Research map / Large language models / Retrieval and knowledge
CABINET: Content Relevance-based Noise Reduction for Table Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CABINET, a table QA framework that softly weights table tokens by an unsupervised relevance scorer trained end-to-end with a BART-based QA model, plus a weakly supervised parsing-statement-driven cell highlighter, achieving SoTA on WikiTQ, FeTaQA and WikiSQL.
Problem
Table question answering is hurt by irrelevant table content acting as noise for LLMs, and the problem worsens for large tables; hard sub-table decomposition methods can discard relevant content irreversibly.
Approach
Instead of removing table content, multiply each table token embedding by a learned relevance score from an unsupervised relevance scorer (trained differentiably via answer loss plus clustering, centroid-separation and sparsity losses), and boost scores of cells highlighted by a parsing statement generated by a fine-tuned Flan-T5-xl and a cell highlighter.
Key results
- CABINET reaches 69.1% accuracy on WikiTQ versus 62.7% for OmniTab and 65.9% for DATER.
- CABINET achieves 40.5 S-BLEU on FeTaQA, above OmniTab's 34.9 and DATER's 30.9.
- On WikiSQL CABINET reaches 89.5% accuracy versus 88.8% for ReasTAP, and it shows smaller performance drops under row/cell perturbations than OmniTab (e.g. ~11.5% vs ~19% for row addition on WikiTQ).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CABINET's relevance weighting and cell predictor instead of the baseline method | Table QA models · WikiTQ test | accuracy: 65.9 -> 69.1 | Helped |
| Apply CABINET's relevance weighting and cell predictor instead of the baseline method | Table QA models · FeTaQA test | S-BLEU: 36.2 -> 40.5 | Helped |
| Apply CABINET's relevance weighting and cell predictor instead of the baseline method | Table QA models · WikiSQL test | accuracy: 88.8 -> 89.5 | Helped |
| Use TAPEX as the QA backbone and to initialize URS in CABINET | TAPEX · WikiTQ | accuracy: 55.5 -> 62.7 | Helped |
| Use TAPEX as the QA backbone and to initialize URS in CABINET | TAPEX · FeTaQA | S-BLEU: 34.6 -> 37.8 | Helped |
| Use TAPEX as the QA backbone and to initialize URS in CABINET | TAPEX · WikiSQL | accuracy: 86.4 -> 87.3 | Helped |
| Add the weakly supervised cell predictor to URS (ablation) | CABINET · WikiTQ | accuracy: 65.6 -> 69.1 | Helped |
| Add the weakly supervised cell predictor to URS (ablation) | CABINET · FeTaQA | S-BLEU: 35.8 -> 40.5 | Helped |
| Add the weakly supervised cell predictor to URS (ablation) | CABINET · WikiSQL | accuracy: 89.2 -> 89.2 | No clear effect |
| Use only the cell predictor and remove URS (ablation) | CABINET · WikiTQ | accuracy: 65.6 -> 37.6 | Hurt |
| Feed the parsing statement to the QA model instead of highlighting corresponding cells (ablation) | CABINET · FeTaQA | S-BLEU: 40.5 -> 34.9 | Hurt |
| Replace URS with BERT-based question-row similarity while retaining the cell highlighter (ablation) | CABINET · WikiTQ | accuracy: 69.1 -> 64.5 | Hurt |
Surprising findings
- Using only the cell-based relevance score (no unsupervised scorer) collapses performance (37.6 WikiTQ), and feeding the parsing statement directly to the QA model hurts FeTaQA performance relative to OmniTab's baseline.
Limitations
- Parsing statement generator relies on only ~300 manually annotated examples for bootstrapping.
- Cell-based highlighting only assigns scores on exact cell-content matches and performs poorly alone.
- Performance on WikiSQL gains are marginal over strong baselines and some ablations show no benefit from the cell predictor.
Experimental setup
Datasets: WikiTableQuestions FeTaQA WikiSQL ToTTo PASTA-style DeBERTa-V2 question clustering samples from WikiTQ
Benchmarks: WikiTQ FeTaQA WikiSQL
Metrics: accuracy Sacre-BLEU relative performance drop under perturbations
Baselines: OmniTab TAPEX ReasTAP TaCube FlanT5-xl DATER Binder LEVER StructGPT
Builds on: OmniTab BART Flan-T5 TAPEX DATER
Closest papers by meaning
- TableBench: A Comprehensive and Complex Benchmark for Table Question Answering (AAAI 2025 · similarity 0.89)
- NeedleInATable: Exploring Long-Context Capability of Large Language Models towards Long-Structured Tables (NeurIPS 2025 · similarity 0.86)
- Accurate and Regret-Aware Numerical Problem Solver for Tabular Question Answering (AAAI 2025 · similarity 0.86)
- Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · similarity 0.86)
- TrustUQA: A Trustful Framework for Unified Structured Data Question Answering (AAAI 2025 · similarity 0.86)
- Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.