On the Effectiveness of Offline RL for Dialogue Response Generation

TL;DR

This paper evaluates offline reinforcement learning methods for dialogue response generation, demonstrating their effectiveness over traditional teacher forcing without instability or increased training costs.

Contribution

It provides a comprehensive evaluation of offline RL techniques for dialogue, showing their advantages over teacher forcing in multiple settings.

Findings

Offline RL improves dialogue response quality.

No training instability observed with offline RL.

Performance gains achieved without extra training costs.

Abstract

A common training technique for language models is teacher forcing (TF). TF attempts to match human language exactly, even though identical meanings can be expressed in different ways. This motivates use of sequence-level objectives for dialogue response generation. In this paper, we study the efficacy of various offline reinforcement learning (RL) methods to maximize such objectives. We present a comprehensive evaluation across multiple datasets, models, and metrics. Offline RL shows a clear performance improvement over teacher forcing while not inducing training instability or sacrificing practical training budgets.