A method for text line aggregation based on the BERT structure

When performing image and document information extraction tasks, OCR technology is the primary tool used to extract text and layout information from documents. Subsequently, deep learning models such as LSTM, Transformer or BERT are used in order to accomplish specific information extraction tasks. However, in the face of complex document layouts, existing open-source OCR tools fall short in recognizing and labeling complete sentences, often resulting in continuous sentences being incorrectly segmented into multiple parts. This segmentation affects the order of the model's input to the text during the information extraction task, thus indirectly affecting the efficiency of the extraction process. To address these issues, this study proposes a BERT-based language model, BertNPP, which leverages BERT's powerful capabilities in text comprehension to efficiently aggregate text lines by mining and exploiting text features through a specially designed pre-training task. Finally, experiments were conducted on real datasets to validate the effectiveness of the model.