贪心学院面试整理2
Generating Hierarchical Explanations on Text classification via Feature Interaction Detection:
local explanation model: ignore interactions between words and phrases
lime explanation: 单词级别的解释
cd explanation:短语级别的解释
HEDGE:关注interaction
contribution: 有点类似feature importance
- building hierarchical explanation
interaction score计算
contribution set计算
Experiments:
评估指标:
AOPC
Log-odds
Cohesion-score:一致性得分
seq2seq attention
评估方式:
BLEU,比较候选译文和参考译文里的n-gram的重合程度,重合程度越高就认为译文质量越高。
是为了惩罚短的译文
ROUGE:BLEU专注于精度,ROUGE专注召回率
ROUGE-L:最长公共子序列
ROUGE-S:skip n-gram跳词
curriculum learning