2021/06/20 by Chen‐Yu Lee, Chunliang Li, Lee, Chen-Yu +13 · 1 citation
Computer Science · #Advanced Graph Neural Networks #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2106.10786
openalex publication_date 2021/06/20 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Natural reading orders of words are crucial for information extraction from\nform-like documents. Despite recent advances in Graph Convolutional Networks\n(GCNs) on modeling spatial layout patterns of documents, they have limited\nability to capture reading orders of given word-level node representations in a\ngraph. We propose Reading Order Equivariant Positional Encoding (ROPE), a new\npositional encoding technique designed to apprehend the sequential presentation\nof words in documents. ROPE generates unique reading order codes for\nneighboring words relative to the target word given a word-level graph\nconnectivity. We study two fundamental document entity extraction tasks\nincluding word labeling and word grouping on the public FUNSD dataset and a\nlarge-scale payment dataset. We show that ROPE consistently improves existing\nGCNs with a margin up to 8.4% F1-score.\n