-training.
We use clip-vit-base-patch16 to extract image feature.
parse_coco.py and parse_twitter.py is used to pre-process data.
We use flan-t5-base and t5-base to initialize our model.