您的位置:首页 > 产品设计 > UI/UE

Online Sequence-to-Sequence Active Learning for Open-Domain Dialogue Generation

2017-06-28 11:36 633 查看

introduction

Seq2Seq drawback: generate short, dull and inconsistent responses.

DRL:

reward function: most hand-crafted

this paper propose an end-to-end, neural network based generative conversational model that learns open-domain conversation skills via online interaction with human users.

Model

Offline Two-Phase Supervised Learning

responses are short and dull

use Online Active Learning to tackle this issue

Online Active Learning

interacts with real users and learns incrementally from their feedback at each turn of dialog

datasets: considerably small (300K and 8K

resp.)
内容来自用户分享和网络整理,不保证内容的准确性,如有侵权内容,可联系管理员处理 点击这里给我发消息
标签: 
相关文章推荐