ECAI 2025

From Surface to Semantics: Semantic Structure Parsing for Table-Centric Document Analysis

A table-anchored framework for deep, context-aware document understanding

Existing PDF analysis stops at surface-level layout and OCR. DoTabler goes one layer deeper — using tables as anchors to recover their semantically associated text, enabling table-centric document parsing and domain-specific table retrieval at orders-of-magnitude lower latency than GPT-4o.

Xuan Li, Jialiang Dong^*, Raymond Wong^*

University of New South Wales, Sydney, Australia | ^*Corresponding authors

Paper PDF Poster Code

Figure 1: Overall workflow of DoTabler. Document Structure Preprocessing produces parsed blocks; Relation-Aware Annotation builds a table-text dataset; the Table-Text Association Model (TTAM) is trained on this data and powers two downstream capabilities — Table-Centric Semantic Parsing and Domain-Specific Table Retrieval.

Abstract

Documents are core carriers of information across finance, healthcare, and scientific research, with tables being the primary medium for structured data. Existing studies focus on surface-level tasks — layout analysis, table detection, data extraction — and stop short of deep semantic parsing of tables and their contextual associations. This limits advanced tasks such as cross-paragraph data interpretation and context-consistent analysis.

We propose DoTabler, a table-centric semantic document parsing framework designed to uncover deep semantic links between tables and their context. DoTabler leverages a custom-built dataset and domain-specific fine-tuning of pretrained models, integrating a complete parsing pipeline that identifies the context segments semantically tied to each table. Built on this semantic understanding, DoTabler delivers two core functionalities: table-centric document structure parsing and domain-specific table retrieval.

Evaluated on nearly 4,000 pages with over 1,000 tables from real-world PDFs, DoTabler achieves over 90% Precision and F1, outperforming GPT-4o, Gemini-2.0, and Claude-3.5 in table-context semantic analysis while running two orders of magnitude faster.

Two Use Cases

Figure 2a — Table-Centric Parsing. Given a PDF page, DoTabler identifies each table block and recovers the surrounding paragraphs that semantically describe or interpret it.

Figure 2b — Domain-Specific Table Retrieval. Given a natural-language query, DoTabler returns the tables in a document that are semantically relevant, together with their associated context.

Key Contributions

First Table-Centric Semantic Dataset

A new PDF semantic-level dataset modeled around table-centric structures, paired with the Table-Text Association Model (TTAM) for fine-grained table–context relation analysis.

DoTabler Framework

An end-to-end pipeline that combines preprocessing (segmentation, layout, OCR) with TTAM, enabling both document structure parsing and domain-specific table retrieval in a single framework.

Comprehensive Evaluation

Evaluated on nearly 4,000 pages with 1,000+ tables from arXiv and PubMed Central, beating GPT-4o, Gemini-2.0, and Claude-3.5 on accuracy and running orders of magnitude faster.

Method

DoTabler decomposes table-centric semantic parsing into a three-stage pipeline. Each stage is modular and replaceable, making the framework extensible to new pretrained encoders or document domains.

1. Document Structure Preprocessing

Each PDF is rendered to images, segmented into pages, and passed through a Faster R-CNN layout detector fine-tuned on PubLayNet. Tesseract OCR extracts text from Text, List, Title, and Table blocks.

2. Table-Text Association Model (TTAM)

A sentence-pair classifier built on a pretrained encoder (BERT / BART / RoBERTa). For each (table, text) pair, TTAM outputs a binary "related / unrelated" label, trained with cross-entropy over balanced positive/negative pairs.

3a. Table-Centric Semantic Parsing

For every Table block in a document, TTAM scores all candidate Text/List blocks and returns the semantically associated subset, yielding a structured (table, related_paragraphs) view of the document.

3b. Domain-Specific Table Retrieval

A fine-tuned RoBERTa cross-encoder jointly encodes (query, table) pairs and produces a relevance score; tables are ranked and the top-K returned with their context. Trained with margin-based ranking loss.

Dataset & Experimental Setup

No public dataset existed for document-level semantic structure analysis with table-centric anchors, so we built the first one. Documents were sampled from arXiv and PubMed Central; two senior annotators independently labeled table-text relations and reconciled disagreements via expert consensus.

227

Source PDFs (arXiv + PubMed Central)

3,952

Pages parsed and annotated

1,061

Table blocks · 1,624 Text blocks

Splits. TTAM evaluation: 3,248 balanced table-text pairs split 7:3 (2,273 train / 975 test). Retrieval evaluation: 200 query-table pairs filtered to 129 train / 53 test. Baselines. GPT-4o, Gemini-2.0 Flash, Claude-3.5 — all prompted with the same task instruction so the comparison is fair.

Results

Table-Text Relation Classification (TTAM)

90.01%

F1 on table–text linking (RoBERTa-based TTAM)

+10 pts

F1 over the strongest LLM baseline (Gemini-2.0)

+41 pts

F1 over GPT-4o (48.48 → 90.01)

Document-Level & Retrieval Performance

128 / 193

Documents fully correct (vs. 109–114 for LLM baselines)

71.7%

Retrieval Recall@1 (88.7% at K=3)

~270×

Faster than GPT-4o (0.0035s vs 0.80s avg)

Figure 3: Mean and median time overhead per (table, text) pair, batched across 10 evenly-split test groups. DoTabler stays flat near zero while the three LLM baselines fluctuate around 0.6–1.7s.

Why DoTabler wins. Decoder-only LLMs are tuned for fluent generation and adopt conservative decision strategies — they keep precision high but produce many false negatives, missing real table–text associations. RoBERTa's bidirectional encoder, fine-tuned on the new dataset, captures fine-grained contextual dependencies and is far less prone to hallucination, which is exactly what this task rewards.

Citation

@inproceedings{li2025dotabler, title={From Surface to Semantics: Semantic Structure Parsing for Table-Centric Document Analysis}, author={Li, Xuan and Dong, Jialiang and Wong, Raymond}, booktitle={Proceedings of the 28th European Conference on Artificial Intelligence (ECAI)}, year={2025} }

Acknowledgments: This work was conducted at the University of New South Wales. We thank our annotators, Xiao Chen and Kuang He, for their careful work on the table–text association dataset.