Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HAD: HAllucination Detection Language Models Based on a Comprehensive Hallucination Taxonomy

Fan Xu, Xinyu Hu, Zhenghan Yu, Li Lin, Xu Zhang, Yang Zhang, Wei Zhou, Jinjie Gu, Xiaojun Wan

TL;DR

This work tackles hallucination in natural language generation by introducing a fine-grained, three-level taxonomy that captures 11 hallucination types across faithfulness and factuality. The authors build HAD, a unified pipeline that jointly classifies, localizes, and corrects hallucinations, trained on a large synthetic corpus (~90K samples) and validated on HADTest (2,248 samples). HAD achieves state-of-the-art results on multiple benchmarks (HaluEval, FactCHD, FaithBench) and demonstrates robust out-of-domain performance, with knowledge augmentation further enhancing factuality detection. The approach offers a versatile, task-agnostic framework for improving reliability of LLM outputs across diverse NLG tasks, while also highlighting limitations such as single-span handling and reliance on synthetic data.

Abstract

The increasing reliance on natural language generation (NLG) models, particularly large language models, has raised concerns about the reliability and accuracy of their outputs. A key challenge is hallucination, where models produce plausible but incorrect information. As a result, hallucination detection has become a critical task. In this work, we introduce a comprehensive hallucination taxonomy with 11 categories across various NLG tasks and propose the HAllucination Detection (HAD) models https://github.com/pku0xff/HAD, which integrate hallucination detection, span-level identification, and correction into a single inference process. Trained on an elaborate synthetic dataset of about 90K samples, our HAD models are versatile and can be applied to various NLG tasks. We also carefully annotate a test set for hallucination detection, called HADTest, which contains 2,248 samples. Evaluations on in-domain and out-of-domain test sets show that our HAD models generally outperform the existing baselines, achieving state-of-the-art results on HaluEval, FactCHD, and FaithBench, confirming their robustness and versatility.

HAD: HAllucination Detection Language Models Based on a Comprehensive Hallucination Taxonomy

TL;DR

This work tackles hallucination in natural language generation by introducing a fine-grained, three-level taxonomy that captures 11 hallucination types across faithfulness and factuality. The authors build HAD, a unified pipeline that jointly classifies, localizes, and corrects hallucinations, trained on a large synthetic corpus (~90K samples) and validated on HADTest (2,248 samples). HAD achieves state-of-the-art results on multiple benchmarks (HaluEval, FactCHD, FaithBench) and demonstrates robust out-of-domain performance, with knowledge augmentation further enhancing factuality detection. The approach offers a versatile, task-agnostic framework for improving reliability of LLM outputs across diverse NLG tasks, while also highlighting limitations such as single-span handling and reliance on synthetic data.

Abstract

The increasing reliance on natural language generation (NLG) models, particularly large language models, has raised concerns about the reliability and accuracy of their outputs. A key challenge is hallucination, where models produce plausible but incorrect information. As a result, hallucination detection has become a critical task. In this work, we introduce a comprehensive hallucination taxonomy with 11 categories across various NLG tasks and propose the HAllucination Detection (HAD) models https://github.com/pku0xff/HAD, which integrate hallucination detection, span-level identification, and correction into a single inference process. Trained on an elaborate synthetic dataset of about 90K samples, our HAD models are versatile and can be applied to various NLG tasks. We also carefully annotate a test set for hallucination detection, called HADTest, which contains 2,248 samples. Evaluations on in-domain and out-of-domain test sets show that our HAD models generally outperform the existing baselines, achieving state-of-the-art results on HaluEval, FactCHD, and FaithBench, confirming their robustness and versatility.
Paper Structure (36 sections, 4 figures, 14 tables)

This paper contains 36 sections, 4 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Hallucination Taxonomy
  3. Faithfulness Hallucination
  4. A. Instruction Inconsistency
  5. B. Input Context Inconsistency
  6. C. Internal Inconsistency
  7. Factuality Hallucination
  8. D. Fact Contradiction
  9. E. Fact Fabrication
  10. Data Construction
  11. Source Data Selection
  12. Hallucination Synthesis
  13. Hallucination Injection
  14. Automatic Filtering
  15. Test Data Annotation
  16. ...and 21 more sections

Figures (4)

  • Figure 1: Example of the hallucination detection process.
  • Figure 2: Taxonomy of hallucinations, comprising three levels and 11 fine-grained categories.
  • Figure 3: Statistics of our constructed HADTest.
  • Figure 4: Confusion matrix of the test results generated by HAD-14B.