Application of BERT-Based Semantic Matching Algorithm for Cross-Page Table Recognition

In the financial field, PDF documents often contain complex cross-page tables that pose challenges for conventional recognition methods. Existing research focuses on extracting information and merging tables at a high level, overlooking the text segmentation issues within individual table cells. This paper presents a two-step approach for cross-page table recognition: detecting cross-page tables and merging them. To address the challenge of merging cross-page tables, a BERT-based semantic matching algorithm is introduced by leveraging BERT’s NSP (Next Sentence Prediction) pre-training task for fine-tuning. To address the challenge of collecting cross-page table data, datasets for the matching task are successfully generated using the pdfplumber. Experimental results demonstrate the high accuracy of the proposed method, validating the effectiveness of the dataset generation approach.