A Text Style and Theme-guided Framework for Children's News Generation with Large Language Models

Mainstream news content is mostly designed for adults, which is not easy for children to understand. This paper proposes a topic-based children's news discourse structure framework (TNC-LLM). The framework integrates two core modules, Text Style Definition (TSD) and Theme Category Definition (TCD). The TSD module adopts various machine learning algorithms to analyze the characteristics of text style and paragraph layout from different granularities, and the TCD module conducts content analysis for different topics to reveal the writing characteristics and content tendency of children's news and to ensure that the content is educational and appropriate. The experiments in this paper mainly evaluate the performance of four models, including ChatGPT3.5, in converting adult news into child-oriented news. The experimental results show that TNC-LLM significantly improves the key dimensions of content accuracy, text interest, and educational appropriateness in the children's news content generation task.