TableBank-Word: Weakly Supervised Table Detection

• TableBank-Word is a corpus built from .docx files, featuring 163,417 table detection annotations and 56,866 structure recognition labels derived through weak supervision.
• It employs a document-centric pipeline with pixel-level comparison between original and annotated PDFs to extract reliable table bounding boxes and HTML-like structural sequences.
• Semi-supervised and active learning studies on TableBank-Word demonstrate its value as a challenging benchmark for annotation-efficient table detection with diverse, irregular layouts.

TableBank-Word is the Word-derived subset of TableBank, an image-based corpus for table detection and table structure recognition constructed with weak supervision from .docx documents. In the original TableBank release, the Word subset contains 163,417 labeled tables for detection and 56,866 structure-recognition instances; later work also treats the Word partition as a benchmark for single-class table detection with bounding boxes, including semi-supervised Deformable-DETR experiments and active-learning studies (Li et al., 2019, Ehsan et al., 2024, Gautam et al., 24 Sep 2025).

1. Corpus definition and scope

Within the full TableBank corpus, the Word subset is the portion built from crawled .docx Word documents. The original dataset paper reports 417,234 labeled tables for detection across Word and LaTeX, with the Word share equal to 163,417; for table structure recognition, the combined total is 145,463, of which the Word share is 56,866 (Li et al., 2019). The dataset therefore supports two distinct tasks: table detection, where the output is page-level table rectangles, and table structure recognition, where the output is an HTML-like structural sequence.

Later papers reuse the same subset under different experimental protocols, so the reported counts are not always directly comparable. The 2024 semi-supervised detector paper treats TableBank-Word as the Word partition with 163,417 images, while the 2025 active-learning paper reports a dedicated split of 73,383 training images, 2,735 validation images, and 2,281 test images for TableBank-Word (Ehsan et al., 2024, Gautam et al., 24 Sep 2025). This difference reflects distinct task setups and split definitions rather than a single canonical count.

Source	Quantity	Meaning
TableBank (2019)	163,417	Word detection annotations
TableBank (2019)	56,866	Word structure-recognition instances
Semi-supervised detection (2024)	163,417	Word partition used in detection experiments
Active learning (2025)	73,383 / 2,735 / 2,281	Train / validation / test images

The original paper does not explicitly report the number of Word source documents or the number of Word page images overall. It reports counts of labeled tables for detection and sampled evaluation sets for detection and recognition instead.

2. Weak-supervision pipeline for Word documents

The Word subset is built by crawling Word documents from the internet and using only .docx files, because .docx files can be edited internally through Office XML (Li et al., 2019). In the decompressed folder, the relevant content resides in document.xml, where table regions are represented by markup between <w:tbl> and </w:tbl>. These native structural tags provide the basis for weak supervision.

For table detection, the construction pipeline is explicitly document-centric. The .docx file is decompressed; the XML snippets corresponding to tables are modified so that table borders become a distinguishable color relative to the rest of the document; the example in the paper uses a green bounding box; the modified document is rendered to PDF; and only PDF pages that contain at least one table are retained. Ground-truth boxes are then recovered by comparing the original page and the annotated page at the pixel level. The differing pixels reveal the inserted table borders, from which the annotation extractor recovers the upper-left point, the width, and the height.

The original paper also describes an alignment step: for each annotated PDF page, it adds the bounding boxes to the tables in the original page with white color so that the two pages align in the same position. This makes the pixel-level comparison operationally stable for extracting the rectangular box coordinates.

The reported annotation quality is high but not perfect. A manual sample of 1,000 examples contained only 5 incorrectly labeled, corresponding to 0.5% incorrect and 99.5% correct in that spot check (Li et al., 2019). The paper also illustrates failure modes, including a missed table, a mixed case where two tables are annotated as one, an incomplete annotation, and overlapped annotations. Those examples establish that TableBank-Word is weakly supervised rather than manually exhaustively curated.

For table structure recognition, the labeling mechanism differs. The target is an HTML tag sequence representing the arrangement of rows and columns and cell types, and for Word documents the labels are created by transforming the original XML information from documents into an HTML tag sequence. The output vocabulary contains 12 structure tokens: <tabular>, </tabular>, <thead>, </thead>, <tbody>, </tbody>, <tr>, </tr>, <td>, </td>, <cell_y>, and <cell_n>. The semantic distinction is that <cell_y> denotes a cell with content and <cell_n> a cell without content.

3. Tasks, splits, and evaluation protocols

In the original TableBank formulation, table detection takes a page image from a Word document and predicts table bounding boxes on the page, while table structure recognition takes a cropped table image and predicts an HTML-like structure sequence (Li et al., 2019). Detection annotations are rectangular boxes; recognition annotations are sequence labels derived from Word XML.

The original Word detection evaluation protocol samples 2,000 document images from Word documents, split into 1,000 validation and 1,000 test images, with each image containing at least one table. Recognition evaluation samples 500 tables for validation and 500 tables for testing from Word documents. For detection, the paper uses area-based precision, recall, and F1 Score:

$$Precision=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Detected table regions}}$$

$$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$

$$F1\ Score=\frac{2 \times Precision \times Recall}{Precision + Recall}$$

For recognition, the reported metric is 4-gram BLEU with a single reference.

Later studies adopt different evaluation protocols. The 2024 semi-supervised detector paper evaluates table detection only, not structure recognition, and frames the outputs as single-class object-detection bounding boxes. It reports mAP, AP$^ {50}$, AP$^ {75}$, and AR$_L$, and performs semi-supervised training with 10%, 30%, and 50% of the data labeled and the remaining data treated as unlabeled; it does not report a conventional train/validation/test split for TableBank-Word (Ehsan et al., 2024). The 2025 active-learning paper likewise uses TableBank-Word for single-class table detection with bounding boxes, and reports mAP at annotation budgets of 2k, 4k, 6k, 8k, and 10k on its 73,383 / 2,735 / 2,281 train/validation/test split (Gautam et al., 24 Sep 2025).

This multiplicity of protocols means that “TableBank-Word” names a stable data source but not a single universal evaluation regime. A plausible implication is that numerical results are directly comparable only within a paper’s own split and metric conventions.

4. Visual domain characteristics and relation to TableBank-LaTeX

TableBank-Word is explicitly differentiated from the LaTeX portion of TableBank by source, language coverage, and visual regularity. The Word subset is crawled from the internet, uses .docx files, and includes English, Chinese, Japanese, Arabic, and other languages because the authors do not filter document language (Li et al., 2019). In contrast, the LaTeX subset comes mainly from arXiv, spans source code from 2014 to 2018, and is mainly English.

Later work sharpens the domain characterization. The 2025 active-learning paper describes TableBank-Word as a subset of images from Microsoft Word documents, including business reports, forms, and various document types, and states that its tables have more diverse and irregular table layouts, varied formatting styles, fonts, and colors, complex table structures with merged cells and nested elements, and greater variability in table appearance due to user-defined formatting (Gautam et al., 24 Sep 2025). It contrasts this with TableBank-LaTeX, whose tables are described as more structured, grid-like, consistent, and uniform.

The original TableBank baselines quantify the domain gap. For table detection on the Word test set, the best same-domain result is the Word-trained ResNeXt-152 with P 0.9530, R 0.8829, and F1 0.9166; the best LaTeX-trained model evaluated on Word reaches F1 0.8863 (Li et al., 2019). For structure recognition on Word, the Word-trained image-to-text model reaches BLEU 0.7507, whereas the LaTeX-trained model reaches 0.4048. These results support the paper’s conclusion that models perform well on the same domain and that tables from different document types have different visual appearance.

The cross-domain behavior also indicates that mixed training improves generalization without necessarily maximizing Word-only accuracy. On Word detection, the best mixed Word+LaTeX-trained model reaches F1 0.9067, which is below the 0.9166 of the Word-trained model; on Word structure recognition, the mixed model reaches 0.7121, below the Word-trained model’s 0.7507 (Li et al., 2019). This suggests that TableBank-Word is not merely a smaller variant of TableBank-LaTeX but a visually and statistically distinct operating domain.

5. Semi-supervised table detection on TableBank-Word

In the 2024 paper, TableBank-Word is a principal benchmark for semi-supervised table detection in document images (Ehsan et al., 2024). The method is an end-to-end semi-supervised deformable-DETR-style detector built around two query sets,

$Q=\left\{q_{1}, q_{2}, \cdots, q_{n} \right\}$

and

$$\hat{Q}=\left\{ \hat{q}_{1}, \hat{q}_{2}, \cdots, \hat{q}_{T} \right\},$$

corresponding respectively to one-to-one and one-to-many assignment branches. The stated motivation is that CNN-based semi-supervised detectors rely on anchors and NMS, while transformer-based DETR-style semi-supervised methods suffer from strict one-to-one assignment that can yield weak or noisy pseudo-label supervision, especially early in training.

The detector uses a student and a teacher. The teacher processes only unlabeled images under weak augmentation and produces pseudo-labels; the student processes both labeled and unlabeled images, uses strong augmentation for unlabeled images, learns from labeled data and teacher-generated pseudo-labels, and the teacher is updated by EMA from the student. The paper conceptually defines labeled and unlabeled sets as

$$D_{l} = \left\{ x^{l}_{i}, y^{l}_{i} \right\}^{N_{l}}_{i=1}$$

and

$D_{u} = \left\{ x^{u}_{i} \right\}^{N_{u}}_{i=1}.$

Pseudo-labels are filtered by a confidence threshold of 0.7.

For TableBank-Word, the reported implementation details are: ResNet-50 backbone pre-trained on ImageNet; Deformable DETR-based detector; 10%, 30%, and 50% labeled ratios; 150 epochs of training; learning rate reduced by a factor of 0.1 at epoch 140; strong augmentations consisting of horizontal flips, resizing, patch removal, cropping, grayscale, Gaussian blur; weak augmentation consisting of horizontal flipping; 30 one-to-one queries; 400 one-to-many queries; ground-truth replication factor 6; loss weights $$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$0 for classification and $$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$1 for box regression; input image size 600 pixels in the efficient setup and 800 pixels for comparison with prior work; and one-to-one branch only at inference, preserving NMS-free DETR behavior.

The title refers to “modulated object queries,” but the paper does not provide a standalone mathematical modulation mechanism such as an explicit equation of the form $$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$2. What is concretely specified is the dual-query, dual-assignment training design. The strongest technically grounded reading is therefore that the “modulated” aspect is operationalized through dual query assignment and filtering, while the main novelty lies in the hybrid matching strategy.

The TableBank-Word results are reported partition-specifically in Table 3. With 10% labeled data, the method reaches mAP 84.1, AP50 87.1, AP75 85.5, and AR$$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$3 96.2. With 30% labeled data, it reaches mAP 95.7, AP50 96.9, AP75 96.2, and AR$$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$4 98.1. With 50% labeled data, it reaches mAP 95.5, AP50 98.5, AP75 97.2, and AR$$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$5 98.7. The best mAP is therefore 95.7 at 30% labeled, whereas the best AP50 is 98.5 at 50% labeled. Within TableBank-Word, the change from 10% to 30% labeled yields a +11.6 mAP increase, while 30% to 50% yields a -0.2 change in mAP and a +1.6 increase in AP50.

The paper argues more broadly that its method surpasses prior supervised and semi-supervised methods on TableBank benchmarks, but its comparison tables are not cleanly resolved to the Word partition. The most direct Word-specific conclusion is therefore limited to the 95.7 mAP result at 30% labeled data and the accompanying AP50, AP75, and AR$$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$6 values.

6. Active learning on TableBank-Word

The 2025 paper studies active learning for single-class table detection on TableBank-Word using CascadeTabNet and YOLOv9 (Gautam et al., 24 Sep 2025). The setup uses the split 73,383 train, 2,735 validation, 2,281 test. The paper emphasizes that the single-class nature of the problem changes the informativeness of conventional uncertainty measures, and positions TableBank-Word as the harder domain because Word-generated tables are less regular than LaTeX-generated ones.

Four selection strategies are evaluated in addition to a Random baseline: Prediction Uncertainty, Mask Ambiguity, Bounding-Box Ambiguity, and Table Count. The active-learning loop begins with an initial labeled set, trains a detector, performs inference on the unlabeled pool, scores or filters unlabeled examples, annotates selected samples, continues from the previous state, and evaluates. The reported hyperparameters specific to TableBank-Word are sparse: the uncertainty threshold is 0.95, and the IoU threshold for the Bounding-Box Ambiguity score on TableBank-Word is $$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$7, chosen “to ensure a sufficiently large sample set under budget constraints.”

The methods are defined operationally. Prediction Uncertainty bins images by confidence and prioritizes lower-confidence bins; for images with multiple predicted tables, the image-level score is the average confidence across predicted tables. Bounding-Box Ambiguity targets images with multiple overlapping predicted boxes. Mask Ambiguity uses the score $$Recall=\frac{\text{Area of Ground truth regions in Detected regions}}{\text{Area of all Ground truth table regions}}$$8 between a detection mask and a segmentation mask, computed over all bounding-box-occupied and segmentation-mask-occupied regions in the page. Table Count prioritizes images containing more than one table, giving higher sampling probability proportionate to the number of tables. The paper also notes that some printed formulas are corrupted, especially the confidence-rate expression and an entropy-style uncertainty expression, so only the qualitative selection logic is unambiguous.

For CascadeTabNet on TableBank-Word, the reported mAP values at budgets 2k / 4k / 6k / 8k / 10k are: Random 82.2 / 85.1 / 86.8 / 87.8 / 88.8; Prediction Uncertainty 83.1 / 87.3 / 88.1 / 89.4 / 89.8; Mask Ambiguity 52.1 / 86.6 / 88.4 / 89.3 / 89.5; Bounding-Box Ambiguity 69.7 / 84.2 / 87.4 / 88.9 / 90.3; Table Count 80.3 / 85.5 / 86.9 / 88.6 / 89.3. The strongest CascadeTabNet result is therefore 90.3 mAP at 10k using Bounding-Box Ambiguity. The paper also highlights that Prediction Uncertainty reaches 89.4 mAP at 8k, which it interprets as a 20% reduction in annotation relative to random sampling.

For YOLOv9 on TableBank-Word, the reported mAP values are: Random 64.6 / 84.2 / 88.0 / 89.6 / 88.6; Prediction Uncertainty 67.3 / 86.3 / 89.2 / 89.2 / 90.1; Mask Ambiguity 66.4 / 84.6 / 89.4 / 89.6 / 90.5; Bounding-Box Ambiguity 68.2 / 87.9 / 90.2 / 91.7 / 90.2; Table Count 67.0 / 89.3 / 90.7 / 90.3 / 91.0. The strongest mid-budget result is 91.7 mAP at 8k using Bounding-Box Ambiguity, while the best 10k result is 91.0 using Table Count. The paper explicitly concludes that “Sampling based on table count provides a more significant performance boost for YOLOv9, whereas ambiguity-based sampling methods perform better with CascadeTabNet.”

Several caveats accompany these results. The paper does not provide a numeric full-supervision baseline for TableBank-Word, does not specify optimizer, learning rate, batch size, augmentation policy, image size, exact mAP protocol, or seed averaging, and does not fully define the active-learning round design. It also reports unstable low-budget behavior for some heuristics, especially Mask Ambiguity at 2k for CascadeTabNet (52.1 mAP) and Bounding-Box Ambiguity at 2k for CascadeTabNet (69.7). These omissions and instabilities limit strict reproducibility, but the budgeted comparisons against Random remain explicit.

The active-learning study therefore positions TableBank-Word as a challenging, high-variability benchmark where careful sample selection can improve mAP under a fixed annotation budget. The semi-supervised study positions it as a benchmark where improved pseudo-label assignment can yield 95.7 mAP with 30% labeled data. Taken together, these results indicate that TableBank-Word functions not only as a source of weakly supervised labels from Word documents, but also as a stress test for annotation-efficient table detection methods under substantial visual heterogeneity.