NLP-对话评估指标

image.png

The quality of gener-
ated responses is evaluated with different metrics
from three aspects: relevance, diversity, and nov-
elty.

Human Evaluation.

• Average Score and Best@1
  
  表给出了人类裁判的平均得分和Best@1的比率。第一种是五名法官的平均值;后者计算四个系统中认为对应响应最好的法官的比例。

ACL 2019Target-Guided Open-Domain Conversation