Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Controlled Automatic Task-Specific Synthetic Data Generation for Hallucination Detection

Yong Xie, Karan Aggarwal, Aitzaz Ahmad, Stephen Lau

TL;DR

The paper tackles the challenge of task-specific hallucinations by proposing a Generation-Selection pipeline that uses Hallucination Pattern Guidance (HPG) and Language Style Alignment (LSA) to create high-quality synthetic datasets for training post-hoc detectors. It further augments robustness with a data mixture strategy across multiple LLM generators, enabling cross-generator, cross-pattern, and cross-task generalization. Empirical evaluation on OpenDialKG, ReDial, and SalesBot shows detectors trained on synthetic data outperform in-context learning detectors by a substantial margin and maintain robust generalization across generators and tasks. While promising, the approach relies on human-curated hallucination patterns and balanced data, suggesting opportunities for improving pattern discovery and distribution-aware mixing in future work.

Abstract

We present a novel approach to automatically generate non-trivial task-specific synthetic datasets for hallucination detection. Our approach features a two-step generation-selection pipeline, using hallucination pattern guidance and a language style alignment during generation. Hallucination pattern guidance leverages the most important task-specific hallucination patterns while language style alignment aligns the style of the synthetic dataset with benchmark text. To obtain robust supervised detectors from synthetic datasets, we also adopt a data mixture strategy to improve performance robustness and generalization. Our results on three datasets show that our generated hallucination text is more closely aligned with non-hallucinated text versus baselines, to train hallucination detectors with better generalization. Our hallucination detectors trained on synthetic datasets outperform in-context-learning (ICL)-based detectors by a large margin of 32%. Our extensive experiments confirm the benefits of our approach with cross-task and cross-generator generalization. Our data-mixture-based training further improves the generalization and robustness of hallucination detection.

Controlled Automatic Task-Specific Synthetic Data Generation for Hallucination Detection

TL;DR

The paper tackles the challenge of task-specific hallucinations by proposing a Generation-Selection pipeline that uses Hallucination Pattern Guidance (HPG) and Language Style Alignment (LSA) to create high-quality synthetic datasets for training post-hoc detectors. It further augments robustness with a data mixture strategy across multiple LLM generators, enabling cross-generator, cross-pattern, and cross-task generalization. Empirical evaluation on OpenDialKG, ReDial, and SalesBot shows detectors trained on synthetic data outperform in-context learning detectors by a substantial margin and maintain robust generalization across generators and tasks. While promising, the approach relies on human-curated hallucination patterns and balanced data, suggesting opportunities for improving pattern discovery and distribution-aware mixing in future work.

Abstract

We present a novel approach to automatically generate non-trivial task-specific synthetic datasets for hallucination detection. Our approach features a two-step generation-selection pipeline, using hallucination pattern guidance and a language style alignment during generation. Hallucination pattern guidance leverages the most important task-specific hallucination patterns while language style alignment aligns the style of the synthetic dataset with benchmark text. To obtain robust supervised detectors from synthetic datasets, we also adopt a data mixture strategy to improve performance robustness and generalization. Our results on three datasets show that our generated hallucination text is more closely aligned with non-hallucinated text versus baselines, to train hallucination detectors with better generalization. Our hallucination detectors trained on synthetic datasets outperform in-context-learning (ICL)-based detectors by a large margin of 32%. Our extensive experiments confirm the benefits of our approach with cross-task and cross-generator generalization. Our data-mixture-based training further improves the generalization and robustness of hallucination detection.

Paper Structure

This paper contains 46 sections, 1 figure, 11 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Problem Setting.
  4. Generation-Selection Pipeline
  5. Generator.
  6. Judge.
  7. Hallucination Pattern Guidance
  8. Language Style Alignment
  9. Data Mixture
  10. Experiments
  11. Setup
  12. Benchmarks.
  13. Setup.
  14. Evaluation.
  15. Synthetic Dataset Analysis
  16. ...and 31 more sections

Figures (1)

  • Figure 1: Automatic generation pipeline. We use non-hallucinated samples to generate the synthetic hallucination dataset with two inputs to the generator: human defined Hallucination Patterns and Language Style Features with language style features, like text tone. These are used by the Generator LLM to generate hallucinated samples, which are then judged by a LLM Judge to finally select the most plausible hallucinated samples.