Balancing AI and Human Judgment for Text Refinement: A Step Toward Intercultural Competence in Language Processing

Human writing often exhibits a variety of styles and levels of sophistication, yet automated text generation systems typically struggle to produce nuanced and culturally sensitive prose. Achieving a balance between AI-driven auto-generation and human judgment is essential for refining text in ways that respect diverse cultural contexts. This study addresses the challenges inherent in text refinement, a task that is complex due to the one-to-many relationship between inputs and outputs in natural language generation, making annotation consistency difficult. Our research proposes a semi-automatic data construction method that combines the strengths of both AI and human judgment to generate more elegant expressions while preserving the original semantics and cultural relevance of the input sentences. Initially, the method employs back translation to convert elegant expressions into more neutral ones, followed by an iterative quality control process. This process involves data filtering and human judgment to ensure that the auto-generated text adheres to cultural norms and quality standards. By involving minimal human effort in each iteration, this approach significantly reduces the annotation workload while producing a large-scale, high-quality dataset for text refinement. Ultimately, this method contributes to the development of more culturally aware AI systems that facilitate ethical and effective intercultural communication in the age of globalization.