Prompt-Based Editing for Text Attribute Transfer

Abstract

Text attribute transfer (TAT) is a natural language processing task that involves transforming some attributes of a given text while preserving other attributes. Recently, prompting approaches have been explored in TAT with the emergence of various pretrained language models (PLMs), where a textual prompt is used to query a PLM to generate attribute-transferred texts word by word in an autoregressive manner. However, such a generation process is less controllable and early prediction errors may negatively affect future word predictions. Consequently, these issues will lead to low performance in general. In this thesis, we propose a prompt-based editing approach to text attribute transfer. Specifically, we prompt a PLM for attribute classification and use the classification probability to compute an attribute score. Then, we perform discrete search with word-level editing to maximize a comprehensive scoring function for an attribute transfer task. In this way, we transform a prompt-based generation problem into a classification one, which does not suffer from the error accumulation problem and is more controllable than the autoregressive generation of sentences. In our experiments, we perform automatic evaluation on four attribute transfer benchmark datasets, and we show that our approach largely outperforms the existing systems that have 20 times more parameters. Additional empirical analyses further demonstrate the effectiveness of our approach.