Sequence-to-Sequence Generation for Spoken Dialogue via Deep Syntax Trees and Strings

TL;DR

This paper introduces a sequence-to-sequence natural language generator that produces both syntax trees and strings from dialogue acts, demonstrating that a joint generation approach outperforms traditional two-step methods with limited training data.

Contribution

It presents a novel joint sequence-to-sequence model for dialogue generation that outperforms separate planning and realization stages, even with minimal training data.

Findings

Joint approach surpasses state-of-the-art n-gram scores

Joint model produces more relevant outputs

Effective with limited training data

Abstract

We present a natural language generator based on the sequence-to-sequence approach that can be trained to produce natural language strings as well as deep syntax dependency trees from input dialogue acts, and we use it to directly compare two-step generation with separate sentence planning and surface realization stages to a joint, one-step approach. We were able to train both setups successfully using very little training data. The joint setup offers better performance, surpassing state-of-the-art with regards to n-gram-based scores while providing more relevant outputs.