Are LLM-based Evaluators Confusing NLG Quality Criteria?

Xinyu Hu, Mingqi Gao, Sen Hu, Yang Zhang, Yicheng Chen, Teng Xu, Xiaojun Wan

Abstract

Some prior work has shown that LLMs perform well in NLG evaluation for different tasks. However, we discover that LLMs seem to confuse different evaluation criteria, which reduces their reliability. For further verification, we first consider avoiding issues of inconsistent conceptualization and vague expression in existing NLG quality criteria themselves. So we summarize a clear hierarchical classification system for 11 common aspects with corresponding different criteria from previous studies involved. Inspired by behavioral testing, we elaborately design 18 types of aspect-targeted perturbation attacks for fine-grained analysis of the evaluation behaviors of different LLMs. We also conduct human annotations beyond the guidance of the classification system to validate the impact of the perturbations. Our experimental results reveal confusion issues inherent in LLMs, as well as other noteworthy phenomena, and necessitate further research and improvements for LLM-based evaluation.

Introduction

Evaluating Natural Language Generation (NLG) systems—responsible for tasks like summarization, translation, and data-to-text—is a persistent challenge. Traditional metrics like BLEU or ROUGE often fail to capture semantic nuances, leading the field toward "LLM-as-a-judge" paradigms. In these setups, models like GPT-4 are prompted to score text based on specific quality criteria. However, a critical flaw has emerged: Aspect Confusion. Unlike human evaluators, who can distinguish between a text being ungrammatical yet factually accurate, LLMs often suffer from a "halo effect." A failure in one dimension (e.g., a typo) frequently leads to lower scores across unrelated dimensions (e.g., relevance), making them unreliable for fine-grained debugging and quality control.

Methodology

To systematically investigate this phenomenon, the researchers employed a behavioral testing approach. They first established a rigorous Hierarchical Taxonomy of 11 common NLG quality aspects, such as Coherence, Fluency, and Informativeness, to eliminate vague definitions and inconsistent conceptualization.

Hierarchical taxonomy of 11 common NLG quality aspects.

Using this taxonomy, 18 specific perturbation attacks were designed to isolate failures. For instance, "Grammaticality" was tested by introducing spelling errors, while "Faithfulness" was tested by negating facts. The study utilized datasets including SummEval, CNN/Dailymail, and SAMSum. Evaluation was conducted via:

  • Directional Testing: Ensuring a targeted error (e.g., grammar) correctly drops the relevant score (e.g., Fluency).
  • Invariance Testing: Ensuring a targeted error does not affect unrelated scores (e.g., Faithfulness).

Results and Analysis

The experiments confirmed that LLM evaluators struggle with disentanglement. A primary finding is the significant correlation gap between LLM and human scores.

Correlation heatmap between LLM and human evaluations.

As shown in the heatmap, LLM scores across different aspects (like Fluency and Coherence) are much more highly correlated with each other than human scores are. This suggests that LLMs often fail to provide independent assessments. Key observations include:

  • Surface-Level Sensitivity: LLMs are overly sensitive to surface errors. A simple spelling mistake can cause a model to penalize "Non-hallucination," even if the content remains factually grounded.
  • The "Halo Effect": Fact-based perturbations (e.g., changing a name) caused score drops in nearly every category, including Readability and Fluency, which should theoretically remain stable.
  • Prompting Limitations: Providing more detailed definitions or using different prompting styles (Detailed vs. Simplified) did not significantly mitigate this confusion. LLMs appear to rely more on the internal "vibe" of terms like "Fluency" than on the specific instructions provided in the prompt.

Discussion

The study highlights that while GPT-4 is a more sensitive error detector than GPT-3.5, it is also more prone to "bleeding" penalties across categories. Its high sensitivity makes it a harsher judge, but one that lacks specificity. Open-source models like Prometheus fared even worse, struggling to differentiate between aspects effectively. This suggests that the current reliance on general-purpose LLMs for evaluation may be fundamentally limited by their inability to disentangle style from content.

Conclusion

The findings necessitate a shift in how we build NLG evaluators. Simply prompting existing LLMs is insufficient for reliable, fine-grained quality control. Future research should focus on training or fine-tuning models specifically for disentangled evaluation, ensuring that "judges" can accurately distinguish between the "what" (adequacy) and the "how" (readability) of generated text.

Summary updated

Abstract

Some prior work has shown that LLMs perform well in NLG evaluation for different tasks. However, we discover that LLMs seem to confuse different evaluation criteria, which reduces their reliability. For further verification, we first consider avoiding issues of inconsistent conceptualization and vague expression in existing NLG quality criteria themselves. So we summarize a clear hierarchical classification system for 11 common aspects with corresponding different criteria from previous studies involved. Inspired by behavioral testing, we elaborately design 18 types of aspect-targeted perturbation attacks for fine-grained analysis of the evaluation behaviors of different LLMs. We also conduct human annotations beyond the guidance of the classification system to validate the impact of the perturbations. Our experimental results reveal confusion issues inherent in LLMs, as well as other noteworthy phenomena, and necessitate further research and improvements for LLM-based evaluation.
BibTeX
@inproceedings{hu2024are,
  author = {Xinyu Hu and Mingqi Gao and Sen Hu and Yang Zhang and Yicheng Chen and Teng Xu and Xiaojun Wan},
  title = {{Are LLM-based Evaluators Confusing NLG Quality Criteria?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.acl-long.516},
}
Figure