Stochastic Language Generation in Dialogue using Recurrent Neural Networks with Convolutional Sentence Reranking

TL;DR

This paper introduces a neural network-based language generator for dialogue systems that learns directly from data, producing high-quality, varied responses without predefined rules, and outperforms previous methods in objective and human evaluations.

Contribution

It proposes a novel joint recurrent and convolutional neural network model for dialogue generation that eliminates the need for semantic alignments or predefined grammars.

Findings

Outperforms previous methods on objective metrics

Produces linguistically varied and high-quality responses

Preferred by human judges over n-gram and rule-based systems

Abstract

The natural language generation (NLG) component of a spoken dialogue system (SDS) usually needs a substantial amount of handcrafting or a well-labeled dataset to be trained on. These limitations add significantly to development costs and make cross-domain, multi-lingual dialogue systems intractable. Moreover, human languages are context-aware. The most natural response should be directly learned from data rather than depending on predefined syntaxes or rules. This paper presents a statistical language generator based on a joint recurrent and convolutional neural network structure which can be trained on dialogue act-utterance pairs without any semantic alignments or predefined grammar trees. Objective metrics suggest that this new model outperforms previous methods under the same experimental conditions. Results of an evaluation by human judges indicate that it produces not only high quality but linguistically varied utterances which are preferred compared to n-gram and rule-based systems.