基于文档顺序与多模态模型的金融票据信息抽取OACSTPCD

The current methods for extracting information from documents mostly work well for simple documents but are not effective for extracting information from complex financial documents that contain background noise and structural complexity.To ad-dress the problem of matching entity relationships in complex financial documents,a sequential reconstruction method and the Lay-outLMv3-GRU information extraction model are proposed.It creates a complex financial document dataset that incorporates text,layout,and image modalities for information extraction.Using the Layout-Parser tool,it designs a sorting module to arrange text in-formation based on contextual relationships and rearrange words that are far apart spatially but closely related logically.By combin-ing the improved LayoutLMv3 model with the GRU network,it further improves the accuracy of the model.It conducts experiments on the public dataset FUNSD and the self-built complex financial dataset.The results show that our method achieves a 2.37%im-provement in F1 score compared to the LayoutLMv3 model.Particularly on the self-built complex financial dataset,the model achieves an F1 score of 88.36%,demonstrating the superiority of the method in extracting information from complex documents and its general applicability in handling various types of documents.