ADL100(2)-Wuwei-检索式对话
作者:互联网
1. 检索
1)(单轮回复选择)匹配模型
a. sentence 生成定长向量 进行匹配 f(g(Q), g(R))
f和g各种变种
f: MLP Neural Tensor Cosine
g: CNN LSTM+Att
b. query 与 candidate response 匹配
g: Interaction Representation(Att / Sim)
c. 效力和效率
2)多轮回复选择
a. 挑战:
层次结构:word->utterance->session
信息冗余
逻辑:句子的顺序?词句长期依赖,适合句子的约束
3)knowledge enhanced
方向:deep wide pre-training -> hhh yes
Learning method
from unlabelled data (weak supervision) (acl 18)
denoise with Peer (co-teaching) (两个网络互相加权,互相trach acl19)
teach with dynamic margin
teach with dynamic instance weight
teach with dynamic data curriculum(选高置信度的)
Summary:
matching model
two framework
vec -> match
rep -> math -> aggregation
external knowledge
Deep, wide(better representation), learning method
标签:检索,wide,knowledge,dynamic,ADL100,Att,teach,Wuwei,method 来源: https://www.cnblogs.com/zh-liu/p/ADL100-2-Retrival-Chat.html