PaperlayerView exact revision on arXiv

Source-linked AI summary

Fast and Accurate Factual Inconsistency Detection Over Long Documents

Barrett Martin Lattimer, Patrick Chen, Xinyuan Zhang, Yi Yang

arXiv:2310.13189v2cs.CLcs.AI

TL;DR

Long-document factual inconsistency detection remains difficult because existing approaches often struggle with long inputs and limited calibration or efficiency. SCALE uses NLI with large source chunks and chunk-based retrieval to detect inconsistencies and explain decisions. It achieves strong results across standard benchmarks and long-form dialogue evaluation while improving efficiency and retrieval capability.

  • Problem

    Existing factual inconsistency detectors struggle with long inputs, while prior benchmarks predominantly use short documents.

    Source: §1 · pp. 1–2, §3.1 · p. 3

  • Method

    SCALE uses an NLI-based detector with large source chunks and a chunk-enabled retrieval strategy for relevant source sentences.

    Source: §1 · pp. 1–2, §3.1 · p. 3

  • Results

    SCALE achieves state-of-the-art results across four NLG tasks and 11 TRUE datasets, and significantly outperforms all models other than GPT-4 on ScreenEval.

    Source: §7 · p. 9

  • Takeaways & Limitations

    SCALE provides interpretable calibrated scores and faster, more cost-effective long-dialogue evaluation than GPT-4, while improving relevant-utterance retrieval recall.

    Source: §7 · p. 9

    Using only the “Yes” and “No” logits may lose information carried by similar or other logits and reduce accuracy.

    Source: §8 · p. 9

Abstract

from arXiv · show

Generative AI models exhibit remarkable potential; however, hallucinations across various tasks present a significant challenge, particularly for longer inputs that current approaches struggle to address effectively. We introduce SCALE (Source Chunking Approach for Large-scale inconsistency Evaluation), a task-agnostic model for detecting factual inconsistencies using a novel chunking strategy. Specifically, SCALE is a Natural Language Inference (NLI) based model that uses large text chunks to condition over long texts. This approach achieves state-of-the-art performance in factual inconsistency detection for diverse tasks and long inputs. Additionally, we leverage the chunking mechanism and employ a novel algorithm to explain SCALE's decisions through relevant source sentence retrieval. Our evaluations reveal that SCALE outperforms existing methods on both standard benchmarks and a new long-form dialogue dataset ScreenEval we constructed. Moreover, SCALE surpasses competitive systems in efficiency and model explanation evaluations. We have released our code and data publicly to GitHub.

1 Introduction

SCALE addresses the difficulty of detecting factual inconsistencies in long documents by combining NLI with source chunking and relevant-text retrieval. It achieves broad benchmark performance and is evaluated on the long-dialogue ScreenEval dataset.

  • Prior methods focused mainly on short documents and sentence-level matching, creating speed and calibration challenges for long-document online use.The growth of LLM context sizes increases the need for efficient long-document inconsistency detection.Source: §1 · p. 1
  • SCALE combines an NLI-based detector with a chunking mechanism and source-text retrieval to support efficient factual inconsistency detection and explanation.The method is designed for diverse NLG tasks and online use.Source: §1 · pp. 1–2
  • ScreenEval contains 52 long-form dialogues averaging over 6,000 tokens, targeting factual inconsistency in generated summaries.The dataset includes summaries generated by humans, Longformer, and GPT-4.Source: §1 · p. 2
  • On ScreenEval, SCALE is evaluated for accuracy, speed, and model explanation against strong competitors.The experiments report that SCALE surpasses strong competitors in the majority of tests.Source: §1 · p. 2

2 Related Work

Prior inconsistency-detection work includes NLI- and QA-based approaches, but long-document evaluation remains limited. SCALE uses larger premise chunks, while ScreenEval targets long-form dialogue factual inconsistency.

  • SCALE uses larger NLI premise chunks, which the authors report improves efficiency and outperforms sentence decomposition.Earlier systems such as SeNtLI and SummaC rely on sentence decomposition, while SWING reports mixed outcomes.Source: §2 · p. 2
  • QA-based methods generate questions from summaries and assess consistency through answer overlap; QuestEval and QAFactEval are considered competitive examples.Source: §2 · p. 2
  • Existing inconsistency-detection methods largely target short documents, while long-document datasets have limited domain or evaluation scope.ContractNLI focuses on legal documents, and LongEval emphasizes human evaluation without dialogues.Source: §2 · p. 2
  • ScreenEval is presented as the first dataset for evaluating factual inconsistency in long-form dialogues.Source: §2 · p. 2

3 SCALE

SCALE resolves NLI granularity and efficiency problems by comparing generated sentences with larger source chunks. It also uses chunk-based search to retrieve supporting source sentences with logarithmic model-call complexity.

3.1 Chunking Mechanism for NLI based Model

  • The model defines factual consistency as the generated facts being a subset of the source facts.Source: §3.1 · p. 3
  • Sentence-level pairwise NLI can lose cross-sentence context and requires |D| · |G| model runs for one generated sentence.Multiple source sentences may jointly support a single claim.Source: §3.1 · p. 3
  • SCALE partitions the source document into larger chunks while splitting generated text into sentences, improving context capture and reducing model runs.Chunks can have arbitrary lengths subject to memory requirements.Source: §3.1 · pp. 3–4
  • SCALE computes entailment between each source chunk and generated sentence using Flan-T5 prompts, then aggregates the chunk results into a factual-consistency score.The prompt treats the source chunk as premise and the generated sentence as hypothesis.Source: §3.1 · p. 4

3.2 Model Explanation via Relevant Source Text Retrieval

  • For explanations, SCALE greedily descends a binary search tree of source chunks and selects the highest-scoring chunk until reaching a supporting sentence.This retrieves relevant source text in context for a target hypothesis.Source: §3.2 · p. 4, Figure 2 · p. 4
  • Retrieval requires O(log(n)) model calls instead of O(n) sentence-level calls for a source with n sentences, although larger split factors may be needed under GPU-memory constraints.Source: §3.2 · p. 4

4 ScreenEval Dataset

ScreenEval is a publicly released dataset for evaluating factual inconsistency detection in long-form dialogues, using human and generated summaries with supporting-utterance annotations.

  • ScreenEval evaluates factual inconsistency in long-form dialogues using TV scripts and summaries from the SummScreen dataset.The dataset includes human summaries plus summaries generated with Longformer and GPT-4.Source: §4 · p. 4
  • Human annotators classify inconsistency for each summary sentence and identify relevant supporting utterances for factually consistent sentences.ScreenEval is released publicly, with construction and annotation details described in the appendix.Source: §4 · p. 5
  • ScreenEval uses the Forever-Dreaming subset of SummScreen because of its manageable summary size and diversity across 21 show genres.The source dataset provides human-written summaries from Wikipedia and TVMaze.Source: §4 · p. 5
  • The dataset contains 52 source documents averaging 6,073 tokens, 624 summary sentences, and 455 artificially generated sentences.Summaries average 101 tokens and cover major plot points and character developments.Source: §4 · p. 5

5 Experiments

The experiments evaluate SCALE across the multi-task TRUE benchmark and long-dialogue ScreenEval using diverse competitive systems and metrics for accuracy, efficiency, calibration, and retrieval.

5.1 Datasets

  • TRUE contains 11 datasets spanning four NLG tasks, with source documents, generated text, and factual inconsistency labels.Its datasets cover summarization, dialogue, fact verification, and paraphrasing tasks.Source: §5.1 · p. 5, Table 2 · p. 5

5.2 Competitive Systems

  • ScreenEval comparisons include NLI, question-answering, GPT, semantic-similarity, and n-gram systems, while retrieval is compared with SuperPal and seNtLI.SCALE uses Flan-T5 base, XL, and XXL backbones.Source: §5.2 · pp. 5–6, §5.1 · p. 5

5.3 Metrics

  • Evaluation measures accuracy with ROC_AUC, Pearson, Kendall_Tau, and F1_Macro, efficiency with wall-clock time, and explanations with calibration and relevant-source retrieval.Calibration uses Expected Calibration Error, which compares bin accuracy and confidence; lower ECE indicates better calibration, while retrieval reports recall.Source: §5.3 · p. 6

6 Results

SCALE generalizes across factual inconsistency tasks, with strong accuracy and calibration on TRUE and superior accuracy, speed, and retrieval performance on long-document ScreenEval.

6.1 TRUE

  • SCALEXXL outperforms non-SCALE models on 10 of 11 TRUE datasets, while SCALEXL leads on 8 datasets.These results demonstrate performance across domains and diverse model types.Source: §6.1 · p. 6, Table 3 · p. 7
  • SCALE provides the best average calibration on TRUE, and SCALEXL beats other non-SCALE models on over half the datasets.Lower ECE indicates better calibration.Source: §6.1 · pp. 6–7, Table 4 · p. 7
  • SCALE stays close to the perfectly calibrated line, whereas most models underestimate positives and SummaCConv overestimates them below 0.4.The authors attribute SCALE’s calibration to its larger context window and use 512-token chunks in subsequent experiments.Source: §6.1 · p. 7, Figure 3 · p. 7

6.2 ScreenEval

Accuracy and Efficiency Evaluation Results

7 Conclusion

SCALE is presented as a task-agnostic factual inconsistency detector with state-of-the-art performance across diverse NLG tasks and datasets. It also improves calibration, long-document evaluation efficiency, and relevant-utterance retrieval.

  • SCALE achieves state-of-the-art factual inconsistency detection across four NLG tasks and 11 TRUE benchmark datasets.Source: §7 · p. 9
  • SCALE is superior in calibration, producing interpretable scores that enable accurate score comparisons.Source: §7 · p. 9
  • On ScreenEval, SCALE significantly outperforms models other than GPT-4 while being more cost effective and faster than GPT-4.Source: §7 · p. 9
  • SCALE's retrieval strategy significantly reduces relevant-utterance retrieval time while increasing recall accuracy on long documents.Source: §7 · p. 9

8 Limitations

The limitations concern SCALE's entailment scoring and calibration consistency. Using only “Yes” and “No” logits may lose information, and calibration remains difficult on some tasks and model sizes.

8 Limitations

  • Using only the “Yes” and “No” logits to compute entailment may reduce accuracy by discarding information associated with similar tokens.Source: §8 · p. 9
  • Aggregate calibration is better, but calibration still varies across certain NLG tasks and model sizes.Source: §8 · p. 9

A Details of Construction of ScreenEval Dataset

ScreenEval summaries are produced by GPT-4, Longformer, and humans. The model-generated summaries use systems with large context windows to attend globally over long source dialogues.

A.1 Building the Dataset

ScreenEval is constructed from filtered TV scripts and summaries generated by Longformer, GPT-4, and humans. The resulting dataset preserves long documents while accommodating model and annotation constraints.

A.1 Building the Dataset

  • The dataset retains 52 TV scripts with an average length of 6,073 tokens and 624 summary sentences after filtering summary lengths and GPT-4 token limits.Source: §A.1 · p. 12

A.2 Longformer

  • Longformer summaries use a SummScreen-FD-finetuned transformer sequence-to-sequence model that globally attends over dialogues up to 16,384 input tokens.Source: §A.2 · p. 12

A.3 GPT-4

  • GPT-4 summarizes selected ScreenEval documents with the prompt “Summarize in 5 sentences or less: ”, restricted to documents under 8,000 tokens.Source: §A.3 · p. 12

A.4 Human Annotation

ScreenEval instances were labeled by qualified crowd workers who judged summary–dialogue consistency and identified supporting utterances. The appendix also specifies prompts used for automated ChatGPT/GPT-4 evaluation.

A.4 Human Annotation

  • ScreenEval workers read each TV-script dialogue, judged whether a highlighted summary sentence was consistent, and selected relevant utterances when answering Yes.Each instance received labels from three human annotators, with unanimous agreement reported 61% of the time.Source: §A.4 · p. 12
  • Annotators were restricted to workers in the US or Canada with MTurk Master status, over 95% approval, and more than 5,000 approved tasks.Batches of 30 were monitored, and workers were rejected for omitting or inventing relevant utterances.Source: §A.4 · p. 12

B.1 ChatGPT/GPT-4

  • The ChatGPT/GPT-4 evaluation prompt asks whether the dialogue factually implies the summary sentence and requires a Yes-or-No answer.The appendix provides separate prompts for the evaluation setup in Figures 6 and 7.Source: §B.1 · p. 12, Figure 6 · p. 13, Figure 7 · p. 13
Loading 2310.13189v2…