TEAR: Table Extraction with Attribute Recommendation from Texts via Large Language Models
Tong Li, Shuye Ding, Jiachuan Wang, Yongqi Zhang, Shuangyin Li, Lei Chen, Bo Li
cs.DB · cs.AI
Quick verdict
LLM framework TEAR adds text-driven attribute recommendation to table extraction on natural texts; solid benchmark and gains, but significance testing is thin.
Summary
TEAR targets table extraction (text-to-table) from "naturally occurring texts" — news reports and dialogues lacking the structural cues of specialized documents — and introduces a companion attribute recommendation (AR) task that discovers schema attributes missing from a user's heuristic schema. The framework has two workflows: the Table Extraction Workflow fine-tunes a BART-Large surrogate to produce fuzzy header/value previews, then retrieves demonstrations from a 300-sample labeled pool using three utilities (header overlap, value chrF similarity, dense semantic similarity); the Attribute Recommendation Workflow frames discovery as LLM table expansion, deduplicates proposals via Vendi-Score diversity filtering plus LLM merge judgments, and ranks candidates by a Schema Coherence Score (token probability given the schema). The authors contribute a 3,375-pair benchmark (Incidents 1,369 and Weather 2,006 from CACAPO; Conversation 1,000 from MultiWOZ 2.4), an index-free structured-F1 metric, and experiments with Llama-3-8B/70B and Qwen2.5-14B showing TEAR best in every Table 2 cell and on 48/54 AR comparisons. Limitations: no significance tests despite some sub-point margins, a likely decoding-parameter typo (temperature 5), reliance on per-dataset fine-tuned surrogates, benchmark provenance partially inheriting CACAPO annotations, and code not yet released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- TEAR is the first framework to support automated, text-driven attribute recommendation for exploratory schema design in table extraction.
- Preview-guided demonstration retrieval (the Proactive Demonstration Module) beats static and dense-similarity-only example selection for extraction from naturally occurring texts.
- TEAR achieves state-of-the-art extraction, best in every reported cell of Table 2 across three datasets, three backbones, and two metrics under three string similarities.
- The attribute recommendation workflow is best on 48 of 54 comparisons, and user-accepted recommended attributes further improve extraction (TEAR*) in exploratory settings.
Strengths
New task and benchmark with a real annotation effort
The Incidents (1,369) and Weather (2,006) sets are manually annotated from CACAPO texts (82–88% pairwise agreement, A.6), plus a 1,000-text Conversation stress test from MultiWOZ 2.4; three exploratory levels (20/35/50% hidden low-frequency attributes) operationalize the new AR task.
Index-free evaluation metric validated against humans
structured-F1 (§5.2.1, Eq. 21–22) compares tables as sets of header-to-value dictionaries without logical indices; human validation on 200 minimally perturbed predictions shows 90/95/92.5% directional agreement (EM/chrf/BS, §6.6.5).
Consistent, substantial extraction gains
Economical, well-analyzed dedup machinery
Honest task-framing for AR
Weaknesses
Statistical rigor does not support blanket superiority claims
Pipeline complexity sits awkwardly with the data-scarcity motivation
Benchmark provenance weakens the 'naturally occurring' framing
AR baselines do not isolate the discovery contribution
Artifacts not yet available
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TEAR (arXiv 2609.15205) targets table extraction (text-to-table) from "naturally occurring texts" — news reports and dialogues lacking the structural cues of specialized documents — and introduces a companion task, attribute recommendation (AR): automatically discovering attributes absent from a user's heuristic schema (§1, Challenges 1–2). Claimed contributions: the first text-driven AR framework; a Proactive Demonstration Module for extraction; a three-part AR workflow (Discovery Mechanism, Hybrid Integration Strategy, Schema Coherence Score); and a 3,375-pair benchmark with a new index-free metric. This is a method paper with a substantial benchmark component, evaluated with Llama-3-8B/70B-Instruct and Qwen2.5-14B-Instruct.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 TEAR and what does it do?
TEAR is a framework that (1) extracts tables from free-form texts such as news reports and dialogues, and (2) recommends new attributes the user's initial schema is missing. Its extraction workflow generates fuzzy header/value previews with a small fine-tuned model, retrieves the most useful demonstration examples per input, and prompts a backbone LLM; its recommendation workflow asks the LLM to expand the extracted table with new columns, deduplicates proposals, and ranks them by how coherently they extend the known schema.
What data does the TEAR paper use?
It uses 3,375 text-table pairs: Incidents (1,369 gun-violence news reports) and Weather (2,006 forecasts), both built on texts from the public CACAPO corpus with new manual table annotations, plus Conversation (1,000 dialogues adapted from MultiWOZ 2.4). The authors state the table annotations will be released.
What is structured-F1 in TEAR?
It is a new evaluation metric that compares tables as sets of records, where each record is a header-to-value dictionary, so no logical index column is needed — fitting texts (like news) that provide no natural row keys. Header names are aligned by similarity, then value similarities are aggregated. On 200 manually checked predictions, its preferred direction matched human judgment 90–95% of the time depending on the string similarity used.
Is TEAR's code or data available?
Not at review time. The paper links to a repository but a footnote states it is 'being prepared' for the camera-ready version; the benchmark table annotations are promised for release. Prompts, models (Llama-3-8B/70B, Qwen2.5-14B, BART-Large, SFR-Embedding-Mistral) and most hyperparameters are documented in the paper and appendix.
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 4–18 are vector plots whose numeric content is largely unrecoverable from extracted text; AR results (Fig. 4, 11) and some ablation curves rest on the authors' textual assertions only.
- Code link is a placeholder; footnote 7 says the repository is 'being prepared', so availability could not be verified.
- Digit runs embedded from Figures 7–10 lack labels in the extraction; mappings to metrics were inferred from surrounding axis labels and may be imperfect.
- The paper uses the ACM placeholder template ('Conference acronym XX'), suggesting an under-review author-kit version; some details may change for the camera-ready.
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.