G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment

By Yang Liu, Dan Iter, Yichong Xu, Shuo Wang, Ruochen Xu, Chenguang Zhu. Conference on Empirical Methods in Natural Language Processing, 2023

2,735 citations recorded by the source corpus.

The quality of texts generated by natural language generation (NLG) systems is hard to measure automatically. Conventional reference-based metrics, such as BLEU and ROUGE, have been shown to have relatively low correlation with human judgments, especially for tasks that require creativity and diversity. Recent studies suggest using large language models (LLMs) as reference-free metrics for NLG evaluation, which have the benefit of being applicable to new tasks that lack human references. However, these LLM-based evaluators still have lower human correspondence than medium-size neural evaluators. In this work, we present G-Eval, a framework of using large language models with chain-of-thoughts (CoT) and a form-filling paradigm, to assess the quality of NLG outputs. We experiment with two generation tasks, text summarization and dialogue generation. We show that G-Eval with GPT-4 as the backbone model achieves a Spearman correlation of 0.514 with human on summarization task, outperforming all previous methods by a large margin. We also propose preliminary analysis on the behavior of LLM-based evaluators, and highlight the potential issue of LLM-based evaluators having a bias towards the LLM-generated texts. The code is at https://github.com/nlpyang/geval

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/381ab7a640f5b46b62f7e08d1af4a8e0d3eadd55

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

Prometheus: Inducing Fine-grained Evaluation Capability in Language Models (2023, methodology): “Recent work has used GPT-4 or a fine-tuned critique LLM as an evaluator along a single dimension of “preference” (Chiang & yi Lee, 2023; Li et al., 2023; Dubois et al., 2023; Zheng et al., 2023; Liu et al., 2023).”
Understanding the Effects of RLHF on LLM Generalisation and Diversity (2023, methodology): “…tasks include translation, question-answering, and other standard NLP tasks, as well as newer tasks such as using LLMs to simulate human annotators (Dubois et al., 2023; Mao et al., 2023; Liu et al., 2023b) or as content generators for improving other models (Peng et al., 2023; Wang et al., 2023).”
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate (2023, methodology): “The second part and the third part of Table 1 show the results of FairEval’s method and the results of our proposed method respectively.”
Augmenting large language models with chemistry tools (2023, methodology): “GPT-4 has been recently presented and used as a self-evaluation method 5,62 , but these results indicate that when it lacks the required understanding to answer a prompt, it also lacks information to evaluate the prompt completions and thus fails to provide a trustworthy assessment, rendering it…”
CRAFT Your Dataset: Task-Specific Synthetic Dataset Generation Through Corpus Retrieval and Augmentation (2024, neutral): “Further implementation details and filtering statistics are provided in Ap-pendices A and D, respectively.”
Aligning Large Language Models with Human: A Survey (2023, background): “Chain of Hindsight (CoH) (Liu et al., 2023b), on the other hand, directly incorporates human preference as a pair of parallel responses discriminated as low-quality or high-quality using natural language prefixes.”
AlignScore: Evaluating Factual Consistency with A Unified Alignment Function (2023, background): “Concurrent work proposes to combine large language models (LLMs) with prompting to evaluate different aspects of generated text, including factual consistency (Fu et al., 2023; Liu et al., 2023; Gao et al., 2023).”
Evaluating Large Language Models at Evaluating Instruction Following (2023, neutral): “…traditional natural language generation tasks (Chiang & Lee, 2023; Fu et al., 2023; Wang et al., 2023a; Kocmi & Federmann, 2023; Chen et al., 2023b; Liu et al., 2023b), which has been demonstrated to score higher correlations with humans than using conventional reference-based evaluation, e.g.,…”
PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits (2023, background): “…research has investigated the use of LLMs for automatic personality prediction (Gane-san et al., 2023; Rao et al., 2023; Cao and Kosinski, 2023; Yang et al., 2023), assessing the ability of LLMs to express certain personality traits (Li et al., 2022; Pan and Zeng, 2023; Safdari et al., 2023)…”
Revisiting the Gold Standard: Grounding Summarization Evaluation with Robust Human Evaluation (2022, methodology): “Without targeted evaluation, these prior preferences can affect the annotators’ assessments for task-specific quality of summaries and make them favor LLM-summaries; therefore the superiority of LLMs in conditional text generation tasks may be overestimated.”
CMMLU: Measuring massive multitask language understanding in Chinese (2023, background): “However, some recent work (Goyal et al., 2022; Liu et al., 2023b) have demonstrate that LLM can perform even better than human or human annotators on some tasks such as summarization, leading to a re-evaluation of the appropriateness of using these benchmarks.”
FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation (2023, methodology): “A series of concurrent work verifies the precision of the citations (attributions) provided by the model (Gao et al., 2022; Liu et al., 2023a; Yue et al., 2023; Gao et al., 2023).”
Large Language Model Alignment: A Survey (2023, background): “, 2022) can be explored to significantly enhance the reasoning ability of LLMs, thereby improving evaluations that demand reasoning skills (Wang et al., 2023a; Liu et al., 2023c; Zheng et al., 2023a).”
AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents (2024, background): “LLM-based “judges” used to evaluate natural language generation on open-ended questions [7, 71, 32] are emerging as a solution to dynamic real-world environments [10, 35] and while showing promise also for computer control agents [39, 16] they remain imperfect evaluators.”
DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (2023, methodology): “We conduct an additional study of the quality of generated text using GPT4, given the fact that several prior studies Chiang & Lee (2023a); Liu et al. (2023) have shown the great potential of GPT-4 to serve as an alternative to human evaluation.”