Fine-grained Controllable Text Generation through In-context Learning with Feedback

TL;DR

This paper introduces a novel in-context learning approach for fine-grained controllable text generation, enabling rewriting sentences to match specific linguistic features without finetuning, effective even with limited data.

Contribution

It presents a new method that leverages in-context learning for controllable text rewriting based on linguistic features, outperforming previous finetuning-based approaches.

Findings

Achieves accurate sentence rewriting matching specific linguistic features.

Matches state-of-the-art performance on school grade level rewriting.

Effective in low-data scenarios.

Abstract

We present a method for rewriting an input sentence to match specific values of nontrivial linguistic features, such as dependency depth. In contrast to earlier work, our method uses in-context learning rather than finetuning, making it applicable in use cases where data is sparse. We show that our model performs accurate rewrites and matches the state of the art on rewriting sentences to a specified school grade level.