GPTScore: Evaluate as You Desire

TL;DR

GPTScore introduces a flexible, instruction-based evaluation framework leveraging pre-trained models' emergent abilities to assess generated texts across multiple facets without requiring annotated data.

Contribution

The paper presents GPTScore, a novel evaluation method that uses pre-trained models' zero-shot capabilities for customizable, multi-dimensional text quality assessment.

Findings

Effective evaluation across four text generation tasks

Ability to customize evaluation criteria via natural language instructions

No need for annotated datasets for different evaluation aspects

Abstract

Generative Artificial Intelligence (AI) has enabled the development of sophisticated models that are capable of producing high-caliber text, images, and other outputs through the utilization of large pre-trained models. Nevertheless, assessing the quality of the generation is an even more arduous task than the generation itself, and this issue has not been given adequate consideration recently. This paper proposes a novel evaluation framework, GPTScore, which utilizes the emergent abilities (e.g., zero-shot instruction) of generative pre-trained models to score generated texts. There are 19 pre-trained models explored in this paper, ranging in size from 80M (e.g., FLAN-T5-small) to 175B (e.g., GPT3). Experimental results on four text generation tasks, 22 evaluation aspects, and corresponding 37 datasets demonstrate that this approach can effectively allow us to achieve what one desires to evaluate for texts simply by natural language instructions. This nature helps us overcome several long-standing challenges in text evaluation--how to achieve customized, multi-faceted evaluation without the need for annotated samples. We make our code publicly available at https://github.com/jinlanfu/GPTScore.