中国科学院自动化研究所机构知识库

Self-supervised multi-modal document pre-training for document knowledge learning shows superiority in various downstream tasks. However, due to the diversity of document languages and structures, there is still room to better model various document layouts while efficiently utilizing the pre-trained language models. To this goal, this paper proposes a Graph-based Multi-level Layout Language-independent Model (GraphMLLM) which uses dual-stream structure to explore textual and layout information separately and cooperatively. Specifically, GraphMLLM consists of a text stream which uses off-the-shelf pre-trained language model to explore textual semantics and a layout stream which uses multi-level graph neural network (GNN) to model hierarchical page layouts. Through the cooperation of the text stream and layout stream, GraphMLLM can model multi-level page layouts more comprehensively and improve the performance of language-independent document pretrained model. Experimental results show that compared with previous state-of-the-art methods, GraphMLLM yields higher performance on downstream visual information extraction (VIE) tasks after pre-training on less documents. Code and model will be available at https://github.com/HSDai/GraphMLLM.