ICDAR 2026

Meaning Lies in Structure: Fine-Grained Table-Centric Document Semantic Parsing

Fine-grained table-text semantic alignment for reliable document understanding

Tables carry decision-critical quantitative evidence, but their meaning often lives in surrounding text. AITabler models table-text relationships as a five-level semantic hierarchy, integrating PDF structural reconstruction, layout-aware context retrieval, and LLM-based reasoning for table-centric document parsing and question answering.

Xuan Li, Mengfei Xiao, Jingtian Wei, Jialiang Dong, Raymond Wong^*

University of New South Wales, Australia | University of Technology Sydney, Australia | University of Wollongong, Australia | ^*Corresponding author

Paper PDF Poster Code

Figure 1: Overall workflow of AITabler. The system reconstructs PDF structure, retrieves layout-aware spatial context, performs fine-grained table-text semantic relevance grading, and uses the recognized associations to improve downstream table question answering.

Abstract

Tables encode quantitative evidence for financial reporting, risk assessment, and policy evaluation. In these settings, automated analysis depends on accurate semantic integration between tabular data and explanatory text; misalignment can distort interpretation and downstream decisions.

Existing PDF table research focuses mostly on visual structure recovery, while fine-grained semantic alignment across document components remains underexplored. We address this gap by formulating table-text semantic alignment as a five-level hierarchical framework and propose AITabler, an end-to-end system that integrates PDF structural reconstruction, layout-aware spatial context modeling, and LLM-based semantic reasoning.

We construct two benchmarks: a fine-grained relevance corpus with 29,138 annotated pairs and a Table-QA dataset with 6,000 questions, derived from over 3,000 PDF pages. Experiments show consistent improvements in relevance grading and substantial gains in downstream Table-QA.

Key Contributions

Five-Level Semantic Alignment

We model table-text relationships as a structured hierarchy, capturing caption reference, content alignment, topical match, contextual relevance, and unrelated context beyond coarse binary relevance.

End-to-End AITabler Pipeline

AITabler combines PDF structural reconstruction, spatial context retrieval, and LLM-based reasoning to recover table-text semantic associations at document scale.

Benchmarks for Parsing and QA

We release a fine-grained table-text relevance corpus and a purpose-built Table-QA dataset to support reproducible table-centric document analysis research.

Method

1. Document Structural Reconstruction

PDF pages are parsed into layout blocks so that tables, captions, and surrounding paragraphs can be represented as structured document components.

2. Spatial Context Retrieval

AITabler retrieves nearby candidate paragraphs using layout-aware spatial cues, reducing the search space before semantic grading.

3. Five-Level Relevance Grading

LLM-based reasoning assigns each table-text pair to one of five relevance levels, producing fine-grained associations instead of a single related/unrelated label.

4. Semantic-Enhanced Table-QA

The recognized associations provide high-value context for downstream table question answering, improving accuracy without indiscriminately adding noisy paragraphs.

Dataset & Results

29,138

Annotated table-text pairs

6,000

Table-QA questions

3,000+

PDF pages used to derive benchmarks

Experiments show that AITabler reaches over 98% accuracy on strongly related samples, exceeds 75% on weakly related cases, and achieves more than 70% accuracy on complex QA tasks, substantially outperforming coarse-grained retrieval baselines.

Citation

@inproceedings{li2026aitabler,
  title={Meaning Lies in Structure: Fine-Grained Table-Centric Document Semantic Parsing},
  author={Li, Xuan and Xiao, Mengfei and Wei, Jingtian and Dong, Jialiang and Wong, Raymond},
  booktitle={International Conference on Document Analysis and Recognition},
  year={2026}
}