Abstract
Human writing often exhibits a range of styles and levels of sophistication. However, automated text generation systems typically lack the nuanced understanding required to produce refined and elegant prose. Due to the inherent one-to-many relationship between inputs and outputs in natural language generation tasks, achieving annotator consistency is challenging. This complexity makes the annotation process considerably more difficult compared to tasks focused on natural language understanding. Our study focuses on the typical task of text refinement, which faces annotation difficulties, aiming to generate sentences with more elegant expressions while preserving the original semantics of the input sentence. This paper proposes a semi-automatic data construction method that combines auto-generation with human judgment. Initially, this method translates collected sentences containing elegant expressions into ordinary expressions through back translation. Subsequently, in an iterative quality control process, data filtering and human judgment are introduced to screen the auto-generated data based on quality standards, resulting in a large-scale text refinement dataset. By replacing manual annotation with human judgment and involving only a small amount of data for human judgment in each iteration, this method significantly reduces annotation difficulty and workload. With minimal human effort, it acquires a substantial amount of labeled data for text refinement, laying a foundation for further research in the field.