Abstract
Natural language generation (NLG) is a critical component in spoken dialoguesystems. Classic NLG can be divided into two phases: (1) sentence planning:deciding on the overall sentence structure, (2) surface realization:determining specific word forms and flattening the sentence structure into astring. Many simple NLG models are based on recurrent neural networks (RNN) andsequence-to-sequence (seq2seq) model, which basically contains anencoder-decoder structure; these NLG models generate sentences from scratch byjointly optimizing sentence planning and surface realization using a simplecross entropy loss training criterion. However, the simple encoder-decoderarchitecture usually suffers from generating complex and long sentences,because the decoder has to learn all grammar and diction knowledge. This paperintroduces a hierarchical decoding NLG model based on linguistic patterns indifferent levels, and shows that the proposed method outperforms thetraditional one with a smaller model size. Furthermore, the design of thehierarchical decoding is flexible and easily-extensible in various NLG systems.