Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance (2410.18889v1)

Abstract: NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models. While crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in LLMs offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples. Through a case study of four datasets from the TRUE benchmark, covering different tasks and domains, we empirically analyze the labeling quality of existing datasets, and compare expert, crowd-sourced, and our LLM-based annotations in terms of agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve model performance.

Analysis of LLM-Annotated Label Correction in NLP Benchmarks

The paper under discussion, "Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance," presents an empirical investigation into the efficacy of LLMs for the correction of label errors in NLP datasets. This paper emphasizes the significance of high-quality annotations in benchmarks, crucial for both model training and evaluation.

Context and Motivation

The advent of LLMs has catalyzed significant advances in NLP, necessitating larger and more diverse datasets. Traditional annotation methods, relying either on domain experts or crowd-sourcing, exhibit limitations in terms of scalability and label consistency. This paper posits that LLMs can play a pivotal role in identifying and correcting label errors, offering an alternative that balances scale and precision.

Methodology

The authors propose a method termed "LLM-as-a-judge," wherein an ensemble of LLMs is employed to detect mislabeled instances. This approach involves:

1. LLM Ensemble Creation: Deploying multiple LLMs diversified by prompts to enhance reliability in prediction.
2. Flagging Protocol: Instances showing strong disagreement between LLM predictions and original labels are flagged for potential mislabeling.

The paper focuses on four datasets from the TRUE benchmark, examining annotation quality across multiple tasks, including summarization and dialogue. The detection process is augmented by comparative analyses with expert and crowd-sourced annotations to establish a gold label standard.

Key Findings

• Label Error Prevalence: Existing datasets reveal label error rates between 6% to 21%, indicating a significant room for improvement in current benchmarks.
• Impact on Performance: Correcting these label errors resulted in noticeable increases in model performance, suggesting that many reported errors are due to flawed labels rather than model shortcomings.
• LLM Capabilities: The precision of LLMs in detecting label errors improves as their confidence increases. Instances with high LLM confidence yielded a 15% improvement in model performance after correction.
• Comparison with Human Annotation: LLMs outperformed crowd-sourced annotations, offering better trade-offs between quality and efficiency. However, they matched experts only when methods addressed their limitations in accuracy.

Implications and Future Directions

The correction of label errors using LLMs not only enhances model performance but also ensures the reliability of NLP benchmarks. The findings suggest substantial implications for model evaluation protocols, urging a reassessment of previously established performance baselines.

Practically, LLM-based annotation offers a scalable and cost-effective solution to dataset creation, potentially reducing the reliance on manual annotations. Theoretical advancements may focus on refining LLM ensemble techniques to further improve error detection accuracy and address biases.

As future research, extending this methodology across varied tasks and examining long-term impacts on model generalization and transfer learning would be valuable. Additionally, exploring hybrid models that combine LLMs and human intervention via active learning might provide a nuanced approach to dataset refinement.

Conclusion

This paper provides compelling evidence for the integration of LLMs into the annotation pipeline, presenting a sophisticated method to leverage their capabilities in improving dataset quality. The approach delineated offers a nuanced understanding of label errors, facilitating more accurate and effective NLP model training and evaluation in the future.