Fast and Accurate Factual Inconsistency Detection Over Long Documents

TL;DR

The paper addresses factual inconsistency in long-form generations by introducing SCALE, a chunking-based, NLI-driven detector that conditions on large source chunks and uses a BST-based retrieval strategy for explanations. SCALE achieves state-of-the-art results on the TRUE benchmark and demonstrates strong calibration, efficiency, and explainability on a newly released ScreenEval long-dialogue dataset. The approach balances accuracy, speed, and interpretability, enabling practical online deployment for long documents. The work also contributes ScreenEval as a long-form dialogue inconsistency dataset and provides public code and data to foster reproducibility and further research.

Abstract

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.

Figures (7)

• Figure 1: Chunking mechanism for SCALE to produce a score given a source document and generated text. The source document is broken into chunks (represented by dashed lines) and each chunk is fed into to a prompt as the premise. The highlighted generated text is fed into all prompts as the hypothesis. Each prompt is then run through Flan-T5 and the resulting logits are used to compute the entailment score.
• Figure 2: Visualization of SCALE's in context retrieval using chunks to find the most relevant source utterance given a sentence. Each chunk is scored by SCALE, as shown by the gray boxes.
• Figure 3: Calibration curves on the PAWS benchmark
• Figure 4: Effect of different chunk sizes on calibration performance on ScreenEval dataset.
• Figure 5: Effect of different chunk sizes on $\text{SCALE}_{large}$ performance and time on ScreenEval dataset. ROC$\_$AUC score is multiplied by 100 for readability.