Scholar - SciOpen

To enhance the quality of test paper generation by teachers and the efficiency of teaching as well as optimize teaching resources, numerous scholars at home and abroad have conducted in-depth research on automatic test paper generation techniques. Previously proposed algorithms for automatic test paper generation include random selection, backtracking-based automatic test paper generation, and optimization algorithm–based automatic test paper generation. These traditional algorithms did not consider the overall cognitive level of classes, resulting in low efficiency in test paper generation and uneven quality of test papers, which affected the accurate assessment of students’ learning outcomes. To address these issues, this paper proposes an automatic test paper generation method based on cognitive diagnosis and a large language model (LLM)-optimized genetic algorithm.

[Methods]

First, leveraging the powerful data processing and analysis capabilities of pretrained LLMs, this paper proposes the use of fine-tuning techniques and question-answering mechanisms of LLMs for automatic test paper generation. This study explores the impact of different prompts, model parameter sizes, and fine-tuning techniques on automatic test paper generation tasks, ultimately selecting the Qwen LLM with 7 billion parameters and using the QLoRA fine-tuning technique for automatic test paper generation. This method can generate test papers efficiently, but the quality of the test papers is somewhat inferior to that of traditional genetic algorithms. Therefore, an improved automatic test paper generation method is proposed based on cognitive diagnosis and an LLM-optimized genetic algorithm. The test papers generated via the Qwen LLM are used as the initial population, and a fitness calculation method for test paper generation is developed. The genetic operators are designed with the idea of “elitism”, and the design of the selection, crossover, and mutation operators is improved to always retain the best individuals in the population.

[Results]

Experimental results show that the traditional genetic algorithm takes 34.586 s to generate a test paper with a success rate of 83%, while the LLM-based automatic test paper generation algorithm completes the task in 8.565 s but with a slight decrease in the success rate to 81%. In comparison, the proposed genetic algorithm–based automatic test paper generation method combining cognitive diagnosis and LLM optimization not only significantly improves the efficiency of test paper generation to just 0.673 s but also achieves a success rate of 100%, fully meeting the needs of practical applications. This method significantly outperforms the aforementioned two algorithms, demonstrating dual advantages in efficiency and success rate.

[Conclusions]

By combining emerging LLM technology, cognitive diagnosis methods, and genetic algorithms for the requirements of automatic test paper generation, this method improves the efficiency of test paper generation while ensuring its quality and success rate. This innovative combination effectively circumvents the limitations of traditional test paper generation methods, providing strong technical support for teachers. It not only marks a significant advancement in information technology in the field of test paper generation but also adds robust momentum to educational reform, leading us toward a more efficient and intelligent educational future.

Regular Paper Issue

Self-Supervised Music Motion Synchronization Learning for Music-Driven Conducting Motion Generation

Fan Liu, De-Long Chen, Rui-Zhi Zhou, Sai Yang, Feng Xu

Journal of Computer Science and Technology 2022, 37(3): 539-558

Published: 31 May 2022

Abstract Collect Collected

The correlation between music and human motion has attracted widespread research attention. Although recent studies have successfully generated motion for singers, dancers, and musicians, few have explored motion generation for orchestral conductors. The generation of music-driven conducting motion should consider not only the basic music beats, but also mid-level music structures, high-level music semantic expressions, and hints for different parts of orchestras (strings, woodwind, etc.). However, most existing conducting motion generation methods rely heavily on human-designed rules, which significantly limits the quality of generated motion. Therefore, we propose a novel Music Motion Synchronized Generative Adversarial Network (M²S-GAN), which generates motions according to the automatically learned music representations. More specifically, M²S-GAN is a cross-modal generative network comprising four components: 1) a music encoder that encodes the music signal; 2) a generator that generates conducting motion from the music codes; 3) a motion encoder that encodes the motion; 4) a discriminator that differentiates the real and generated motions. These four components respectively imitate four key aspects of human conductors: understanding music, interpreting music, precision and elegance. The music and motion encoders are first jointly trained by a self-supervised contrastive loss, and can thus help to facilitate the music motion synchronization during the following adversarial learning process. To verify the effectiveness of our method, we construct a large-scale dataset, named ConductorMotion100, which consists of unprecedented 100 hours of conducting motion data. Extensive experiments on ConductorMotion100 demonstrate the effectiveness of M²S-GAN. Our proposed approach outperforms various comparison methods both quantitatively and qualitatively. Through visualization, we show that our approach can generate plausible, diverse, and music-synchronized conducting motion.

Total 2