Improved Evidence Extraction and Metrics for Document Inconsistency Detection with LLMs

1. Problem Definition

The paper addresses the challenge of Document Inconsistency Detection, defined as identifying contradictions within a single input document. Unlike Natural Language Inference (NLI), which typically evaluates the relationship between a premise and a hypothesis, this task requires the model to find internal contradictions within a long, complex text.

The problem is formulated as a dual-output task:

1. Classification: Determine if an inconsistency exists ($y \in \{Yes, No\}$).
2. Evidence Extraction: If an inconsistency exists, identify the specific set of sentences ($E = \{s_1, ..., s_k\}$) that constitute the contradiction.

The Gap: While classification metrics are well-established, existing research lacks comprehensive metrics for evaluating evidence extraction. Furthermore, current datasets often contain documents with only a single inconsistent sentence, failing to capture the complexity of multiple contradictions found in real-world scenarios. Humans also struggle with this task in long documents, necessitating automated solutions.

2. Methodology

The authors propose a novel framework and a suite of new evaluation metrics to address these gaps.

A. New Evaluation Metrics

The paper argues that standard metrics (Accuracy, F1) are insufficient for evidence extraction. They introduce a comprehensive set of metrics:

• Evidence Hit (EH): A strict metric requiring the entire ground truth evidence set to be correctly identified ($E \subseteq \hat{E}$).
• Evidence Precision (EP): Measures the ratio of correctly identified evidence to the total number of sentences extracted. This penalizes models that output excessive false positives.
• Evidence Recall (ER): Measures the ratio of correctly identified evidence to the total ground truth evidence. This rewards partial extraction.
• Derived Rates: The authors define rates for the full dataset (e.g., EHR, EPR, ERR) and rates specifically for correctly classified documents (e.g., EHRC, EPRC, ERRC). They establish mathematical identities connecting these to True Positive Rate (TPR).

B. The Redact-and-Retry with Constrained Filtering (RnR+CF) Framework

To improve extraction performance, the authors propose an iterative algorithm:

1. Redact-and-Retry:
   • The LLM is prompted to identify inconsistencies in the document.
   • If inconsistencies are found, the identified sentences are redacted (removed) from the document.
   • The LLM is prompted again on the modified document to find remaining inconsistencies.
   • This loop continues until no new inconsistencies are found.
   • Intuition: Redacting found evidence reduces the search space and prevents the model from "fixating" on the same contradiction, allowing it to detect subsequent ones.
2. Constrained Filtering:
   • After the redact-and-retry loop, the accumulated evidence set is passed to a filter LLM call.
   • Constraint: The filter is explicitly instructed to output at least one sentence if the initial classification was "Yes."
   • Goal: The filter refines the evidence set by removing false positives while ensuring the system does not revert to a "No" classification (which would happen with an unconstrained filter if the evidence was deemed weak).

3. Key Contributions

1. Comprehensive Metrics: Introduction of Evidence Precision, Recall, and Hit metrics specifically designed to evaluate the quality of extracted evidence, not just the binary classification.
2. RnR+CF Framework: A novel iterative prompting strategy that significantly outperforms standard Direct Prompting (DP), Self-Consistency, and Self-Criticism in evidence extraction tasks.
3. ContraDocPaired Dataset: The creation of a new semi-synthetic dataset derived from the existing ContraDoc dataset.
   • Motivation: Existing datasets only contained documents with exactly one inconsistent sentence ($|E|=1$).
   • Innovation: ContraDocPaired combines pairs of documents to create instances with exactly two inconsistent sentences ($|E|=2$), enabling the evaluation of metrics like Evidence Hit vs. Evidence Recall, which are indistinguishable when $|E|=1$.

4. Experimental Results

The authors evaluated their approach using GPT-4o (strong model) and LLaMA3.2-90B (weaker model) on both ContraDoc and ContraDocPaired.

• Evidence Extraction Performance:
  • RnR+CF consistently achieved the best performance across all evidence metrics (EHR, EPR, ERR) for both models and datasets.
  • It successfully balanced high recall (finding all contradictions) with high precision (avoiding false positives).
  • In contrast, Self-Criticism (S-cri) showed lower True Positive Rates (TPR) and F1 scores, particularly on the weaker LLaMA model.
• Classification Performance:
  • There was no clear "winner" in binary classification accuracy among the methods; Direct Prompting and RnR+CF performed comparably.
  • This highlights that the primary value of RnR+CF lies in the quality of the evidence, not just the classification decision.
• Efficiency:
  • RnR+CF extracted significantly fewer sentences than the unconstrained Redact-and-Retry (RnR) or Self-Consistency approaches, reducing noise for the end-user.
  • The number of retries remained low (median < 3), indicating good scalability even when the evidence set size doubled in ContraDocPaired.
• Filter Analysis:
  • Ablation studies showed that the Constrained Filter (CF) was superior to the Unconstrained Filter (UF). The CF prevented the system from incorrectly flipping a "Yes" classification to "No" when the evidence was ambiguous, while still effectively pruning false positives.

5. Significance

• Practical Utility: The paper demonstrates that simply asking an LLM to "find contradictions" is insufficient. The proposed Redact-and-Retry mechanism mimics human cognitive processes (iteratively scanning and removing known issues) to improve detection in long contexts.
• Evaluation Standard: By introducing metrics that penalize both missing evidence (Recall) and hallucinated evidence (Precision), the paper sets a new standard for evaluating LLMs in document analysis tasks.
• Dataset Advancement: The release of ContraDocPaired fills a critical gap in the literature, allowing researchers to test models on multi-contradiction scenarios, which are more representative of real-world financial or legal documents.
• Trustworthiness: The ability to provide precise, verified evidence (rather than just a "Yes/No" label) is crucial for high-stakes domains like finance and law, where users must verify the source of a system's conclusion.

In conclusion, the paper establishes that iterative evidence extraction with constrained filtering is a robust method for document inconsistency detection, significantly outperforming standard prompting strategies in both accuracy and the quality of explanatory evidence.