Source-linked AI summary
Is ChatGPT a Good NLG Evaluator? A Preliminary Study
Jiaan Wang, Yunlong Liang, Fandong Meng, Zengkui Sun, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, Jie Zhou
TL;DR
The paper asks whether ChatGPT can serve as a reliable evaluator for NLG outputs, a question motivated by limited evidence about its agreement with human judgments. It prompts ChatGPT as a human evaluator with task- and aspect-specific instructions and tests it across five meta-evaluation datasets. ChatGPT shows state-of-the-art or competitive correlations in most cases, while performance depends on prompts and dataset biases.
Problem
Whether ChatGPT can reliably evaluate NLG outputs remains underexplored, despite the importance of metrics that correlate with human judgments.
Method
The study prompts ChatGPT as a human evaluator using task-specific and aspect-specific instructions, with reference-free and reference-based settings, across five NLG meta-evaluation datasets.
Results
ChatGPT achieves state-of-the-art correlations on SummEval and OpenMEVA and competitive results on NewsRoom and BAGEL, with high human correlation in most cases.
Takeaways & Limitations
ChatGPT shows potential as a general NLG metric, especially for creative tasks such as story generation where multiple outputs can satisfy humans.
Takeaways & Limitations
Performance depends on prompt design and reference-related dataset biases, and the preliminary study covers only English NLG tasks and limited task types.
Abstract
from arXiv · showhide
Recently, the emergence of ChatGPT has attracted wide attention from the computational linguistics community. Many prior studies have shown that ChatGPT achieves remarkable performance on various NLP tasks in terms of automatic evaluation metrics. However, the ability of ChatGPT to serve as an evaluation metric is still underexplored. Considering assessing the quality of natural language generation (NLG) models is an arduous task and NLG metrics notoriously show their poor correlation with human judgments, we wonder whether ChatGPT is a good NLG evaluation metric. In this report, we provide a preliminary meta-evaluation on ChatGPT to show its reliability as an NLG metric. In detail, we regard ChatGPT as a human evaluator and give task-specific (e.g., summarization) and aspect-specific (e.g., relevance) instruction to prompt ChatGPT to evaluate the generated results of NLG models. We conduct experiments on five NLG meta-evaluation datasets (including summarization, story generation and data-to-text tasks). Experimental results show that compared with previous automatic metrics, ChatGPT achieves state-of-the-art or competitive correlation with human judgments in most cases. In addition, we find that the effectiveness of the ChatGPT evaluator might be influenced by the creation method of the meta-evaluation datasets. For the meta-evaluation datasets which are created greatly depending on the reference and thus are biased, the ChatGPT evaluator might lose its effectiveness. We hope our preliminary study could prompt the emergence of a general-purposed reliable NLG metric.
1 Introduction
Although ChatGPT performs well on many NLG tasks, whether it can reliably evaluate generated text remains unclear. This study tests ChatGPT as a human-like NLG evaluator and finds strong performance in most settings, with sensitivity to prompts and dataset construction.
- The study asks whether ChatGPT can reliably evaluate the quality of natural-language generations.
- ChatGPT is prompted with task-specific and aspect-specific instructions to assess NLG outputs.The experiments also vary scoring criteria and whether golden references are provided.
- Experiments cover five NLG meta-evaluation datasets spanning summarization, story generation, and data-to-text tasks.
- ChatGPT correlates highly with human judgments in most cases, especially for story generation and other creative NLG tasks.The contribution summary attributes this strength partly to settings where multiple generations can satisfy humans.
- ChatGPT evaluation is prompt-sensitive, so prompts should be designed carefully for different tasks and aspects.
- When meta-evaluation datasets depend heavily on references, ChatGPT might lose effectiveness because similarity-based metrics already strongly reflect human judgments.
2 Related Work
NLG metrics include reference-overlap, embedding-based, LLM-based, and task-specific approaches. Existing methods improve semantic or human-related evaluation in different ways, but ChatGPT had not yet been studied as a general evaluator of textual outputs.
- Automatic NLG metrics compare generated texts with reference texts to estimate generation quality and reduce human-evaluation labor.
- N-gram-based Metrics: N-gram-based metrics measure lexical overlap, including ROUGE, BLEU, Distinct-n, and METEOR.They can correlate well with references but may miss equivalent meanings expressed with different surface forms.
- Embedding-based Metrics: Embedding-based metrics such as BERTScore and MoverScore target semantic similarity using word or sentence embeddings.Their potential may be limited by embedding quality.
- LLM-based Metrics: LLM-based metrics such as BARTScore and GPTScore can correlate strongly with human judgments, but ChatGPT had not been evaluated as a general NLG metric.
- Other Metrics: Task-specific metrics include PRISM for translation, StoryER for story generation, and PARENT for table-to-text generation.
- ChatGPT as an Evaluator: Prior ChatGPT evaluator studies addressed translation quality and human personalities, whereas this work targets general textual outputs.
3 ChatGPT for NLG Evaluation
The paper prompts ChatGPT to act as a reference-free or reference-based NLG metric by specifying the task and evaluation aspect. It tests continuous scores and star ratings, extracts numerical judgments, and can include human references in the prompt.
- Experimental Evaluation: The evaluation framework is tested on summarization, with Table 1 reporting sample-level Spearman, Pearson, and Kendall correlations across aspects on SummEval.
- Prompt Design: ChatGPT is instructed to evaluate an NLG output by specifying the task and the aspect being assessed.Examples include summarization and fluency.
- Scoring Criteria: Direct-assessment prompts request a continuous score from 0 to 100, with endpoints defined by the evaluated aspect and its antonym.
- Scoring Criteria: Star prompts request a one-to-five-star rating using the same aspect-specific endpoint definitions.
- Output Processing: ChatGPT returns a judgment and explanatory description, after which numerical scores are extracted using heuristic rules.
- Reference Conditions: Reference-free prompts provide the conditioned input and generated output, while reference-based prompts additionally provide a human reference.
4 Experiments
The experiments evaluate ChatGPT as an NLG metric using correlation with human judgments across summarization, story generation, and data-to-text datasets. Results are generally strong but depend on dataset construction and prompt design.
- Experimental Setup: The study compares automatic metrics with human judgments using Spearman, Pearson, and Kendall’s Tau correlations at sample and dataset levels.The evaluation strategies aggregate automatic and human scores either across model outputs for each condition or across the dataset.
- Datasets: Experiments cover SummEval, NewsRoom, and RealSumm for summarization, OpenMEVA-ROC for story generation, and BAGEL for data-to-text generation.The datasets contain human judgments of generated outputs across multiple NLG tasks and evaluation aspects.
- Text Summarization: ChatGPT achieves state-of-the-art correlation in most SummEval aspects and outperforms ROUGE and BERTScore by a large margin on NewsRoom.The reported results concern sample-level and dataset-level evaluations, with similar trends across the corresponding tables.
- Text Summarization: On RealSumm, ChatGPT (Stars w/ ref) reaches 0.195 sample-level Spearman correlation, versus 0.498 for ROUGE-1.The authors attribute this weaker performance to RealSumm’s pyramid-based, reference-oriented human annotation procedure.
- Story Generation: ChatGPT achieves the best correlations on all OpenMEVA-ROC measures, including 0.507 sample-level Spearman correlation versus 0.324 for PPL.Similarity-based metrics show weak correlations on this open-ended story-generation task, while ChatGPT results vary substantially across prompts.
- Data-to-Text Generation: ChatGPT achieves competitive correlations with previous state-of-the-art baselines on BAGEL, but reference-free informativeness is not evaluated because BAGEL defines informativeness against the gold reference.The reference is therefore required for evaluating that aspect.
5 Conclusion
The report finds that ChatGPT can serve as an NLG evaluator, achieving state-of-the-art or competitive correlations with human judgments across several tasks, while its effectiveness depends on dataset and prompt design.
- ChatGPT achieves new state-of-the-art correlations with human judgments on SummEval and OpenMEVA, and competitive results on NewsRoom and BAGEL.
- Lexical biases in meta-evaluation datasets can limit ChatGPT evaluator performance when datasets depend heavily on reference similarity.
- ChatGPT evaluator performance is sensitive to prompt format, so prompts should be carefully designed for different tasks and aspects.
- The report suggests that ChatGPT may exceed its current performance and provide a reliable NLG metric in the near future.
Limitations
The preliminary study identifies prompt dependence, incomplete task coverage, and the lack of an official ChatGPT API at the time of experimentation as limitations.
- ChatGPT’s performance as an NLG metric depends on prompt design, motivating future work on more powerful prompts.
- The report does not evaluate some mainstream NLG tasks, including dialogue generation and report generation.
- The experiments were conducted before OpenAI released an official ChatGPT API.