What is shared today is:AI large model seriesIn-depth Research Report:AI Large Model Topic: Research Report on Large Models for Education
Report produced by the Institute of Digital Education, Chinese Academy of Educational Sciences).
Report total:Page.
Featured Report**: The School of Artificial Intelligence
At present, the rapid development of a new generation of artificial intelligence technology has attracted widespread attention around the world. In November 2023, OpenAI held its first developer conference and released GPTS and the latest development technology, allowing everyone to create custom large models through natural language human-computer dialogue. At the same time, domestic technology companies such as Alibaba, Huawei, and iFLYTEK have successively released a variety of artificial intelligence models to provide Chinese solutions. In the future, the AI model will be deeply integrated into all fields and links, enabling the intelligent upgrading of thousands of industries and helping social productivity to leap. From a general model to a special model for education, it is an important trend for the deepening and development of artificial intelligence model technology. This is not just a fine-tuning and optimization based on a general large model, but a major innovation based on the education scenario and an open model architecture. The education-specific model refers to a multi-level open technology architecture based on the general model, driven by a variety of educational digital applications, and provides professional capabilities to teachers, students and social learners through a unified interactive dialogue interface. It has a wealth of diverse educational expertise and is constantly being upgraded in its applications.
"Large models" are AI models with billions to tens of billions or more trainable parameters, which are the product of a combination of technologies such as deep learning, GPU hardware, and large-scale datasets. The powerful capabilities of large models are essentially the result of "quantitative change causing qualitative change" in deep learning. When the number of model parameters reaches a certain scale, the model accuracy undergoes a qualitative leap, a process called "emergent ability", that is, the ability to automatically learn from the original training data and discover new, higher-level features and patterns. These capabilities are prominently manifested in the ability to understand general user intent, the ability to continuously dialogue in a large range of contexts, the ability to intelligently correct interactions, the ability to edit, classify, and summarize content, the ability to generate new content, and the ability to multimodal.
As one of the key R&D areas, the education model is an integrated application of large model technology, knowledge base technology and various intelligent education technologies, which will train and form new educational scenarios in the education process and realize the two-way construction of human learning and machine learning. At present, there are two main technical routes for education-specific large models: one is to directly call the general large model to make it have certain professional capabilities through fine-tuning or prompt learning; The second is to use professional data in the field of education to train large models for solving educational tasks. Some progress has been made in both technical routes, but the implementation effect still needs to be improved. Due to the lack of sufficient professional data training, insufficient in-depth knowledge and weak intelligence, it is difficult to flexibly deal with complex and changeable educational tasks.
How to develop a large model for education, the solution lies in the integration of the two technical routes. This is not a simple add-on, but a technological breakthrough is achieved by establishing a new open technology architecture that systematically integrates general problem-solving capabilities and education-specific problem-solving capabilities. Specifically, it is necessary to combine the general large model and professional data in the field of education, break the data island, and continuously obtain data from normalized education applications through open data interfaces. It is necessary to use the expert knowledge base as a supplement to the large model, and consciously "teach" the subject knowledge and various rules in education and teaching to the large model; It is necessary to integrate and apply all kinds of intelligent education technologies, and integrate all kinds of intelligent education technologies that have been developed into the special large model for education.
Large models bring great potential and possibilities for education in terms of speaking practice, math learning, sentiment analysis, and personalized recommendations. We have sorted out five typical application cases of large models of education majors, and conducted an in-depth analysis of their development background, application scenarios, and design characteristics.
From the perspective of application scenariosXinghuo Language Companion is mainly used for language learning, supports real-time translation of multilingual text, voice and **, and can correct grammatical errors and provide oral sparring. EMOGPT is used to provide mental health services, identify and respond to user emotions, and provide ongoing psychological support. Aimed at math enthusiasts and research institutions around the world, MATHGPT provides problem-solving and problem-solving algorithms to support users in math problem solving and practice. Zhihai-Sanle is used for AI professional education, providing functions such as search engine, calculation engine, and local knowledge base, and supporting intelligent Q&A and test question generation. Khanmigo provides learners with personalized learning solutions through conversational AI chatbots, while also supporting career planning services, teaching method coaching optimization, and more.
Judging by technological progressThe education-specific large model has shown advantages in terms of model performance, application scenarios, and technical characteristics, covering most disciplines, mainly focusing on self-directed learning scenarios, including knowledge question answering, language learning, learning guidance, and teaching assistance. On the technical route, the path of "general + fine-tuning" has proven its effectiveness, and many technical solutions are based on general large models to achieve effective answers to specific subject knowledge through instruction fine-tuning and other methods.
Judging from the existing shortcomingsThe existing education-specific large models still have limitations in terms of accuracy, diversity of teaching content, support for core education scenarios, and inclusion of learner diversity, with high error rates and lack of empathetic comprehension ability, mainly focusing on subject knowledge teaching and test-oriented education situations, and still lacking in interdisciplinary learning, students' comprehensive ability and higher-order thinking cultivation. Mainly focusing on supporting self-directed learning, how to give full play to the role of large models in scenarios such as real classrooms, peer collaboration, and hybrid teaching has not been effectively explored.
In short, the application of large models in the field of education has made significant progress, but it is still facing practical problems, and it is necessary to further improve the quality and scale of training data, especially to deeply embed advanced education concepts, in-depth knowledge of education and the real needs of education core scenarios into technical design, and carry out multiple rounds of iterations based on user feedback to form a more intelligent and flexible large model for education.
A case study of a large model for education.
Report total:Page.
Featured Report**: The School of Artificial Intelligence