Controlled automatic task-specific synthetic data generation for hallucination detection

We present a novel approach to automatically generate task-specific synthetic datasets for hallucination detection. Our approach features a two-step generation-selection pipeline, where the generation step integrates a hallucination pattern guidance module and a language style alignment module. Hallucination pattern guidance makes it possible to curate synthetic datasets covering the most important hallucination patterns specific to target applications. Language style alignment improves the dataset quality by aligning the style of the synthetic dataset with benchmark text. To obtain robust supervised detectors from synthetic datasets, we also propose a data mixture strategy to improve performance robustness and model generalization. Our supervised hallucination detectors trained on synthetic datasets outperform in-context-learning (ICL)-based detectors by a large margin. Our extensive experiments confirm the benefits of our two-staged generation pipeline with cross-task and cross-hallucination pattern generalization. Our data-mixture-based training further improves generalization and the robustness of hallucination detection.