What question did this study set out to answer?

This study aims to evaluate generated text quality by modeling it as a text generation problem using BART.

June 22, 2021Open Access

BARTScore: Evaluating Generated Text as Text Generation

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial evaluates text generated by BART, highlighting improved scoring in various metrics.

This study aims to evaluate generated text quality by modeling it as a text generation problem using BART.
Utilized pre-trained encoder-decoder model BART for evaluation of generated text.
Proposed a new metric called BARTScore, applicable in an unsupervised manner.
Tested BARTScore across 16 datasets and 7 evaluation perspectives.
BARTScore outperformed existing metrics in 16 of 22 test settings.
Demonstrated effectiveness across various perspectives including informativeness and factuality.
Enhances understanding of strengths and weaknesses of different evaluation metrics.

Yuan et al. (2021) studied this question.

View Full Paper

Ask AI

Bookmark