《What makes a good conversation?How controllable attributes affect human judgments》阅读笔记

动机

一个好的对话需要保持平衡：简洁与细节 持续主题与更换主题 问问题和回答问题

对应四种属性：重复性、独特性、回复相关性和问与答

lowlevel model attributes—Conditional Training

•控制加入句子表示的特征

•将输入x与属性概率z作为解码器的输入生成y

lowlevel model attributes—weighted decoding

•在测试时将特征加入目标函数

•这个特征只能是词的特征

 

Controlling conversational attributes——重复——weighted decoding

•分三种：对话中的自我重复、非对话中的自我重复、重复对方的话（赋予负的权重）

 

Controlling conversational attributes——独特

R是所有响应个数，cw是是包含词w的响应个数，然后将它标准化

Controlling conversational attributes——回复相关

•weighted decoding

•the cosine similarity between the GloVe embedding for the word w, and the sentence embedding for the partner’s last utterance ‘

•conditional training

the overall cosine similarity between the partner’s last utterance ‘ and the model’s response y 发现没有效果

Controlling conversational attributes——问与答的平衡

•weighted decoding

•是否有问词

•副作用：负权重影响这种句子的生成I’m learning how to knit

•正权重：正权重容易生成What??????? or Who? When? How?)这种句子

•conditional training,

•与以往工作中在句子层面进行控制不同，本文在对话层面进行控制，比如e.g.20% questions or 70% questions

We train our CT modelon a control variable z with 11 possible values:
（1……10）i/10是概率

结果：

conditional training方式更好

疑问词：how, what, when, where, which, who, whom,whose,why

另建立了一个交流群，感兴趣的朋友可以加下