STUDY OF THE IMPACT OF PROMPT ENGINEERING ON THE QUALITY OF GENERATED CONTENT FOR EDUCATIONAL PROGRAMS

Abstract

The presented study is devoted to the analysis of the influence of prompt engineering on the quality of content generated by language models for educational programs. The paper provides a comprehensive analysis of the effectiveness of various prompting methods in interpreting educational program components using a wide range of modern language models - from compact to large-scale. Based on the processing of a large corpus of 8269 educational programs, a structural and semantic gradation of their components according to the degree of interpretability by artificial intelligence was identified. Highly interpretable components (goals of educational programs, information about disciplines, generated learning outcomes) demonstrate consistently high values ​​of cosine similarity, while low-interpretable components (Atlas of the profession) show consistently low values. The study found different efficiency of prompting methods for different components: few-shot learning is most effective for the goals of educational programs, and chain of thought is most effective for professional standards. The identified "specialized advantage" is of particular scientific value: the model adapted using the LoRA technique significantly outperforms larger non-specialized analogs in interpreting educational content, which refutes the hypothesis of a linear relationship between the parametric volume of the model and the quality of its work. The obtained results form the basis for developing optimal strategies for using language models in educational analytics and open up prospects for further research in the field of interpreting complex semantic structures of educational programs by artificial intelligence. The practical significance of the study lies in the development of recommendations for choosing optimal models and prompting methods for various tasks of analyzing educational programs, as well as strategies for specialized customization of language models to improve the efficiency of interpreting specific components.