Tables Decoded: DELTA for Structure, TARQA for Understanding
Jahanvi Rajput, Dhruv Kudale, Saikiran Kasturi, Utkarsh Verma, Ganesh Ramakrishnan
cs.CV · cs.LG
Quick verdict
Modular TATR+SPRINT+OCR pipeline (DELTA) emits OTSL; OTSL-tuned LLM (TARQA) gives strong TEDS-S and TabVQA gains, but full TEDS ~55 lags SOTA due to OCR.
Summary
The paper presents a modular pipeline for table understanding split into two components. DELTA reconstructs tables by doubly decoupling structure recognition: TATR predicts physical rows and columns, SPRINT predicts the logical cell arrangement, and EasyOCR fills in cell text; the resulting HTML is losslessly converted to the compact OTSL format. TARQA is LLaMA-3-8B-Instruct fine-tuned on OTSL-serialized tables for question answering. On structure recognition, DELTA reports TEDS-S of 98.2 (FinTabNet), 97.5 (PubTabNet), and 97.7 (PubTables-1M), comparable to specialized TSR systems, but full TEDS (55.9/54.4/54.8) falls far below end-to-end and PDF-parsing competitors; a ground-truth-content ablation attributes nearly the whole gap to OCR errors. On WTQ, fine-tuning on OTSL beats identical-config HTML (ANLS 56.5 vs 42.3) and plain-text (52.8) variants. DELTA+TARQA reaches 45.2 relieved accuracy on FinTabNetQA, above the best prior VLM (36.0), and ranks second on the new Hindi benchmark TORQUE (27.49 vs Qwen-2.5-VL's 47.40). Main limitations: the system assembles existing components with limited architectural novelty; headline comparisons mix evaluation protocols; no variance or significance reporting; FinTabNetQA has only 250 test samples; and TORQUE ground truth derives from LLM outputs with unclear manual-verification depth. Code, models, and benchmark are stated to be released.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Doubly decoupling physical TSR (TATR), logical TSR (SPRINT), and OCR (EasyOCR) yields table reconstruction with TEDS-S comparable to state of the art on FinTabNet, PubTabNet, and PubTables-1M.
- Fine-tuning an LLM on OTSL-serialized tables outperforms identical-config HTML and plain-text serializations (WTQ ANLS 56.5 vs 42.3 vs 52.8).
- DELTA+TARQA achieves 45.2 relieved accuracy on FinTabNetQA, surpassing all evaluated VLM baselines by 9.2 p.p. over the best (36.0).
- The pipeline transfers zero-shot to Hindi, ranking second on the new TORQUE benchmark (27.49) behind Qwen-2.5-VL-Instruct (47.40).
- With ground-truth content mapped to predicted structure, TEDS approaches 90+, showing OCR is the sole remaining bottleneck.
Strengths
Transparent bottleneck analysis
The GT-mapped ablation (Table 11: TEDS 91.2/87.8/92.5/83.5 with perfect content vs 54-64 with OCR) cleanly isolates OCR as the cause of low full TEDS, and Section 11.4 adds an engine comparison with latency (Tesseract 13.9s vs EasyOCR 4.3s per image).
Controlled representation ablation
Table 3 fine-tunes the same LLaMA-3-8B configuration on OTSL, HTML, and plain text; OTSL wins (ANLS 56.5 vs 42.3 vs 52.8; EM 54.0 vs 39.9 vs 50.2), directly supporting the core format claim rather than relying only on external baselines.
New multilingual benchmark
Competitive end-to-end TabVQA results
Reproducibility posture
Weaknesses
Limited architectural novelty
Full-reconstruction performance lags SOTA badly
Cross-protocol baseline comparisons
No statistical rigor
Reporting inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the DELTA table recognition pipeline?
DELTA ('Doubly dEcoupled tabLe reconsTruction Approach') recognizes tables by separating three stages: TATR detects physical rows and columns, SPRINT predicts the logical cell arrangement (spans, merged cells), and EasyOCR reads the text inside each cell. The combined HTML is converted to OTSL, a compact text format encoding structure and content. It reaches structure-only TEDS-S scores of roughly 97-98 on FinTabNet, PubTabNet, and PubTables-1M, though full scores including text (TEDS) are around 55 because of OCR errors.
Why does the paper use OTSL instead of HTML for table question answering?
OTSL (Optimised Table Structure Language) encodes a table far more compactly than HTML using a few special tokens per cell (<fcel>, <ecel>, <lcel>, <ucel>, <nl>). In a controlled experiment fine-tuning the same LLaMA-3-8B-Instruct model on WikiTableQuestions, the OTSL variant scored 56.5 ANLS versus 42.3 for the HTML variant and 52.8 for plain text, suggesting the compact, structure-explicit format helps the LLM answer questions.
How accurate is DELTA+TARQA on table question answering?
On the FinTabNetQA visual table QA benchmark the full pipeline scores 45.2 relieved accuracy, above the best prior vision-language baseline (SPHINX-MoE, 36.0). When given ground-truth table sequences instead of pipeline output, the same QA model reaches 69.2, so reconstruction and OCR quality account for a large share of the remaining gap.
What is the TORQUE benchmark?
TORQUE is a new Hindi (Devanagari) table understanding benchmark introduced in this paper: 210 tables (109 scanned, 101 digital-born, sourced from Indian government circulars and spiritual books) plus 422 manually verified question-answer pairs. It tests both table reconstruction and Hindi table QA; the paper's zero-shot pipeline scores 27.49 relieved accuracy, second to Qwen-2.5-VL-Instruct at 47.40.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are embedded as raw extracted text and could not be visually analyzed; qualitative examples were judged from text only.
- Some table values (e.g., the SmolDocling PubTables row) may be extraction artifacts of this text dump rather than errors in the original PDF; flagged as inconsistencies with that qualification.
- The GitHub repository's actual contents (code, models, TORQUE) could not be verified; the reproducibility rating relies on the paper's release statement plus the included configurations and pseudocode.
- Supplementary material is only partially embedded; edge-case behavior of the TATR-SPRINT fusion (Algorithm 2) could not be fully audited.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.