Bridging the Gap Between Training and Inference of Bayesian Controllable Language Models

TL;DR

This paper introduces the Gemini Discriminator to improve the training-inference alignment in Bayesian Controllable Language Models, enhancing controllable text generation without significant computational overhead.

Contribution

The paper proposes a novel Gemini Discriminator that reduces training-inference mismatch in BCLMs, achieving state-of-the-art results in sentiment and topic control tasks.

Findings

Achieved new state-of-the-art results in sentiment control.

Achieved new state-of-the-art results in topic control.

Demonstrated effectiveness with low computational cost.

Abstract

Large-scale pre-trained language models have achieved great success on natural language generation tasks. However, it is difficult to control the pre-trained language models to generate sentences with the desired attribute such as topic and sentiment, etc. Recently, Bayesian Controllable Language Models (BCLMs) have been shown to be efficient in controllable language generation. Rather than fine-tuning the parameters of pre-trained language models, BCLMs use external discriminators to guide the generation of pre-trained language models. However, the mismatch between training and inference of BCLMs limits the performance of the models. To address the problem, in this work we propose a "Gemini Discriminator" for controllable language generation which alleviates the mismatch problem with a small computational cost. We tested our method on two controllable language generation tasks: sentiment control and topic control. On both tasks, our method reached achieved new state-of-the-art results in automatic and human evaluations.