HRDoc: Dataset and Baseline Method toward Hierarchical Reconstruction of Document Structures

The problem of document structure reconstruction refers to converting digital or scanned documents into corresponding semantic structures. Most existing works mainly focus on splitting the boundary each element in a single page, neglecting multi-page documents. This paper introduces hierarchical structures as novel task suitable for NLP and CV fields. To better evaluate system performance new task, we built large-scale dataset named HRDoc, which consists 2,500 with nearly 2 million units. Every HRDoc has line-level annotations including categories relations obtained from rule-based extractors human annotators. Moreover, proposed an encoder-decoder-based parsing (DSPS) tackle this problem. By adopting multi-modal bidirectional encoder structure-aware GRU decoder soft-mask operation, DSPS model surpass baseline method by large margin. All scripts datasets will be made publicly available at https://github.com/jfma-USTC/HRDoc.