[Paper Review] HalluDial: A Large-Scale Benchmark for Automatic Dialogue-Level Hallucination Evaluation

Wen Luo, Tianshu Shen+5 more|arXiv (Cornell University)|Jun 11, 2024

Schizophrenia research and treatmentMedicine3 citations

TL;DR

HalluDial introduces the first large-scale benchmark for automatic dialogue-level hallucination evaluation, covering both factuality and faithfulness hallucinations across 4,094 dialogues (146,856 samples). It enables comprehensive evaluation through hallucination detection, localization, and rationale generation, and supports the development of specialized judge models like HalluJudge, which achieves superior performance in automatic hallucination assessment.

ABSTRACT

Large Language Models (LLMs) have significantly advanced the field of Natural Language Processing (NLP), achieving remarkable performance across diverse tasks and enabling widespread real-world applications. However, LLMs are prone to hallucination, generating content that either conflicts with established knowledge or is unfaithful to the original sources. Existing hallucination benchmarks primarily focus on sentence- or passage-level hallucination detection, neglecting dialogue-level evaluation, hallucination localization, and rationale provision. They also predominantly target factuality hallucinations while underestimating faithfulness hallucinations, often relying on labor-intensive or non-specialized evaluators. To address these limitations, we propose HalluDial, the first comprehensive large-scale benchmark for automatic dialogue-level hallucination evaluation. HalluDial encompasses both spontaneous and induced hallucination scenarios, covering factuality and faithfulness hallucinations. The benchmark includes 4,094 dialogues with a total of 146,856 samples. Leveraging HalluDial, we conduct a comprehensive meta-evaluation of LLMs' hallucination evaluation capabilities in information-seeking dialogues and introduce a specialized judge language model, HalluJudge. The high data quality of HalluDial enables HalluJudge to achieve superior or competitive performance in hallucination evaluation, facilitating the automatic assessment of dialogue-level hallucinations in LLMs and providing valuable insights into this phenomenon. The dataset and the code are available at https://github.com/FlagOpen/HalluDial.

Motivation & Objective

• To address the lack of dialogue-level hallucination evaluation benchmarks that cover both factuality and faithfulness hallucinations.
• To overcome limitations in existing benchmarks, such as focusing only on sentence- or passage-level hallucinations and neglecting localization and rationale generation.
• To reduce reliance on labor-intensive or non-specialized evaluators by enabling automated, high-quality hallucination evaluation.
• To provide a comprehensive resource for training and evaluating specialized judge models for dialogue hallucination detection.
• To support the development of more reliable and trustworthy LLM-based dialogue systems through systematic hallucination assessment.

Proposed method

• The benchmark is constructed from an information-seeking dialogue dataset, incorporating both spontaneous and induced hallucination scenarios.
• Spontaneous hallucinations are generated via a two-step pipeline: diverse dialogue sampling followed by automatic annotation using LLMs.
• Induced hallucinations are created by prompting GPT-4 with task-specific instructions to generate hallucinated responses with explicit explanations.
• Each sample includes hallucination detection, localization (span-level), and rationale (natural language explanation), ensuring interpretability.
• The dataset is structured to support multi-dimensional evaluation: detection, localization, and justification of hallucinations.
• A specialized judge model, HalluJudge, is trained on HalluDial to enable high-performance automatic hallucination evaluation in dialogue systems.

Experimental results

Research questions

• RQ1 To what extent do existing hallucination benchmarks fail to evaluate dialogue-level hallucinations comprehensively?
• RQ2 How effective are current methods at localizing hallucinations and providing interpretable rationales in dialogue contexts?
• RQ3 Can a specialized judge model trained on a large-scale, multi-dimensional hallucination benchmark outperform general LLMs in hallucination evaluation?
• RQ4 What is the comparative performance of automatic evaluation using HalluDial versus human-annotated baselines?
• RQ5 How do factuality and faithfulness hallucinations differ in their manifestation and detectability in dialogue systems?

Key findings

• HalluDial comprises 4,094 dialogues and 146,856 samples, making it the largest publicly available benchmark for dialogue-level hallucination evaluation.
• The benchmark supports three key evaluation tasks: hallucination detection, localization, and rationale generation, enhancing interpretability and reliability.
• HalluJudge, the specialized judge model trained on HalluDial, achieves superior or competitive performance compared to general LLMs in hallucination evaluation.
• The induced hallucination generation method effectively produces diverse, high-quality hallucinated samples with clear explanations, improving benchmark coverage.
• The dataset enables consistent and reproducible evaluation, reducing reliance on costly human annotations or inconsistent API-based LLM evaluators.
• HalluDial is publicly available under a CC BY-NC-SA license on GitHub, supporting ongoing research and model development.