Attribute Alignment: Controlling Text Generation from Pre-trained Language Models

TL;DR

This paper introduces a flexible method for controlling text generation in pre-trained language models by aligning attribute representations, achieving improved attribute control without altering the original model.

Contribution

It presents a novel alignment-based approach for attribute control in text generation that outperforms previous discriminator-based methods.

Findings

Significant improvements in sentiment and topic control accuracy.

Retention of fluency and diversity in generated texts.

No need to modify original language model parameters.

Abstract

Large language models benefit from training with a large amount of unlabeled text, which gives them increasingly fluent and diverse generation capabilities. However, using these models for text generation that takes into account target attributes, such as sentiment polarity or specific topics, remains a challenge. We propose a simple and flexible method for controlling text generation by aligning disentangled attribute representations. In contrast to recent efforts on training a discriminator to perturb the token level distribution for an attribute, we use the same data to learn an alignment function to guide the pre-trained, non-controlled language model to generate texts with the target attribute without changing the original language model parameters. We evaluate our method on sentiment- and topic-controlled generation, and show large performance gains over previous methods while retaining fluency and diversity.