FUDGE: Controlled Text Generation With Future Discriminators

TL;DR

FUDGE is a modular method for controlled text generation that adjusts a base model’s output probabilities using learned attribute predictors, enabling flexible conditioning on multiple attributes across various tasks.

Contribution

FUDGE introduces a novel approach that uses future discriminators to condition text generation on desired attributes with minimal access to the base model's internals.

Findings

Improves control over text attributes in multiple tasks

Easily combines multiple attribute predictors

Achieves measurable gains in poetry, topic control, and translation

Abstract

We propose Future Discriminators for Generation (FUDGE), a flexible and modular method for controlled text generation. Given a pre-existing model G for generating text from a distribution of interest, FUDGE enables conditioning on a desired attribute a (for example, formality) while requiring access only to G's output logits. FUDGE learns an attribute predictor operating on a partial sequence, and uses this predictor's outputs to adjust G's original probabilities. We show that FUDGE models terms corresponding to a Bayesian decomposition of the conditional distribution of G given attribute a. Moreover, FUDGE can easily compose predictors for multiple desired attributes. We evaluate FUDGE on three tasks -- couplet completion in poetry, topic control in language generation, and formality change in machine translation -- and observe gains in all three tasks.