PaperlayerView exact revision on arXiv

Source-linked AI summary

When TableQA Meets Noise: A Dual Denoising Framework for Complex Questions and Large-scale Tables

Shenghao Ye, Yu Guo, Dong Jin, Yikai Shen, Yunpeng Hou, Shuangwu Chen, Jian Yang, Xiaofeng Jiang

arXiv:2509.17680v2cs.CL

TL;DR

As TableQA questions grow more complex and tables larger, noisy data increases and degrades reasoning performance. EnoTab filters question semantic units using consistency and table usability, then uses an evidence tree to prune tables with rollback to protect answer-related data. EnoTab achieves substantial gains across multiple benchmarks for complex and large-scale TableQA tasks.

  • Problem

    As TableQA questions grow more complex and tables larger, noisy data increases and degrades reasoning performance.

    Source: §1 · p. 1

  • Method

    EnoTab filters question semantic units using consistency and table usability, then uses an evidence tree to prune tables with rollback to protect answer-related data.

    Source: §1 · p. 2, §2.4 · pp. 4–5

  • Results

    EnoTab achieves substantial gains across multiple benchmarks for complex and large-scale TableQA tasks.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    The authors conclude EnoTab is effective and general for complex and large-scale TableQA tasks.

    Source: §5 · p. 9

    EnoTab is evaluated for single-table question answering; its performance in multi-table settings remains unclear, and validation struggles with specialized formats such as "1-1".

    Source: Limitations · p. 9

Abstract

from arXiv · show

Table question answering (TableQA) is a fundamental task in natural language processing (NLP). The strong reasoning capabilities of large language models (LLMs) have brought significant advances in this field. However, as real-world applications involve increasingly complex questions and larger tables, substantial noisy data is introduced, which severely degrades reasoning performance. To address this challenge, we focus on improving two core capabilities: Relevance Filtering, which identifies and retains information truly relevant to reasoning, and Table Pruning, which reduces table size while preserving essential content. Based on these principles, we propose EnoTab, a dual denoising framework for complex questions and large-scale tables. Specifically, we first perform Evidence-based Question Denoising by decomposing the question into minimal semantic units and filtering out those irrelevant to answer reasoning based on consistency and usability criteria. Then, we propose Evidence Tree-guided Table Denoising, which constructs an explicit and transparent table pruning path to remove irrelevant data step by step. At each pruning step, we observe the intermediate state of the table and apply a post-order node rollback mechanism to handle abnormal table states, ultimately producing a highly reliable sub-table for final answer reasoning. Finally, extensive experiments show that EnoTab achieves outstanding performance on TableQA tasks with complex questions and large-scale tables, confirming its effectiveness.

1 Introduction

Noisy data in complex questions and large tables degrades TableQA reasoning, while existing approaches struggle to filter irrelevant question content and prune tables without losing answer-related data. EnoTab pairs question denoising with observable, rollback-capable table pruning; experiments report outstanding performance on complex-question and large-table tasks.

1 Introduction

  • Complex questions and large tables introduce noisy data that degrades reasoning, while table pruning can lose answer data and require a full redo.Source: §1 · pp. 1–2
  • Effective reasoning under noise requires filtering spurious question correlations and pruning irrelevant table rows while preserving answer-relevant data.The motivating example reduces a 260-row table to three relevant rows.Source: §1 · p. 2
  • EnoTab assesses minimal question units for consistency and table usability, then uses observable evidence-tree pruning steps and post-order rollback to protect answer data.Source: §1 · pp. 2–3
  • Extensive experiments show EnoTab achieves outstanding performance on TableQA tasks involving complex questions and large-scale tables.Source: §1 · p. 3

2 EnoTab

EnoTab addresses irrelevant question units and table content by filtering both before answer reasoning. It combines evidence-based question denoising with an explicit tree-guided pruning path and rollback.

2.1 Problem Definition

  • The task targets irrelevant question units and table cells, producing retained evidence and a subtable before answer reasoning.The stated goal is to enable more efficient and accurate answer derivation.Source: §2.1 · p. 3

2.2 Model Overview

  • EnoTab couples EQD, which removes irrelevant question units, with ETD, which prunes irrelevant table values; answer generation uses the retained evidence and resulting subtable.The framework presents the pruning path as observable, with post-order rollback intended to prevent loss of answer-related data.Source: §2.2 · p. 3

2.3 Evidence-based Question Denoising

  • EQD decomposes complex questions into single-unit evidence, retaining evidence only when it is consistent across generations and grounded in table data.Unlike decomposition for planning reasoning paths, this decomposition extracts individual semantic units for subsequent relevance assessment.Source: §2.3 · pp. 3–4

2.4 Evidence Tree-guided Table Denoising

  • ETD represents evidence as leaves and AND/OR relations as internal nodes, merging evidence-filtered subtables through a binary tree.When pruning produces an empty subtable, And2Or replaces AND with OR; a verifier checks completeness and can roll back to a previous subtable, with the full table used after two failed attempts.Source: §2.4 · pp. 4–5

3 Experiments

Experiments compare EnoTab with established methods across standard and large-scale TableQA datasets, and examine complex questions and framework behavior. EnoTab outperforms baselines on large-scale datasets and across WikiTQ difficulty levels, while analyses assess its components, adaptability, and pruning.

3.1 Experimental Setup

  • The evaluation covers WikiTQ, TabFact, and two Spider-based large-table datasets, with comparisons against generic, decomposition-based, and pruning-based methods.The study uses exact match accuracy for STQA-N, STQA-L, and WikiTQ, and binary classification accuracy for TabFact.Source: §3.1 · p. 5

3.2 Main Results

  • EnoTab significantly outperforms all baselines on STQA-N and STQA-L, with greater advantages on STQA-L.The authors attribute STQA-L’s greater pruning demands to its naturally large tables, complex structures, and data formats.Source: Table 1 · p. 6, §3.2 · p. 6
  • On WikiTQ and TabFact, EnoTab consistently outperforms existing methods, extending its reported performance to standard TableQA benchmarks.Evaluations use GPT-4o-mini.Source: §3.2 · p. 6, Table 2 · p. 7
  • EnoTab achieves higher accuracy than Chain-of-Table and Dater at every WikiTQ difficulty level, including Extra Hard.The analysis also finds a positive correlation between question complexity and the number of evidences generated by EnoTab.Source: §3.2 · p. 6, Figure 4 · p. 7, Figure 3 · p. 7

3.3 Experimental Analysis

  • Disabling any tested module reduces performance on STQA-N and STQA-L; removing either assessment module causes the most noticeable drop.The ablation examines Consistency Assessment, Usability Assessment, And2Or Operation, and Table Verifier.Source: Table 3 · p. 7, §3.3 · p. 7
  • Across closed- and open-source models on WikiTQ, EnoTab maintains stable performance while End-to-End QA drops when switching to open-source models.The authors attribute EnoTab’s robustness to relevance filtering and table pruning, which reduce reasoning burden.Source: §3.3 · pp. 7–8, Figure 5 · p. 8
  • On large-scale datasets, EnoTab substantially compresses tables; on WikiTQ and TabFact, it avoids removing answer-related data despite less noise.The analysis reports average token counts and compression rates before and after pruning.Source: §3.3 · p. 8

4 Related Work

Prior TableQA research has explored question decomposition and table pruning. The paper contrasts EnoTab’s fine-grained question assessment and observable pruning steps with limitations it identifies in those approaches.

Related Work

  • Existing work in these directions aims to simplify reasoning through intermediate steps or smaller subtables.The paper frames both as prominent directions in TableQA research.Source: §4 · p. 8

Question Decomposition

  • Question decomposition methods can rely heavily on intermediate-step correctness, while spurious correlations in complex questions may derail reasoning.EnoTab uses decomposition to assess semantic units before reasoning rather than primarily to construct a reasoning path.Source: §4 · p. 8

Table Pruning

  • Program-based table pruning can be opaque, making errors harder to detect and potentially risking the loss of answer-related data.EnoTab instead makes each pruning step observable and verifiable so erroneous states can be detected and corrected.Source: §4 · p. 8

5 Conclusion

The conclusion identifies limited relevance filtering and table pruning as bottlenecks for complex questions and noisy large tables. It presents EnoTab’s dual-denoising approach and reports gains on those TableQA tasks.

5 Conclusion

  • The paper identifies limited relevance filtering and table pruning as bottlenecks when TableQA handles complex questions and noisy large tables.The conclusion frames these as the sources of the identified bottlenecks.Source: §5 · p. 8
  • EnoTab decomposes questions into semantic units, grounds them in tables, and builds explicit reasoning paths with rollback to preserve answer-related information.The conclusion characterizes this as a dual-denoising framework.Source: §5 · p. 8
  • Experiments across multiple benchmarks show substantial gains on complex-question and large-scale TableQA tasks.The conclusion describes the approach as effective and general.Source: §5 · pp. 8–9

Limitations

The reported evaluation is in single-table question answering; multi-table performance and validation on some specialized formats remain limitations.

Limitations

  • Evaluation currently covers single-table question answering, so performance in multi-table settings remains unclear.The paper says the method can handle multiple tables but leaves its performance there for future assessment.Source: Limitations · p. 9
  • The validation method struggles with certain specialized table formats, such as 1-1, where the numbers represent the count of wins and losses.The authors aim to address this limitation in future work.Source: Limitations · p. 9
  • Future work will assess EnoTab in multi-table scenarios and address the stated validation difficulty.These are the two limitations named in the paper.Source: Limitations · p. 9

Ethics Statement

The authors acknowledge ethical concerns associated with the large language models used and encourage critical evaluation of EnoTab’s outputs before downstream use.

Ethics Statement

  • GPT-4o and GPT-4o-mini may inherit risks of inaccurate or harmful content, so users are urged to evaluate EnoTab’s outputs before downstream use.The study’s datasets are publicly available through peer-reviewed publications; it also uses open-source models in accordance with their usage policies and license agreements.Source: Ethics Statement · p. 9
  • The authors acknowledge the contributions of open-source models, including Qwen-1.5 and Llama-2, and state that they follow their usage policies and licenses.Source: Ethics Statement · p. 9

A Additional Implementation Details of EnoTab

The appendix details EnoTab’s retrieval, evidence assessment, and tree-guided pruning procedures to clarify the pipeline and support reproducibility.

A.1 Two-Stage Retrieval Process

  • Keyword filtering followed by embedding-based reranking selects representative table rows for evidence generation, avoiding use of the full table.The reranking score combines semantic similarity and lexical similarity based on edit distance.Source: §A.1 · p. 12
  • The retrieval process first uses question keywords and LSH to shrink the candidate rows, then ranks candidates by semantic and lexical similarity.In practice, k = 10, λ = 0.7, and C = min(256, ⌈0.1N⌉); GPT-4o-mini extracts keywords and bge-large-en-v1.5 is the default encoder.Source: §A.1 · p. 12

A.2 Consistency Assessment of Evidence

  • The consistency-assessment algorithm groups evidence by area and action, extracts area-condition-action triples, computes consistency, and adds qualifying evidence to the candidate set.Its stated inputs are evidence sets from n rounds, a threshold α, and a semantic discriminator; its output is candidate evidence Ec.Source: §A.2 · p. 12, §8 · p. 12

A.3 Usability Assessment of Evidence

  • Usability assessment checks whether evidence can be grounded in the table by applying its condition to the target column and retaining it only when the resulting sub-table is non-empty.The process normalizes the condition and invokes an action-specific API, such as string matching, numeric comparison, or date evaluation.Source: §A.3 · pp. 12–13

A.4 Detailed Description of the Evidence Tree

  • The Evidence Tree mirrors the question’s logic and is executed in post-order, with leaf nodes filtering the table and internal AND/OR nodes merging child subtables.The tree is illustrated as having four leaves and three internal nodes.Source: §A.4 · p. 13, Figure 7 · p. 13
  • Observable intermediate subtables allow abnormal states such as empty tables to be detected and support rollback, making pruning decisions inspectable and the trajectory auditable.Leaf-level filtering occurs before internal merging.Source: §A.4 · p. 13

A.5 Rollback Mechanism: Analysis of And2Or

And2Or relaxes an AND node that produces an empty subtable, preserving potentially useful records while accepting possible loss of precision.

A.5 Rollback Mechanism: Analysis of And2Or

  • When an AND node yields an empty subtable, And2Or replaces AND with OR, producing a superset that may contain redundant or irrelevant rows but preserves answer-critical content.The authors prioritize recall because later steps can filter irrelevant rows, whereas missing evidence cannot be recovered.Source: §A.5 · p. 13
  • The authors aim to improve discrimination at conjunction nodes so the method can rely less on And2Or while retaining it as a safeguard.Source: §A.5 · p. 14

A.6 Details of Dataset Construction

The appendix describes semiautomatic construction and manual verification of STQA-L and noisy STQA-N, and reports stable results across repeated runs.

A.6 Details of Dataset Construction

  • More than 80% of generated QA pairs were accepted without edits, while fewer than 20% required manual refinement after verification.Questions are generated from sampled cells, rows, or subtables; the sampled region is not directly used as the final answer.Source: §A.6 · p. 14
  • STQA-N expands tables shorter than 4,096 tokens and perturbs 15% of non-answer cells with type-specific noise while preserving answer correctness.String values are replaced with semantically close alternatives and numeric values with values outside the gold range; all pairs are manually verified.Source: §A.6 · p. 14
  • Answer errors are manually corrected, whereas ill-posed, ambiguous, or table-inconsistent questions are discarded.Source: §A.6 · p. 14

B.1 Hyper-parameter Settings

  • Across datasets, repeated experiments had standard deviation within 1% absolute accuracy, supporting stability of the reported results.Open-source models were run three times and closed-source APIs twice.Source: §B.1 · pp. 14–15

B.2 Error Analysis

The analysis distinguishes common Text2SQL pruning errors and explains how EnoTab handles them. Under injected table noise, EnoTab maintains stable performance while End-to-End QA declines.

B.2 Error Analysis

  • EnoTab independently verifies and executes evidence units, filtering invalid evidence to reduce BinderException and recognizing cases that exceed its table-reasoning capability to avoid ParserException by design.This contrasts with pruning through a single SQL query, which can fail when a condition is invalid or exceeds SQL’s expressive capacity.Source: §B.2 · p. 15
  • EnoTab remains limited on structurally complex tables whose compound cell values require careful extraction to preserve answer-related data.Adding this extraction ability could improve pruning, but also increases the risk of losing critical data.Source: §B.2 · p. 15
  • BinderException is the most frequent Text2SQL pruning error, ParserException is second, and IndexError involves access to out-of-range rows.The error analysis distinguishes semantic reference errors, SQL syntax errors, and row-index errors.Source: Figure 9 · p. 15, §B.2 · p. 15

B.3 Noisy Content

  • End-to-End QA drops significantly under injected noise, while EnoTab maintains stable performance, primarily through table pruning.The comparison evaluates both methods on clean and noisy WikiTQ development sets.Source: Figure 10 · p. 16, §B.3 · p. 16

B.4 Experiments of EnoTab on FetaQA

On FeTaQA, EnoTab leads existing methods on ROUGE-1/2/L but slightly trails Chain-of-Table on BLEU. The accompanying analyses relate evidence counts to question difficulty and describe the prompts, evidence examples, and related work.

B.4 Experiments of EnoTab on FetaQA

  • EnoTab consistently outperforms existing state-of-the-art methods on FeTaQA ROUGE-1/2/L, while slightly underperforming Chain-of-Table on BLEU.The paper attributes the BLEU difference to the lack of explicit optimization for surface phrasing or sequence alignment; manual inspection finds the answers accurate and complete in content.Source: §B.4 · p. 16
  • The FeTaQA BLEU shortfall does not reflect semantic degradation according to manual inspection, though EnoTab does not explicitly optimize surface phrasing or sequence alignment.BLEU is described as sensitive to word order and n-gram precision.Source: §B.4 · p. 16

B.5 Effectiveness Analysis of Evidence Generation

  • Across WikiTQ and TabFact, harder questions generally require more evidence pieces, and the number of samples with higher evidence counts rises with difficulty.The evidence-count distribution closely mirrors the difficulty-level distribution, with the same trend reported for both benchmarks.Source: §B.5 · p. 16

C.1 Prompt of Dataset Construction

C.2 Prompt of EnoTab

C.3 Example of Evidence

D Additional Related Work

  • Related work includes table-focused fine-tuning and methods combining pruning with reasoning plans; EnoTab separates pruning and planning and adds verifiable checkpoints.The paper characterizes some prior planning and pruning approaches as black boxes that make errors difficult to trace or correct.Source: D Additional Related Work · p. 17

Related papers

Loading 2509.17680v2…