TL;DR
This paper investigates whether ChatGPT can serve as a reliable automatic metric for evaluating the quality of Natural Language Generation (NLG) models. By treating ChatGPT as a human evaluator and prompting it with specific instructions, the study finds that ChatGPT achieves state-of-the-art or competitive correlations with human judgments across various NLG tasks, suggesting its potential as a general-purpose NLG evaluation metric.
What Problem Does This Paper Address?
Evaluating the quality of text generated by NLG models is a notoriously difficult and time-consuming task, often requiring extensive human annotation. Existing automatic metrics, while useful, frequently struggle to align well with human perceptions of quality, especially for nuanced aspects like coherence, relevance, and fluency. This gap between automatic metrics and human judgment limits our ability to effectively develop and compare NLG systems. The authors aim to explore if large language models like ChatGPT, which have demonstrated impressive capabilities in various NLP tasks, can bridge this gap and act as a more human-aligned NLG evaluator.
Methods
The core methodology involves treating ChatGPT as a human evaluator. The researchers designed specific prompts to guide ChatGPT in assessing generated text. These prompts were tailored to different NLG tasks (e.g., summarization, story generation, data-to-text) and specific evaluation aspects (e.g., coherence, relevance, consistency, fluency). Two main prompting strategies were employed:
- Direct Assessment (DA): ChatGPT was asked to provide a numerical score on a continuous scale (e.g., 0-100) or a star rating (1-5) based on predefined criteria.
- Reference-based Evaluation: In addition to the generated text and source input, the prompts also included human-written references to guide ChatGPT's evaluation.
The generated scores from ChatGPT were then compared against human judgments using standard correlation metrics: Spearman correlation, Pearson correlation, and Kendall's Tau.
Datasets
The study utilized five widely-used NLG meta-evaluation datasets:
- SummEval: A dataset for text summarization, providing human judgments on coherence, relevance, consistency, and fluency.
- NewsRoom: Another text summarization dataset, offering human judgments on coherence, relevance, informativeness, and fluency.
- RealSumm: A dataset for text summarization that evaluates pyramid recall.
- OpenMEVA-ROC: A dataset for story generation, featuring human annotations for overall quality.
- BAGEL: A dataset for data-to-text generation.
Key Insights and Ideas
- ChatGPT as a Promising NLG Evaluator: The study strongly suggests that ChatGPT can effectively serve as an NLG evaluation metric, often outperforming traditional metrics in its correlation with human judgments.
- Task and Aspect Specificity: ChatGPT's performance is not uniform across all NLG tasks and aspects. It shows particular promise in evaluating more creative tasks like story generation, where human judgment is highly subjective.
- Prompt Sensitivity: The way prompts are designed significantly impacts ChatGPT's evaluation performance. Careful crafting of instructions is crucial for obtaining reliable results.
- Dataset Bias Matters: The characteristics of the meta-evaluation datasets, particularly how human judgments were collected (e.g., reference-dependent vs. reference-free), can influence the effectiveness of NLG metrics, including ChatGPT. Datasets with "lexical biases" (where similarity to references strongly correlates with human scores) can diminish ChatGPT's advantage.
Key Results
Across the evaluated datasets, ChatGPT consistently achieved high correlations with human judgments.
- On SummEval, ChatGPT achieved state-of-the-art correlations in most aspects, outperforming metrics like ROUGE and BERTScore.
- On NewsRoom, ChatGPT also outperformed dominant summarization metrics.
- For story generation (OpenMEVA-ROC), ChatGPT demonstrated its strongest performance, indicating its potential for evaluating open-ended and creative NLG tasks.
- In data-to-text generation (BAGEL), ChatGPT achieved competitive correlations with existing baselines.
- The study found that reference-free prompts generally yielded better results than reference-based prompts on some datasets, while the opposite was true for others, highlighting the importance of prompt design.
Limitations and Future Work
The authors acknowledge several limitations:
- Prompt Engineering: ChatGPT's performance is sensitive to prompts, and future work could explore more powerful and robust prompting strategies.
- Task Coverage: The study focused on summarization, story generation, and data-to-text. Future work could extend the evaluation to other NLG tasks like dialogue generation and report generation.
- Reproducibility: Experiments were conducted on the ChatGPT website without API access, using default temperature settings, which might affect reproducibility. Future work could leverage official APIs for more controlled experiments.
- Dataset Scope: The experiments were primarily conducted on English datasets. Extending to other languages and cross-lingual scenarios is a potential future direction.
- Dataset Quality: The correlation between ChatGPT and human judgments is influenced by the quality and nature of the meta-evaluation datasets. The authors suggest adapting their method to newer, more challenging benchmarks designed to test LLM evaluators.
Summary updated