A Context-aware Natural Language Generator for Dialogue Systems

TL;DR

This work introduces a context-aware natural language generator for dialogue systems that conditions generation on the preceding user utterance to achieve entrainment and more contextually appropriate responses. It extends a baseline seq2seq with attention with a separate context encoder, input prepending, and an $n$-gram match reranker, all trained end-to-end. Automatic metrics show substantial BLEU gains when combining context mechanisms with reranking, and human judgments indicate a modest but significant improvement in naturalness. The approach moves beyond rule-based entrainment by learning contextual adaptation from data and is positioned for evaluation in live SDS settings.

Abstract

We present a novel natural language generation system for spoken dialogue systems capable of entraining (adapting) to users' way of speaking, providing contextually appropriate responses. The generator is based on recurrent neural networks and the sequence-to-sequence approach. It is fully trainable from data which include preceding context along with responses to be generated. We show that the context-aware generator yields significant improvements over the baseline in both automatic metrics and a human pairwise preference test.

Figures (2)

• Figure 1: An example of NLG input and output, with context-aware additions.
• Figure 2: The base Seq2seq generator (black) with our improvements: prepending context (green) and separate context encoder (blue).