Polite Dialogue Generation Without Parallel Data

TL;DR

This paper introduces three weakly-supervised models for generating polite dialogue responses without needing parallel datasets, achieving high politeness levels while maintaining dialogue relevance through innovative training techniques.

Contribution

The paper presents three novel weakly-supervised models—Fusion, LFT, and Polite-RL—for stylistic dialogue generation that do not require parallel data, advancing politeness control in conversational agents.

Findings

LFT and Polite-RL produce more polite responses without losing relevance.

Fusion and retrieval models achieve politeness but with less relevance.

Human evaluation confirms effectiveness of proposed models in politeness control.

Abstract

Stylistic dialogue response generation, with valuable applications in personality-based conversational agents, is a challenging task because the response needs to be fluent, contextually-relevant, as well as paralinguistically accurate. Moreover, parallel datasets for regular-to-stylistic pairs are usually unavailable. We present three weakly-supervised models that can generate diverse polite (or rude) dialogue responses without parallel data. Our late fusion model (Fusion) merges the decoder of an encoder-attention-decoder dialogue model with a language model trained on stand-alone polite utterances. Our label-fine-tuning (LFT) model prepends to each source sequence a politeness-score scaled label (predicted by our state-of-the-art politeness classifier) during training, and at test time is able to generate polite, neutral, and rude responses by simply scaling the label embedding by the corresponding score. Our reinforcement learning model (Polite-RL) encourages politeness generation by assigning rewards proportional to the politeness classifier score of the sampled response. We also present two retrieval-based polite dialogue model baselines. Human evaluation validates that while the Fusion and the retrieval-based models achieve politeness with poorer context-relevance, the LFT and Polite-RL models can produce significantly more polite responses without sacrificing dialogue quality.