[论文笔记]A Scalable Neural Shortlisting-Reranking Approach for Large-Scale Domain Classification in NLU

这篇论文提出一种大规模训练的智能对话模型。模型有两层结构,第一层(Shortlisting)用来选出几个几率最大的domain,并判别intent,第二层(Hypothesis Reranker)用历史的记录,找出几率最大的intent。

[论文笔记]A Scalable Neural Shortlisting-Reranking Approach for Large-Scale Domain Classification in NLU

[论文笔记]A Scalable Neural Shortlisting-Reranking Approach for Large-Scale Domain Classification in NLU

Shortlisting模型,主要是用character和word-level的资讯,放入LSTM中,去选出前k几率大的domain。再分别依据不同的domain去判断intent。








[论文笔记]A Scalable Neural Shortlisting-Reranking Approach for Large-Scale Domain Classification in NLU

Hypothesis Reranker模型,也是利用LSTM来计算。这边的关键在,hypothesis representation上,文章利用NLU interpretation, user preferences 和 domain index来表示,其中user preferences便是使用者的历史资讯。