Human-centric Dialog Training via Offline Reinforcement Learning

TL;DR

This paper introduces a novel offline reinforcement learning approach to train dialog models using human feedback, addressing exploration and overestimation challenges, and demonstrates improved conversational quality in real-world tests.

Contribution

It develops a new offline RL algorithm with KL-control and pessimistic strategies, enabling effective training of dialog models from static human feedback datasets.

Findings

Significant improvement in dialog quality over existing offline RL methods.

Effective use of implicit human feedback cues as reward signals.

Validated approach with 80 user ratings in open-domain conversations.

Abstract

How can we train a dialog model to produce better conversations by learning from human feedback, without the risk of humans teaching it harmful chat behaviors? We start by hosting models online, and gather human feedback from real-time, open-ended conversations, which we then use to train and improve the models using offline reinforcement learning (RL). We identify implicit conversational cues including language similarity, elicitation of laughter, sentiment, and more, which indicate positive human feedback, and embed these in multiple reward functions. A well-known challenge is that learning an RL policy in an offline setting usually fails due to the lack of ability to explore and the tendency to make over-optimistic estimates of future reward. These problems become even harder when using RL for language models, which can easily have a 20,000 action vocabulary and many possible reward functions. We solve the challenge by developing a novel class of offline RL algorithms. These algorithms use KL-control to penalize divergence from a pre-trained prior language model, and use a new strategy to make the algorithm pessimistic, instead of optimistic, in the face of uncertainty. We test the resulting dialog model with ratings from 80 users in an open-domain setting and find it achieves significant improvements over existing deep offline RL approaches. The novel offline RL method is viable for improving any existing generative dialog model using a static dataset of human feedback.