DiaHalu: A Dialogue-level Hallucination Evaluation Benchmark for Large Language Models

TL;DR

DiaHalu introduces a dialogue-level hallucination benchmark for LLMs, addressing gaps in prior work by incorporating faithfulness alongside factuality across four multi-turn domains. The dataset is created via topic collection, dialogue generation with paired LLMs, and expert annotation, including explanations and subtype labeling across five hallucination types. Experimental results show DiaHalu is highly challenging for current detectors and LLMs, with overconfidence in ChatGPT3.5 and faithfulness detection being notably difficult. The benchmark highlights the need for long-context memory and robust reasoning in LLMs to mitigate dialogue-level hallucinations and snowballing effects, offering a valuable resource for future research and evaluation. Ethical and practical considerations are discussed, along with limitations and directions for extending this work.

Abstract

Since large language models (LLMs) achieve significant success in recent years, the hallucination issue remains a challenge, numerous benchmarks are proposed to detect the hallucination. Nevertheless, some of these benchmarks are not naturally generated by LLMs but are intentionally induced. Also, many merely focus on the factuality hallucination while ignoring the faithfulness hallucination. Additionally, although dialogue pattern is more widely utilized in the era of LLMs, current benchmarks only concentrate on sentence-level and passage-level hallucination. In this study, we propose DiaHalu, the first dialogue-level hallucination evaluation benchmark to our knowledge. Initially, we integrate the collected topics into system prompts and facilitate a dialogue between two ChatGPT3.5. Subsequently, we manually modify the contents that do not adhere to human language conventions and then have LLMs re-generate, simulating authentic human-machine interaction scenarios. Finally, professional scholars annotate all the samples in the dataset. DiaHalu covers four common multi-turn dialogue domains and five hallucination subtypes, extended from factuality and faithfulness hallucination. Experiments through some well-known LLMs and detection methods on the dataset show that DiaHalu is a challenging benchmark, holding significant value for further research.

Figures (15)

• Figure 1: Our benchmark not only includes factuality hallucination but also incorporates faithfulness hallucination at the dialogue level, although most benchmarks overlook the latter one.
• Figure 2: The demonstration of the DiaHalu benchmark, which covers four domains and five hallucination subtypes within dialogue-level scenarios. We also provide explanations and sources in the benchmark.
• Figure 3: The complete process of dialogue generation.
• Figure 4: The distribution of five different hallucination subtypes within the four dialogue domains.
• Figure 5: The proportions of the three dialogue round categories. For example, the three values of R7 denote the proportions of 'these three categories in the 7th round' within 'hallucinated samples that have at least seven rounds dialogues'.