Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improved Evidence Extraction for Document Inconsistency Detection with LLMs

Nelvin Tan, Yaowen Zhang, James Asikin Cheung, Fusheng Liu, Yu-Ching Shih, Dong Yang

TL;DR

This work tackles the problem of detecting inconsistencies within a single document using large language models, with a primary emphasis on extracting supporting evidence rather than merely classifying. It introduces a redact-and-retry framework, augmented by a constrained filtering step, to improve evidence hit and precision across multiple LLMs while keeping the evidence footprint compact. The authors formalize a suite of evidence-extraction metrics (EHR, EHRC, EPR, EPRC, ERR, ERRC, AECR) and demonstrate that the redact-and-retry approach, particularly with a constrained filter, outperforms direct prompting on key metrics. The results on the ContraDoc dataset show that RnR+CF offers a favorable balance of higher evidence quality and controlled evidence size, suggesting practical utility for automated inspection and correction of internal document inconsistencies.

Abstract

Large language models (LLMs) are becoming useful in many domains due to their impressive abilities that arise from large training datasets and large model sizes. However, research on LLM-based approaches to document inconsistency detection is relatively limited. There are two key aspects of document inconsistency detection: (i) classification of whether there exists any inconsistency, and (ii) providing evidence of the inconsistent sentences. We focus on the latter, and introduce new comprehensive evidence-extraction metrics and a redact-and-retry framework with constrained filtering that substantially improves LLM-based document inconsistency detection over direct prompting. We back our claims with promising experimental results.

Improved Evidence Extraction for Document Inconsistency Detection with LLMs

TL;DR

This work tackles the problem of detecting inconsistencies within a single document using large language models, with a primary emphasis on extracting supporting evidence rather than merely classifying. It introduces a redact-and-retry framework, augmented by a constrained filtering step, to improve evidence hit and precision across multiple LLMs while keeping the evidence footprint compact. The authors formalize a suite of evidence-extraction metrics (EHR, EHRC, EPR, EPRC, ERR, ERRC, AECR) and demonstrate that the redact-and-retry approach, particularly with a constrained filter, outperforms direct prompting on key metrics. The results on the ContraDoc dataset show that RnR+CF offers a favorable balance of higher evidence quality and controlled evidence size, suggesting practical utility for automated inspection and correction of internal document inconsistencies.

Abstract

Large language models (LLMs) are becoming useful in many domains due to their impressive abilities that arise from large training datasets and large model sizes. However, research on LLM-based approaches to document inconsistency detection is relatively limited. There are two key aspects of document inconsistency detection: (i) classification of whether there exists any inconsistency, and (ii) providing evidence of the inconsistent sentences. We focus on the latter, and introduce new comprehensive evidence-extraction metrics and a redact-and-retry framework with constrained filtering that substantially improves LLM-based document inconsistency detection over direct prompting. We back our claims with promising experimental results.
Paper Structure (43 sections, 5 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 43 sections, 5 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Motivation.
  3. Problem setup.
  4. Related Work
  5. Sentence-level detection.
  6. Document-level detection.
  7. Applications.
  8. Main contributions.
  9. Metrics for Evidence Extraction
  10. Evidence hit rate (EHR) and evidence hit rate when correct (EHRC).
  11. Evidence precision rate (EPR) and evidence precision rate when correct (EPRC).
  12. Evidence recall rate (ERR) and evidence overlap rate when correct (ERRC).
  13. Average evidence coverage ratio (AECR).
  14. Remark.
  15. Research question.
  16. ...and 28 more sections

Figures (6)

  • Figure 1: Plots of EHR vs. EPR and EHRC vs. EPRC.
  • Figure 2: Radar charts for different approaches and LLMs.
  • Figure 3: Algorithm \ref{['alg:RnR']} visualized.
  • Figure 4: Algorithm \ref{['alg:RnR']} with LLM filter call visualized.
  • Figure 5: Boxplot for number of retries for RnR. 'Pos' refers to datapoints $i$ such that truth $y_i=\text{Yes}$.'Neg' refers to datapoints $i$ such that truth $y_i=\text{No}$.
  • ...and 1 more figures