Response Document Generation for Accommodation Reviews Using Hierarchical Encoder-Decoder Models

Abstract

We examine a system for presenting the draft text to writers of response documents. Since a document is a set of sentences, it is necessary to handle a symbol sequence longer than a sentence conversion such as a conventional machine translation. We propose a method using pre-trained sentence embedding as input of hierarchical RNN encoder-decoder model. It can generate a more detailed response document than the simple non-hierarchical RNN encoder-decoder model.