Document Information Extraction Framework Based on Graph Enhanced Language Model

During the process of intelligent transformation and digital upgrading in the industry, a large amount of document data containing both structured and unstructured information has been generated. Efficiently and accurately extracting valuable information from these documents has become a pressing issue for both industry and academia. Traditional document information extraction models primarily focus on document pre-training models that emphasize entity feature representation within the documents. However, combining the structural information of entity pairs in unstructured documents can help improve the accuracy of information extraction. This paper proposes a Visual-Rich Document Information Extraction Framework based on Graph Neural Network Enhanced Language Model (GELM). This method leverages the document reconstruction capabilities of pre-trained models enhanced by graph neural networks. It uses the visual, textual, and structural aspects of documents as inputs to the graph neural network and integrates the outputs of the graph neural network and the pre-trained model through a biaffine attention mechanism, thereby extracting entities and their relationships from the documents. Comparative experimental results show that the proposed method improves the performance by ${1 \% - 2 \%}$ compared to the baseline pre-trained models, validating its effectiveness.