💻 computer science

PubTables-v2: A new large-scale dataset for full-page and multi-page table extraction

The paper introduces PubTables-v2, a new large-scale dataset designed to advance full-page and multi-page table extraction by addressing the lack of annotated data and establishing baseline performance for vision-language models, particularly highlighting the challenge of multi-page table structure recognition.

Original authors: Brandon Smock, Valerie Faucon-Morin, Max Sokolov, Libin Liang, Tayyibah Khanam, Amrit Ramesh, Maury Courtland

Published 2026-03-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Brandon Smock, Valerie Faucon-Morin, Max Sokolov, Libin Liang, Tayyibah Khanam, Amrit Ramesh, Maury Courtland

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to digitize a massive library of scientific papers. You want to turn the messy, printed tables in these papers into clean, editable spreadsheets that computers can understand. This is the job of Table Extraction.

For a long time, computers were like clumsy librarians. They would look at a page, find a table, cut it out with scissors (cropping it), and then try to figure out the rows and columns. But this approach had two big problems:

It lost context: By cutting the table out, the computer forgot where it came from. Was that "Total" at the bottom a footer for the whole page, or just the end of the table?
It couldn't handle long stories: Some tables are so huge they spill over onto the next page, or even the next ten pages. The old "cut-and-paste" method just got confused and gave up.

Enter PubTables-v2, a new "super-library" created by researchers at Kensho Technologies to teach computers how to do this job better.

Here is the breakdown of what they did, using some everyday analogies:

1. The New "Textbook" (The Dataset)

Think of previous datasets as a collection of single, isolated flashcards. You'd show the computer one card with a table on it, and it had to guess the structure.

PubTables-v2 is like handing the computer an entire textbook.

The "Full Page" View: Instead of just showing the table, they show the whole page, including the title, the footnotes, and the surrounding text. This helps the computer understand the context.
The "Multi-Page" Challenge: This is the big innovation. They included 9,492 tables that stretch across multiple pages. Imagine a table that starts on page 1, continues on page 2, and finishes on page 5. Previous computers didn't know how to stitch these together. PubTables-v2 is the first massive dataset to teach them this skill.
The "Hard Mode" Tables: They specifically included tables that are incredibly long (30+ rows) or incredibly wide (12+ columns), which are the kinds of tables that usually break older software.

2. The "Students" (The Models)

The researchers tested several "students" (AI models) to see how well they could read these new textbooks.

The Specialized Students (VLMs): These are Vision-Language Models. Think of them as smart students who can read text and look at pictures at the same time.
The Result: When asked to read just a single cropped table, these students did okay. But when asked to read a whole page or a multi-page document, they struggled. They often missed tables, got confused by the layout, or couldn't tell where one table ended and another began.

3. The "Glue" Trick (Cross-Page Merging)

One of the biggest gaps found was that computers are terrible at realizing, "Hey, this table on Page 1 is the same one continuing on Page 2." They treat them as two separate, unrelated things.

The researchers invented a clever workaround, like a specialized glue gun:

They trained a simple "detective" (an image classifier) to look at two pages side-by-side.
The detective's only job is to ask: "Does the table at the bottom of Page 1 continue onto Page 2?"
If the detective says "Yes," the system uses "glue" to physically merge the two table outputs into one giant table.
The Result: This simple trick dramatically improved the computer's ability to handle long, multi-page documents. It turned a failing grade into a passing one.

4. Why This Matters

Think of PubTables-v2 as the "Olympics" for table extraction.

Before: We only tested swimmers in a small, calm pool (cropped tables).
Now: We are testing them in the open ocean with waves and currents (full pages and multi-page documents).

The paper shows that while our current AI is getting better at reading text, it still struggles with the complex layout of real-world documents. By providing this massive new dataset and showing that a simple "glue" step can fix multi-page issues, the researchers have given the AI community a clear roadmap for how to build the next generation of document readers.

In short: They built a harder, more realistic training ground for AI, discovered that current AI is bad at connecting the dots across pages, and found a simple trick to help it do so. Now, the data and code are open for everyone to use and improve upon.

1. Problem Statement

Table Extraction (TE) is a critical component of Visual Document Understanding (VDU). While traditional approaches successfully handle isolated, cropped tables, they face significant limitations in real-world scenarios:

Context Loss: Traditional pipelines detect tables first and then process them in isolation, losing crucial document-level context (e.g., captions, footers, surrounding text).
Pipeline Complexity: Multi-stage pipelines (detection $\rightarrow$ structure recognition) are complex to maintain and prone to error propagation.
Multi-Page Challenges: Existing datasets and models struggle with tables that span multiple pages or columns, a common occurrence in scientific literature.
Data Scarcity: There is a lack of large-scale, high-quality benchmarks for end-to-end table extraction at the page and document levels, particularly for multi-page tables. Previous attempts (e.g., combining PubLayNet and PubTables-1M) suffered from data leakage and small scale.

2. Methodology: PubTables-v2 Dataset

The authors introduce PubTables-v2, a large-scale dataset derived from over one million new scientific articles (2023–2025) from PubMed. The dataset construction leverages XML/HTML annotations provided by authors and aligns them with PDF text using sequence alignment.

Key Quality Control Measures:

Strict Alignment: Text in the PDF is compared against the original HTML annotation. A table is discarded if the mean normalized edit distance exceeds $\alpha = 0.02$ (stricter than PubTables-1M's 0.05) or if any cell exceeds $\beta = 0.2$ .
Consistency: If any table in an article fails quality control, the entire article is discarded to ensure no missing annotations.
Canonicalization: Table structures are canonicalized to improve labeling consistency and model performance.

Dataset Collections:
PubTables-v2 is organized into three distinct collections based on context granularity:

Cropped Tables (135k samples): Focuses exclusively on long (≥30 rows) and wide (≥12 columns) tables to challenge models with complex structures.
Single Pages (468k pages, 548k tables): Annotates full pages, including bounding boxes for tables, captions, footers, and hierarchical relationships (parent-child) between these elements.
Full Documents (9,172 documents, 24k tables): The first large-scale dataset for multi-page table recognition. It contains 9,492 multi-page tables, including 17 tables spanning 10+ pages and one spanning 13 pages. Every document contains at least one multi-part table.

3. Key Contributions

First Large-Scale Multi-Page Benchmark: PubTables-v2 is the first dataset to provide a substantial volume of multi-page tables (9,492 instances) for training and evaluating end-to-end document parsing.
Hierarchical Annotations: It introduces large-scale annotations connecting tables to their captions and footers with explicit hierarchical relationships, addressing a gap in document layout analysis.
New Evaluation Metrics: The authors generalize standard metrics (GriTS, TEDS, Exact Match) to handle set-to-set matching (predicting a set of tables for a single input) using the Hungarian algorithm, rather than the traditional one-to-one matching used for cropped tables.
Cross-Page Continuation Prediction: The paper proposes a novel task: predicting whether a table on page $N$ continues to page $N+1$ . They demonstrate that this can be solved with near-perfect accuracy using standard image classifiers.
Baseline Evaluations: They establish baselines for several specialized Vision-Language Models (VLMs) and a modified Table Transformer (POTATR) on these challenging tasks.

4. Experimental Results

The authors evaluated several models, including specialized VLMs (e.g., dots.ocr, GraniteDocling, Qwen2.5-VL) and a modified Table Transformer (POTATR).

Key Findings:

Cropped Tables: Models perform well on isolated tables. TATR-v1.2-Pub (fine-tuned) combined with docTR OCR achieved a GriTSCon of 0.980, significantly outperforming VLMs.
Single Pages: Performance drops when moving to full pages. dots.ocr was the best VLM (GriTSCon: 0.899), but still lagged behind models using direct document text extraction.
Full Documents (The Gap): Multi-page table recognition remains a major bottleneck.
- When processing full documents page-by-page, the best model (dots.ocr) achieved a GriTSCon of only 0.5768.
- When attempting to process full documents in a single pass (context window permitting), performance plummeted (e.g., Qwen2.5-VL-3b dropped to 0.0472), often due to token limits and infinite loops in generation.
Cross-Page Merging Solution:
- The authors trained image classifiers (ResNet-50, ViT-B-16) to predict if a table continues across pages. The ViT model achieved 0.991 F1.
- Impact: Integrating this classifier to merge tables across pages improved the dots.ocr GriTSCon score from 0.5768 to 0.6844, demonstrating a significant performance boost through a data-driven, modular approach.

5. Significance and Future Impact

Bridging the Gap: PubTables-v2 exposes the limitations of current VLMs in handling long-context, multi-page document structures, specifically the inability to naturally "stitch" tables across pages.
Modular Approach: The paper suggests that for complex document parsing, a hybrid approach (using specialized classifiers for structural decisions like page continuation, combined with VLMs for extraction) may be more effective than relying solely on end-to-end generative models.
Resource Availability: The dataset, code, and models are released to the community to foster further research in full-page and multi-page table extraction, a critical need for digitizing scientific literature and enterprise documents.

In conclusion, PubTables-v2 sets a new standard for table extraction research by moving beyond isolated table recognition to the complex, real-world challenge of extracting structured data from full, multi-page documents.