PaperlayerView exact revision on arXiv

Source-linked AI summary

Complicated Table Structure Recognition

Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, Xian-Ling Mao

arXiv:1908.04729v2cs.IRcs.LG

TL;DR

Existing methods struggle to recognize complicated PDF tables with spanning cells, a key obstacle for machine-understandable table structure. The paper proposes GraphTSR, a graph neural network that predicts relations among table cells, and introduces SciTSR with 15,000 labeled tables. GraphTSR outperforms state-of-the-art baselines across benchmark and SciTSR datasets, particularly for complicated tables.

  • Problem

    Existing methods have difficulty accurately recognizing complicated PDF tables containing spanning cells, limiting reliable table-structure recognition.

  • Method

    GraphTSR represents PDF tables as graphs and predicts relations among cells to recover their structure, alongside the SciTSR dataset of 15,000 labeled tables.

  • Results

    GraphTSR outperforms state-of-the-art baselines on benchmark and SciTSR datasets, with at least 2% higher F1 on ICDAR-2013 and SciTSR and at least 7% higher F1 on SciTSR-COMP.

  • Takeaways & Limitations

    Graph-based relation prediction provides a strong approach for recognizing complicated PDF table structures, while SciTSR supplies large-scale labeled data for this task.

  • Takeaways & Limitations

    TableBank cannot be used for this task because its structure labels omit column-coordinate information.

Abstract

from arXiv · show

The task of table structure recognition aims to recognize the internal structure of a table, which is a key step to make machines understand tables. Currently, there are lots of studies on this task for different file formats such as ASCII text and HTML. It also attracts lots of attention to recognize the table structures in PDF files. However, it is hard for the existing methods to accurately recognize the structure of complicated tables in PDF files. The complicated tables contain spanning cells which occupy at least two columns or rows. To address the issue, we propose a novel graph neural network for recognizing the table structure in PDF files, named GraphTSR. Specifically, it takes table cells as input, and then recognizes the table structures by predicting relations among cells. Moreover, to evaluate the task better, we construct a large-scale table structure recognition dataset from scientific papers, named SciTSR, which contains 15,000 tables from PDF files and their corresponding structure labels. Extensive experiments demonstrate that our proposed model is highly effective for complicated tables and outperforms state-of-the-art baselines over a benchmark dataset and our new constructed dataset.

1 Introduction

Table structure recognition helps machines understand tables, but existing PDF methods struggle with complicated tables containing spanning cells. GraphTSR addresses this challenge with a graph-based model, while SciTSR provides 15,000 labeled PDF tables for evaluation.

  • 1 Introduction: The task supports machine understanding of tables for applications including question answering, dialogue systems, and table-to-text.Recognized internal structure is a key step toward machine-understandable tables.
  • 1 Introduction: Existing methods often fail on complicated PDF tables containing spanning cells that occupy multiple rows or columns.Such cells are especially important because they are more likely to be table headers.
  • 1 Introduction: GraphTSR reformulates PDF table structure recognition as graph edge prediction among table cells.The model uses graph neural networks to predict relations among cells.
  • 1 Introduction: Experiments report that GraphTSR outperforms state-of-the-art baselines, especially on complicated tables.The paper evaluates the model on a benchmark dataset and the new SciTSR dataset.
  • 1 Introduction: 15,000 labeled tables from scientific papers form the SciTSR dataset for PDF table structure recognition.The dataset was introduced because suitable large-scale training data were unavailable.

2 Related Work

Prior PDF table-structure methods include rule-based and data-driven approaches, but existing approaches work well mainly on simple grid-like tables. Dataset limitations also motivate the construction of SciTSR.

  • 2 Related Work: Existing table-structure methods are grouped into rule-based and data-driven approaches.The related work covers both traditional heuristics and deep-learning methods.
  • 2 Related Work: Existing approaches generally fail on complicated tables with spanning cells, motivating a graph neural network solution.The paper characterizes prior methods as effective mainly for simple grid-like tables.
  • 2 Related Work: ICDAR-2013 contains only 156 tables, making it too small to support data-driven table-structure models.Its PDF documents are well labeled, but the dataset size is limited.
  • 2 Related Work: The proposed method processes PDF tables through preprocessing, graph construction, relation prediction, and post-processing.The pipeline builds an undirected graph over extracted cells and recovers the table structure from predicted relations.
  • 2 Related Work: TableBank provides image-format tables but omits column coordinates from its structure labels, so it cannot be used for this task.This limitation motivates constructing a new PDF table-structure dataset.

3 Method

GraphTSR recognizes PDF table structure by predicting labeled relations between cells represented in a graph. It combines KNN graph construction with local graph attention over vertex and edge features before classifying edges.

  • 3 Method: The method processes PDF tables through cell extraction, graph construction, relation prediction, and post-processing to recover the table structure.
  • 3.1 Problem Definition: Each table cell becomes a vertex, while vertical and horizontal adjacency relations become labeled edges that encode the target structure.
  • 3.2 Graph Construction: KNN graph construction connects each cell to its K nearest neighbors, reducing candidate edges from quadratic scale to O(K|V|).The resulting edges are classified as vertical, horizontal, or no relation.
  • 3.3 GraphTSR: GraphTSR computes vertex and edge representations through alternating edge-to-vertex and vertex-to-edge graph attention blocks, then classifies graph edges.The architecture supports edge features by representing original edges as additional nodes in a bipartite graph.
  • 3.3 GraphTSR: Vertex features include cell size and absolute or relative locations, while edge features include distances and x- and y-axis overlaps.
  • 3.3 GraphTSR: Graph attention uses local dependencies among neighboring nodes rather than drawing global dependencies between all nodes.The attention blocks use neighboring nodes and update representations through learned transformations and normalization.

4 SciTSR Dataset

SciTSR is a large-scale PDF table-structure dataset built from scientific-paper sources, with labels derived from table content and layout. It includes training, test, and complicated-table subsets for evaluating structure-recognition systems.

  • 4 SciTSR Dataset: SciTSR contains 15,000 PDF tables with corresponding structure labels, constructed from scientific papers.The dataset is introduced as a large-scale resource for table structure recognition.
  • 4 SciTSR Dataset: The construction pipeline extracts LaTeX table snippets, compiles them into PDF files, and derives labels from cell contents and coordinates.The pipeline uses LaTeX source files and exports the resulting cell information in JSON format.
  • 4 SciTSR Dataset: SciTSR is split into 12,000 training tables and 3,000 test tables, averaging about 9 rows, 5 columns, and 48 cells per table.These statistics describe the overall dataset composition.
  • 4 SciTSR Dataset: The dataset supports evaluation against ICDAR-2013 and reports macro- and micro-averaged experimental results on SciTSR and SciTSR-COMP.The reported result tables cover ICDAR-2013, SciTSR, and SciTSR-COMP.
  • 4 SciTSR Dataset: Complicated tables contain at least one spanning cell, and SciTSR provides the 716-table SciTSR-COMP test subset for focused evaluation.The complicated-table subset is extracted from the test set.

5 Experiment

GraphTSR is evaluated against rule-based, data-driven, and commercial baselines on PDF table-structure datasets, including complicated tables. It consistently performs best, with especially strong gains on spanning-cell structures and generalization across datasets.

  • Experimental setup: The evaluation uses adjacent-cell relation precision and recall, summarized as macro- and micro-averaged scores across PDF table datasets.The datasets include SciTSR and ICDAR-2013, while the baselines are Tabby, DeepDeSRT, and Adobe.
  • Overall results: At least 2% higher F1 on ICDAR-2013 and SciTSR, and at least 7% higher F1 on SciTSR-COMP, GraphTSR outperforms all baselines.The results are reported using macro- and micro-averaged scores in Tables 3 and 4.
  • Overall results: GraphTSR remains best when trained on SciTSR and tested on ICDAR-2013, demonstrating stronger cross-dataset generalization than DeepDeSRT and the other baselines.The authors attribute image-input sensitivity to fonts or styles as a reason some methods fail on unseen table appearances.
  • Results on complicated tables: On SciTSR-COMP, every baseline drops at least 4% in performance, whereas GraphTSR maintains high F1 scores for complicated tables.The dataset contains complicated tables, but non-spanning cells remain dominant, so additional spanning-cell-only evaluation is also performed.
  • Case study: In a SciTSR-COMP case study, GraphTSR correctly recognizes the table structure, while Adobe merges cells and DeepDeSRT splits spanning cells into non-spanning cells.The compared outputs are visualized with differently colored cells in Figure 5.
  • Impact of the number of attention blocks: Increasing the number of graph attention blocks improves performance, especially recall, because small values limit how far graph nodes can access surrounding cells.With too few blocks, the model tends to predict “no relation” between cells.

6 Conclusion

The paper proposes GraphTSR for complicated table structure recognition in PDF files and releases SciTSR, a large-scale labeled dataset. Experiments show the model is effective on complicated tables.

  • 6 Conclusion: The paper proposes a graph neural model for complicated table structure recognition in PDF files and releases a dataset containing 15,000 labeled tables.The dataset contains tables and their corresponding structure labels.
Loading 1908.04729v2…