deltaBLEU:一种面向目标内在多样化的生成任务的判别性度量
计算与语言
2015-06-25 v2
摘要
我们介绍了判别性BLEU(deltaBLEU),这是一种用于具有多样化可能输出的生成任务中生成文本内在评估的新度量。参考字符串由人类评分者按[-1, +1]尺度评分质量,以加权多参考BLEU。在涉及对话响应生成的任务中,deltaBLEU与人类判断具有合理的相关性,并且在斯皮尔曼rho和肯德尔tau方面均优于句子级和IBM BLEU。
引用
@article{arxiv.1506.06863,
title = {deltaBLEU: A Discriminative Metric for Generation Tasks with Intrinsically Diverse Targets},
author = {Michel Galley and Chris Brockett and Alessandro Sordoni and Yangfeng Ji and Michael Auli and Chris Quirk and Margaret Mitchell and Jianfeng Gao and Bill Dolan},
journal= {arXiv preprint arXiv:1506.06863},
year = {2015}
}
备注
6 pages, to appear at ACL 2015