Abstract: Innovative automated scoring(IAS) is crucial in education.Traditional scoring is subjective,inefficient,and lacks uniform standards.The fast progress of large language models offers new solutions.This study creates a high-quality dataset WAIS and presents a semantic-driven hierarchical topic extraction algorithm.Through four phases-semantic chunking,basic topic extraction,optimized analysis,and topic fusion-the algorithm improves the model’s ability to extract themes from student answers,enabling automatic topic extraction.It offers a solid basis for automated scoring and establishes an explainable cognitive framework for subsequent scoring.The study compares three prompting strategies:Zero-shot,Few-shot,and Chain-of-Thought(CoT),and evaluates them using several pre-trained models.Results show CoT is superior.The DeepSeek-R1 model achieves 68% accuracy.After fine-tuning,the smaller-parameter model Qwen1.5-7B reaches 83% accuracy,even slightly surpassing the larger-parameter model using only the prompt in innovative scoring tasks.This indicates that using large language models for innovative automated scoring is feasible and has great potential for development.

Key words: Large language models, Innovation, Automated scoring, Prompt engineering, Supervised fine-tuning