生成式文本质量的自动评估方法综述

Human evaluation, as the gold standard for assessing the quality of generated text, is prohibitively expensive. Automatic evaluation, on the other hand, aims to achieve high correlation with manual evaluation, thereby enabling automated analysis and assessment of generated text quality. With the iterative advancement of technologies in the field of natural language processing, the automatic evaluation of generated text quality has undergone several paradigm shifts. However, there is still a lack of systematic summarization of these automatic evaluation techniques in the academic community. Therefore, this paper first systematically summarizes the existing methods for automatic evaluation of generated text. It then analyzes the main development trends of these automatic evaluation methods. Finally, to provide a more comprehensive understanding of automatic evaluation, the paper discusses and anticipates future research directions in the field of automatic evaluation.