Localizing Factual Inconsistencies in Attributable Text Generation

2024-10-09

arXiv.org

AI Summary

Precise localization of factual inconsistencies in attributable text generation remains challenging. Method: We propose QASemConsistency—a framework grounded in Neo-Davidsonian semantics that decomposes generated text into predicate-argument-level question-answer pairs, establishing the first interpretable, fine-grained formalization of consistency. It then employs supervised entailment models and open-source large language models to systematically compare each QA pair against trusted source texts for pinpoint error localization. Contribution/Results: The framework unifies high-agreement human annotation (Cohen’s κ > 0.7) and automated detection within a single modeling paradigm. Experiments demonstrate significant improvements in factual error localization accuracy and validate the effectiveness of multiple automated detection strategies, thereby introducing a novel paradigm for factuality assessment in attributable generation.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

There has been an increasing interest in detecting hallucinations in model-generated texts, both manually and automatically, at varying levels of granularity. However, most existing methods fail to precisely pinpoint the errors. In this work, we introduce QASemConsistency, a new formalism for localizing factual inconsistencies in attributable text generation, at a fine-grained level. Drawing inspiration from Neo-Davidsonian formal semantics, we propose decomposing the generated text into minimal predicate-argument level propositions, expressed as simple question-answer (QA) pairs, and assess whether each individual QA pair is supported by a trusted reference text. As each QA pair corresponds to a single semantic relation between a predicate and an argument, QASemConsistency effectively localizes the unsupported information. We first demonstrate the effectiveness of the QASemConsistency methodology for human annotation, by collecting crowdsourced annotations of granular consistency errors, while achieving a substantial inter-annotator agreement ($kappa>0.7)$. Then, we implement several methods for automatically detecting localized factual inconsistencies, with both supervised entailment models and open-source LLMs.

Problem

• Localizing factual inconsistencies in attributable text generation
• Pinpointing errors at fine-grained semantic level
• Assessing predicate-argument propositions against reference texts

Innovation

• Decomposing text into predicate-argument propositions
• Expressing propositions as simple question-answer pairs
• Assessing each QA pair against reference texts