A discipline structure optimization system with size model collaboration

Abstract

The present application belongs to the technical field of subject structure system optimization, and particularly relates to a subject structure optimization system with large and small models cooperating. The system comprises a system interface, a database, a large model and a small model. The large model is used to undertake complex tasks such as long text semantic understanding and cross-disciplinary correlation analysis, and to perform deep analysis, logical reasoning and explanation on the data preliminarily recalled and screened by the small model. The small model is used to determine a strong correlation subject list of a specified subject by analyzing the correlation between subjects according to subject comprehensive descriptions of different subjects, or to determine a strong correlation subject list of a specified researcher or a strong correlation researcher list of a specified subject according to research information of researchers. The present application reduces the time and cost of the system as a whole by fully combining the logical decision and data analysis capabilities of a general large model and the good fitting effect of a special small model on specific tasks and data distribution.

Description

Technical Field

[0001] This invention belongs to the field of discipline structure system optimization technology, specifically relating to a discipline structure optimization system that coordinates large and small models. Background Technology

[0002] With the increasing accessibility of higher education, the number of disciplines and research personnel in universities has exploded. However, this blind expansion has led to redundancy in disciplines, a dispersion of research directions among researchers, and makes it difficult for universities to develop representative and strong disciplines.

[0003] Optimizing the disciplinary structure through abolition and resource reallocation has become a key focus for universities to enhance their strength. For example, in 2018, Zhejiang University stopped enrolling undergraduate students in Communication Engineering, and related research directions and enrollment plans were merged into the Information Engineering major; in 2020, Tsinghua University stopped enrolling undergraduate students in Accounting, and related research directions and enrollment plans were merged into the Economics and Finance major; in the same year, Fudan University stopped enrolling master's students in Software Engineering, and related research directions and enrollment plans were merged into Computer Science and Technology; in 2022, Liaoning University abolished the School of Public Basic Courses, and transferred the English-related faculty members of the School of Public Basic Courses to the School of Foreign Studies, and so on.

[0004] Existing discipline structure optimization work is often based on people, and has made a series of attempts around national strategic needs, industrial structure, etc., combined with case analysis, interviews, text analysis and other methods. Although human-driven discipline structure optimization can complete the task of discipline abolition and resource reallocation to a certain extent, it inevitably has the following problems when faced with complex discipline data: (1) Lack of professional knowledge. Discipline structure optimization scenarios are highly complex, with many professional terms in different disciplines and frequent overlap between research directions. Relevant decision-makers have difficulty mastering the vast amount of discipline knowledge, which leads to the difficulty in achieving the ideal effect of discipline merging or the allocation of scientific researchers; (2) Subjective bias exists. Subjective bias is difficult to avoid in the human decision-making process, especially when professional knowledge is lacking. Relevant decision-makers often associate existing knowledge to make decisions on unknown content, which will bring huge uncertainty to discipline structure optimization; (3) Huge consumption of human resources. When realizing the fine-grained allocation of scientific researchers under the abolished discipline, relevant decision-makers need to combine data such as projects, papers, and patents to analyze the research direction of each scientific researcher, and then complete the matching with the research direction of the discipline. When the scale of scientific researchers and disciplines is large, mechanical labor will consume huge human resources.

[0005] General-purpose large-scale models, exemplified by DeepSeek, have been successfully applied to all aspects of daily life due to their powerful understanding, generation, and logical reasoning capabilities. Corresponding to the adjustment and optimization of the disciplinary structure in universities, large-scale models, pre-trained on massive amounts of data, can cover the professional knowledge of some disciplines. Furthermore, optimizing the disciplinary structure based on large-scale models will greatly save human resources; currently widely relied upon large-scale pre-trained models have advantages in the breadth of knowledge coverage and reasoning ability. While large-scale models alleviate some of the problems in human-driven disciplinary structure optimization, the highly specialized application scenarios and massive data scales pose significant challenges to model performance. When the context of disciplinary optimization input to a large-scale model is too long (e.g., the large-scale model judges the relevance of all disciplines based on information such as projects and papers), the large-scale model will struggle to capture key content, leading to unexpected output results.

[0006] Meanwhile, due to the scarcity of corpora in some disciplines, large models do not learn enough about these disciplines during the pre-training phase, resulting in limited understanding of their specialized terminology and serious illusion problems. Fine-tuning large models to adapt to discipline structure optimization scenarios not only places high demands on data annotation but also severely reduces the model's general capabilities, such as logical reasoning. Furthermore, the massive parameter scale of large models not only slows down inference speed and lengthies the inference process, but also incurs high computational costs, often requiring high-performance hardware clusters, leading to high single-run costs, which seriously affects the user experience of related systems. In practical applications, there is a clear contradiction between the inference speed of large models and real-time requirements. For example, in the process of discipline structure optimization, rapid calculation and feedback are often required for dynamically updated data, but the lengthy inference chain of large models cannot meet the needs of real-time response, and the high computational cost also limits their sustainability in large-scale applications. Summary of the Invention

[0007] The purpose of this invention is to provide a subject structure optimization system that combines large and small models to solve the problems of high computational overhead, high hardware performance requirements, and slow system response that affect user experience caused by existing systems that only use large models for inference.

[0008] To address the aforementioned technical problems, this invention provides a technical solution for a subject structure optimization system that combines large and small models, as detailed below:

[0009] A subject structure optimization system that integrates large and small models includes a system interface, a database, a large model, and small models. The system interface allows users to select functions and displays the output results of the large and small models accordingly. The database stores subject-related data and researchers' research information. The large model generates the data required for the small model's calculations based on the user-selected functions and the data in the database, and performs in-depth analysis, logical reasoning, and interpretation based on the user-selected functions and the output results of the small model. The small model analyzes the correlation between different subjects based on their comprehensive subject descriptions to determine a list of strongly related subjects for a given subject, or determines a list of strongly related subjects for a given researcher or a list of strongly related researchers for a given subject based on researchers' research information.

[0010] The beneficial effects of the above technical solution are as follows: This invention considers that small-scale models have higher computational efficiency and deployment flexibility compared to large models. Therefore, in addition to large models, small models are also deployed in the entire system. The computational efficiency advantage of small models is utilized to improve the overall system efficiency and reduce the system's hardware performance requirements. Furthermore, considering that small models have limited knowledge reserves, especially in tasks such as interdisciplinary knowledge transfer, complex logical reasoning, and multi-level indicator correlation analysis, they are prone to insufficient accuracy, biased conclusions, or even reasoning errors. Therefore, only some functions are deployed in small models, including using lists of strongly related disciplines for a specific discipline, lists of strongly related disciplines for a specific researcher, or lists of strongly related researchers for a specific discipline. The remaining functions are still implemented by the large model, enabling the large model to handle complex tasks such as long text semantic understanding and interdisciplinary correlation analysis, and to perform in-depth analysis, logical reasoning, and interpretation on the data initially recalled and filtered by the small model. This deployment method based on the collaboration of large and small models leverages the respective advantages of large and small models, forming a complementary and collaborative mechanism, improving system response speed, and enhancing user experience. Overall, this invention reduces the system's time and cost overhead.

[0011] Furthermore, when the user selects the function of subject refinement to reduce the redundancy of subject settings, the data required for the large model to generate the small model includes a comprehensive subject description. The process of generating the comprehensive subject description includes: using a research direction extraction agent built based on the prior knowledge of the large model to read and process the subject-related data stored in the database to generate a list of research directions for the subject; and then using an association analysis agent built based on the logical analysis capability of the large model to semantically combine the list of research directions for the subject to generate a comprehensive subject description.

[0012] Furthermore, the process of analyzing the correlation between disciplines using small models includes: projecting the comprehensive description of each discipline into the latent space using a pre-trained embedding model, and calculating the similarity between disciplines based on the vectors of the comprehensive descriptions of each discipline in the latent space.

[0013] Furthermore, the model used in the small model to determine the list of strongly related disciplines for a given researcher or the list of strongly related researchers for a given discipline is a BERT-based classification model. The BERT-based classification model includes a BERT encoder, an embedding space processing unit, a linear layer, and a Softmax layer connected in sequence.

[0014] Furthermore, when the user selects the coarse-grained merging function under subject refinement, the small model is used to analyze the correlation between the subject to be revoked and other subjects based on the comprehensive subject description of the subject to be revoked and the comprehensive subject description of other sciences input by the user to determine the list of strongly related subjects of the subject to be revoked.

[0015] When the user selects the fine-grained sub-distribution under subject refinement, the small model is used to determine the list of strongly related subjects for a specified researcher based on the researcher's research information. Its input is the researcher's research information, and its output is the predicted list of strongly related subjects.

[0016] When the user selects the discipline leap function, the small model is used to determine the list of strongly related researchers for the specified discipline to be strengthened based on the researchers' research information. Its input is the relevant data of the specified discipline, and its output is the predicted list of strongly related researchers.

[0017] Discipline refinement is used to reduce redundancy in discipline settings, while discipline leapfrogging is used to adjust the disciplines to which researchers belong; coarse-grained merging is used to incorporate abolished disciplines into new disciplines, while fine-grained diversion is used to assign researchers from disciplines to be abolished to new disciplines.

[0018] Furthermore, when the user selects the coarse-grained merging function under subject refinement, the large model provides the following decision-making and analysis process: using a decision-making agent built based on the large model to make a decision interpretation on the final affiliation of the subject to be revoked based on the list of subjects to be revoked and their strongly related subjects;

[0019] When the user selects the fine-grained diversion under the subject refinement function, the big model provides the following decision and analysis process: using the decision-making agent built based on the big model to make a fine-grained analysis and decision on the final affiliation of the researchers under the proposed revoked subject based on the information of the researchers of the proposed revoked subject and the list of strongly related subjects of the researchers of the proposed revoked subject.

[0020] When the user selects the disciplinary leap function, the big model provides the following decision-making and analysis process: using a decision-making agent built on the big model to make resource reallocation decisions based on the specified discipline to be strengthened, its strongly related list of researchers, and the level to be achieved.

[0021] Furthermore, the subject-related data input to the research direction extraction agent is pre-processed subject-related data, and the pre-processing includes noise removal.

[0022] Furthermore, the similarity is cosine similarity.

[0023] Furthermore, the large model is also used to predict the effects of various subject structure optimization schemes proposed by the decision-making agent to help users decide whether to adopt the relevant schemes.

[0024] Furthermore, the system interface is used to display results by calling relevant functions using a visual intelligent agent, which is built based on the tool calling capabilities of a large model; the relevant functions include chord graph creation functions, word cloud creation functions, radar chart creation functions, bar chart creation functions, and heat map creation functions. Attached Figure Description

[0025] Figure 1 This is an overall framework diagram of the method corresponding to the discipline structure optimization system driven by the collaboration of large and small models of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the research direction of this invention: extracting key prompts for intelligent agents.

[0027] Figure 3 This is a schematic diagram of the semantic combination of key prompt words for the research direction of this invention;

[0028] Figure 4 This is a schematic diagram illustrating the research direction of this invention: extracting key prompts for intelligent agents.

[0029] Figure 5 This is a schematic diagram of the key prompts for the current status analysis intelligent agent of this invention;

[0030] Figure 6 This is a schematic diagram of the recall model based on potential information of the present invention;

[0031] Figure 7 This is a schematic diagram of the key prompt words for the decision-making intelligent agent of the present invention;

[0032] Figure 8 This is a schematic diagram of the key prompts for the comparative analysis agent of the present invention;

[0033] Figure 9 This is a class diagram of the visualization tools of this invention;

[0034] Figure 10 This is a sequence diagram of the subject refinement function of the present invention;

[0035] Figure 11 This is a timing diagram of the discipline transition function of the present invention;

[0036] Figure 12 This invention is a chart showing the percentage of different disciplines in the scientific research data of XX University.

[0037] Figure 13 This is a case study diagram of subject refinement based on scientific research data from XX University—coarse-grained merging.

[0038] Figure 14 This invention is a case study diagram of coarse-grained merging of disciplines based on scientific research data from XX University (specific module output).

[0039] Figure 15 This is an interface diagram of the subject structure optimization system based on size model collaborative driving of the present invention;

[0040] Figure 16 This is a comparison chart of the system time overhead of the present invention;

[0041] Figure 17 This is a comparison chart of the system cost of the present invention. Detailed Implementation

[0042] This invention proposes a system design based on a collaborative deployment of large and small models. By leveraging the respective strengths of the large and small models, a complementary and collaborative mechanism is formed. The large model generates the data required for the small model's calculations based on user-selected functions and data from the database, and provides decisions and analyses based on the user-selected functions and the output of the small model. The small model analyzes the correlation between different disciplines based on comprehensive descriptions of various disciplines to determine a list of strongly related disciplines for a given discipline, or determines a list of strongly related disciplines for a given researcher or a list of strongly related researchers for a given discipline based on researchers' research information. The large model ensures the overall performance of the system in terms of knowledge coverage and reasoning accuracy, while the small model improves the overall processing speed and reduces the system's computational and hardware resource consumption.

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0044] To address the challenges of complex disciplinary data, this invention proposes a disciplinary structure optimization system driven by a combination of large and small models. By fully integrating the logical decision-making and data analysis capabilities of a general large model with the excellent fitting performance of specialized small models on specific tasks and data distributions, the system optimizes disciplinary structure based on revocation and resource reallocation. Furthermore, this system, focusing on disciplinary refinement and disciplinary advancement, reduces redundancy in disciplinary settings, rapidly enhances the strength of designated disciplines, and generates multimodal reports to help users understand relevant decisions and their effects. Based on this, the effectiveness of the system has been validated using real-world university data, specifically for disciplinary revocation and resource reallocation.

[0045] This invention focuses on the functions of subject refinement and subject leapfrogging, constructing a subject structure optimization system. Based on the powerful logical capabilities of a general-purpose large model, multiple intelligent agents with distinct functions are built. Furthermore, specialized small models, such as a recall model, are trained based on the good fitting performance of dedicated small models on specific tasks and distributions. Different intelligent agents in this system activate different functions such as data analysis and decision-making through system prompts, and compensate for the disadvantages of the large model, such as difficulty in handling long contexts and high fine-tuning costs, by interacting with the small models. For ease of description, Table 1 defines the main symbols used and their meanings.

[0046] Table 1. Main symbols used and their meanings

[0047]

[0048] This disciplinary structure system, driven by both large and small models, revolves around two main functional modules: disciplinary refinement and disciplinary advancement. Its aim is to reduce redundancy in disciplinary setup and rapidly enhance the strength of specific disciplines. Figure 1 As shown, it consists of the following seven modules:

[0049] 1. Research Direction Extraction Module

[0050] A single discipline in a university typically comprises numerous researchers, including faculty and postdoctoral fellows. This large staff generates a massive amount of research information, such as projects, papers, and patents, posing a significant challenge to model-driven optimization of the discipline structure. Whether it's the overall merging of disciplines or the reallocation of resources based on researchers, the essence lies in the accurate grasp of the discipline's research direction. When extracting research directions based on project, paper, and patent information, existing unsupervised topic modeling techniques struggle to achieve satisfactory results due to the sheer scale of the data and the rarity of some research directions. To address this issue, this invention constructs a research direction extraction agent A based on the extensive prior knowledge of a large model. RThis invention aims to compress massive amounts of university research data and extract research directions from different disciplines. Specifically, it performs preliminary processing on project information by discipline, removing noise such as stop words. Subsequently, research directions are extracted by agent A. R The research information will be further filtered by word to remove unrepresentative subject terms.

[0051]

[0052] In the formula, This is a vocabulary list after preliminary processing of research information such as projects under subject i. P represents the prompt words for extracting agents based on research directions. Some of its content is as follows: Figure 2 As shown.

[0053] Compared to generating research directions for a specific discipline directly from a large model, the research direction extraction method uses agent A. R Having acquired global information, eliminating unreasonable information will greatly improve the model's output. This invention achieves this through the following process for A. R The validity of this can be simply proven.

[0054] The large model predicts the next character based on existing characters. Assuming the number of distinct characters in the large model's character list is M, when directly generating the research direction of subject i from the large model, the entropy for generating the next character is:

[0055]

[0056] In the formula, For each character in the character list, applying the Lagrange multiplier method to formula (2) yields:

[0057]

[0058] For formula (3) Differentiate from λ:

[0059]

[0060] Setting the equation in formula (4) to 0, we get:

[0061]

[0062] From formula (5), The maximum value is logM. Assume... The number of different characters in the model is N. When using a large model... When removing characters from the content, the number of candidate characters changes from M to N. At this point, the entropy for predicting the next character... It is logN. Because Therefore, large-scale model elimination is used. The entropy of irrational research directions will be far less than the entropy of research directions generated based on large models, that is... Therefore, through Eliminate The uncertainty of irrational research directions will be far less than the uncertainty of research directions generated based on large models, using A R Gaining access to research directions from different disciplines will provide a more reliable outcome.

[0063] 2. Association Analysis Module

[0064] Association analysis is the foundation of subject refinement functionality. It calculates the correlation between different subjects using small models to obtain a list of Top-k subjects related to the target subject. This provides potential merger options for disciplines slated for discontinuation. Based on this, a relational analysis intelligent agent is constructed using the logical analysis capabilities of a large-scale model. It provides correlation analysis based on disciplinary research directions to support user decision-making. Specifically, the correlation analysis module is implemented as follows:

[0065] For subject i, the first step is to use a large model to analyze the list of research directions obtained by the data compression module. Semantic combination is used to integrate all research directions in the discipline while avoiding the influence of the order of research directions on the relevance results:

[0066]

[0067] In the formula, For pre-training large models, This is a comprehensive description of the disciplines generated for the large model, where P represents relevant prompt words, and some of its content is as follows: Figure 3 As shown. Among them, Its function is to process subject data, extract features, and prepare for subsequent similarity calculations.

[0068] Subsequently, based on The text uses a pre-trained embedding model The semantic combinations of different disciplines are projected into the latent space, and the relevance of different research directions is calculated based on cosine similarity:

[0069]

[0070] In the formula, Let v represent the relevance between subject i and subject j, and v represent the vector of semantic combination in the latent space.

[0071] It should be noted that the cosine similarity calculation method is used for similarity analysis here. As other implementation methods, other similarity calculation methods in the prior art can also be used, such as distance-based (e.g., Euclidean distance) similarity calculation methods.

[0072] In the subject refinement - coarse-grained merging function (merging revoked subjects into new subjects), the user will specify the subject to be revoked, and according to formula (7), select the k most relevant subjects to form a list of strongly related subjects. This allows users to choose from various options. To assist users in decision-making, this invention constructs an intelligent agent for correlation analysis based on a large model. (Based on an analysis that considers the similarities between various disciplines), this approach provides a specific analysis of the research directions within each discipline to help users understand the relationship between the proposed merging discipline and related disciplines, thereby enabling them to make more informed decisions regarding merging disciplines. Key prompts such as Figure 4 As shown.

[0073] 3. Current Situation Analysis Module

[0074] In the subject refinement function, the current situation analysis aims to provide the specific performance of the subject to be withdrawn under a certain evaluation system, and help the user decide whether to withdraw it. In the subject leap function, the user will specify the subject to be strengthened and the level to be achieved through subject leap. For example, using subject assessment as the specific evaluation basis, the user specifies the subject to be strengthened as B- and the level to be achieved as A-. The subject leap function analyzes the current situation of the subject to be strengthened to clarify the gap between the current level of the subject and the user-specified goal. Based on the logical analysis capabilities of the large model, this invention will construct a current situation analysis intelligent agent. It analyzes the current status of a discipline based on quantified discipline data under a given indicator system, and provides analysis or suggestions based on the functions used. Key words in the function of disciplinary leapfrogging, such as Figure 5 As shown.

[0075] 4. Recall Module

[0076] The recall module is responsible for matching relevant researchers to a specified discipline and matching relevant disciplines to a specified researcher. Similarly, to address the limitations of large models in processing longer contexts, this invention trains a recall model to return a list of disciplines most relevant to the specified input, narrowing the decision-making scope of the large model and allowing it to focus on processing valuable information.

[0077] In classification tasks, the normalized confidence score implicitly reflects the degree of confusion between different categories. For example, in a cat-dog-tree three-class classification task, the classification model outputs a confidence score of [0.6, 0.39, 0.01] for an image of a cat. This result implies that cats and dogs have a stronger correlation, while cats and dogs have a lower correlation with trees. Based on this idea, this invention proposes... Figure 6 The recall model based on latent information shown includes a BERT encoder, a linear layer, and a softmax layer connected in sequence. The output of the BERT encoder is V1, V2, ..., V... n During the training phase, this invention uses researchers and their respective disciplines as labeled data. By inputting research information such as researchers' projects, papers, and patents, it predicts the discipline to which the researchers belong and trains a BERT-based classification model. The loss function used is as follows:

[0078]

[0079] Where c is the probability distribution after BERT and Softmax processing.

[0080] The recall model trained using formula (8) will return the relationship between different research information and corresponding disciplines. During the testing phase, the confidence scores after the Softmax layer will be sorted by value and used for prediction. Figure 6 The top-k related disciplines are highlighted in the red box. Furthermore, since the researchers' disciplines are already known information, therefore, as... Figure 6 As shown, the recall model based on latent information will discard the confidence level corresponding to the original subject. .

[0081] The recall model based on latent information plays different roles in different functions. Specifically, in the fine-grained triage function of subject refinement (assigning researchers in the proposed revoked subject to new subjects), the recall model will return a list of strongly related subjects for each researcher in the revoked subject. In the disciplinary leap function, this invention first uses a recall model (whose input includes relevant data for a specified discipline) to calculate relevant disciplines for schedulable researchers in the database, and returns all researchers whose relevant disciplines include the discipline to be strengthened, in order to construct a list of researchers with strong relevance to the specified discipline. .

[0082] 5. Explainable decision-making module

[0083] The interpretable decision module is a decision-making agent built from a large model. The large model processes the results returned by the smaller model and provides decision-making criteria based on research direction, research information, etc. Specifically, in the subject refinement—coarse-grained merging function, the large model will use a list of subjects to be withdrawn and their related subjects. As input, the model provides a decision interpretation regarding the final allocation of disciplines to be revoked; in the discipline refinement—fine-grained triage function, the large model will use the information of researchers in the disciplines to be revoked and a list of disciplines strongly related to those researchers. As input, a fine-grained analysis and decision-making process is performed on the final affiliation of researchers under the proposed revocation disciplines; in the discipline leap function, the large model uses a list of strongly related researchers and specified disciplines to be strengthened. Using the target level as input, the decision-making agent makes resource reallocation decisions based on the research personnel. During this decision-making process, the decision-making agent... The analysis will use the research directions of different individuals as the basis, supporting the decision-making results by explaining the meaning of the research directions or the degree of matching between them. Taking disciplinary leap as an example, the decision-making agent... Some key prompts such as Figure 7 As shown.

[0084] 6. Comparative Analysis Module

[0085] Comparative analysis agents based on large models Combined with decision-making intelligent agents Provide a solution and predict the expected results after its implementation. This helps users understand the specific changes in the subject matter and assists them in deciding whether to adopt the solution. CA Key prompts such as Figure 8 As shown.

[0086] 7. Visualization Module

[0087] While multimodal large models can generate visual images, they struggle to handle highly detailed subject-specific data. To improve the user's understanding of the content generated by the aforementioned modules, this invention constructs a visualization agent A based on the tool-calling capabilities of large models. T and by calling such Figure 9 The functions shown complete targeted visualizations. Among them, ChordChat is the function for creating chord charts, WordCloud is the function for creating word clouds, RadarChart is the function for creating radar charts, ColumnChart is the function for creating bar charts, and HeatMap is the function for creating heatmaps.

[0088] 8. A subject structure optimization system driven by collaboration between large and small models

[0089] Combining the seven modules mentioned above, this invention constructs a subject structure optimization system driven by the collaboration of large and small models, and uses multimodal reports as the system output to achieve subject structure optimization oriented towards revocation and resource reallocation. This section will provide a detailed description of the subject refinement and subject leap functions in the constructed system.

[0090] The subject refinement function aims to reduce redundancy in subject settings and can be categorized into coarse-grained merging and fine-grained decentralization based on the subsequent processing methods for subjects to be withdrawn. For example... Figure 10 In the subject refinement function shown, the system first imports compressed subject information from the database, and then a dedicated small model extracts agents based on the subject research direction (research direction extraction agent). ) Calculate the correlation between different disciplines, and then use a large model (association analysis agent) The system will analyze the statistical results of the data based on the small model to help users understand the overall relevance distribution of large-scale disciplines. Subsequently, for disciplines slated for revocation, the system will reallocate resources at two levels: coarse-grained discipline merging and fine-grained researcher allocation. For coarse-grained merging, the system will provide the Top-k strongly related disciplines of the slated revocation discipline based on the discipline relevance calculated by the dedicated small model (mentioned in the association analysis module), and provide specific explanations based on the research directions of the slated revocation discipline and the strongly related disciplines (association analysis agent). After the user specifies the merged discipline, to help the user understand the changes before and after the implementation of the solution, the system will build a new merged discipline profile in the database. Simultaneously, it will perform analysis based on a large model and visualize the changes using a visualization module to help the user understand the potential effects of the relevant solution. For fine-grained allocation, the system will assign new disciplines to researchers in the revoked discipline based on their research directions. Specifically, the system will first export the research information, such as projects and papers, of all researchers under the revoked discipline from the database, and then use a dedicated small model (recall module) to provide a list of relevant disciplines for each researcher based on the research information and the research direction of the discipline. Based on this, the large model (decision-making agent)... The system will make more granular decisions based on the results returned by the aforementioned smaller model and subject information, and assign appropriate new subjects to each researcher. Simultaneously, during the decision-making process, the larger model will provide relevant reasoning by combining research direction, helping users understand the proposed solutions and decide whether to adopt them.

[0091] The disciplinary leapfrog function aims to integrate existing resources by adjusting the disciplines to which researchers belong, thereby enhancing the strength of a designated discipline and achieving rapid, leapfrog growth in that discipline within a short period. For example... Figure 11 As shown, after specifying the subject to be strengthened and the target level, the system will import relevant data for the subject to be strengthened from the database, and then use the large model (current situation analysis agent A) to analyze the data. PAThe system analyzes the performance of the discipline to be strengthened under relevant indicators. Then, the user inputs the desired level (e.g., the performance of the same discipline in other universities under the same evaluation indicators), and the large model analyzes the gap between the current level and the desired level of the discipline to be strengthened. Based on this, the system calls upon a small expert model (recall module) to identify the top-k researchers most relevant to the discipline to be strengthened by matching their research directions. The large model (decision agent) then... Further screening will be conducted to make reasonable personnel adjustment decisions. Simultaneously, the relevant adjustment plans and reasons will be provided based on the basic information of the researchers and their research directions. Finally, the system will list the specific changes before and after the optimization of the discipline structure and, by comparing them with the intended level, demonstrate whether the system has achieved the user-specified goals.

[0092] The aforementioned subject structure optimization process is completed collaboratively by large and small models and presented to users in the form of a multimodal report.

[0093] The following experimental analysis will illustrate the effectiveness of the system.

[0094] 1) Experimental data and setup.

[0095] The method and system for optimizing subject structure driven by the collaboration of large and small models proposed in this invention have been validated on research data from 49 disciplines at XX University, where the data from each discipline accounts for a significant proportion. Figure 12 As shown, clinical medicine accounts for the highest proportion of research data at this university, reaching 32.3%. In contrast, research data from disciplines such as psychology accounts for 0.1% or less. This uneven distribution of disciplinary data will significantly impair the model training effect. Therefore, this invention will utilize the powerful generative capabilities of a large model to rewrite the tail-end disciplinary data and increase the proportion of corresponding disciplinary data through resampling to improve the performance of the constructed disciplinary structure optimization system. It should be noted that different universities will have significant differences in the distribution of research data due to their different strong disciplines, and the specific research directions of the same discipline may also differ due to policy and other reasons.

[0096] The system built in this invention uses DeepSeek-V3 as the general-purpose large model and BERT as the dedicated small model. The training of the relevant models is implemented on two A6000 GPUs, while the inference of the general-purpose large model is implemented using relevant APIs.

[0097] First, the effectiveness of the proposed optimization method in extracting disciplinary research directions is verified and compared with the unsupervised topic model LDA. Since real scientific research data lacks labels, this invention uses a powerful large language model to evaluate the reasonableness of the extracted research directions and assigns a score.

[0098] The relevant results are summarized in Table 2. It can be seen that, among the 49 disciplines, the research directions provided by the optimization method for 34 disciplines are more consistent with the expectations of the large model, and the total score obtained by the optimization method is nearly 10 points higher than that of the topic model. This result indicates that the method proposed in this invention is more reasonable in extracting discipline research directions, and that the subsequent discipline refinement and discipline leap functions based on the research directions obtained by the optimization method are more practically significant.

[0099] Table 2. Results of the topic model and the optimization method proposed in this invention in the extraction of subject research directions.

[0100]

[0101] To quantitatively analyze the subject optimization schemes generated by the system, this invention conducted expert evaluations on the subject refinement—fine-grained triage function, and the results are summarized in Table 3. Specifically, the system first generated 63 decision reports, which were then reviewed by experts who provided feedback (positive / negative). Of these, 60 were positive cases, accounting for approximately 95% of the total. This result indicates that most of the cases provided by the proposed subject structure optimization system are consistent with the expert decision schemes, reflecting that the system can effectively achieve the goal of optimization decision-making. Furthermore, by providing subject structure optimization schemes to relevant decision-makers, the consumption of human resources will be greatly reduced. In addition, Table 3 further presents positive cases (top) and negative cases (bottom). The positive cases show that by reducing the decision scope of the large model through a smaller model and decreasing the length of the input context, the large model can better focus on key information and provide reasonable decisions and justifications. The negative cases indicate that the proposed system still experiences some confusion when making decisions based on complex research directions with certain similarities, making the explanation of the large model's decisions somewhat forced. Therefore, even if the proposed system can largely achieve the goal of optimizing the disciplinary structure, the relevant decision-making process still requires human participation in the form of review to ensure the accuracy of the conclusions.

[0102] Table 3. Subject Refinement – ​​Quantitative Results and Related Cases of Fine-Grained Diversion Functionality

[0103]

[0104] This implementation method uses "Philosophy" and "Marxism" at XX University as examples to conduct a case study analysis of the function of discipline refinement—coarse-grained discipline integration. Please note that the disciplines involved in this case are only for verifying the effectiveness of the proposed method and do not represent any level of decision-making or guidance. Furthermore, the discipline data in this case may have inaccurate quantification of discipline indicators due to issues such as collection methods, but this does not affect the demonstration of the optimization process.

[0105] like Figure 13As shown in (a), when a user inputs "Philosophy" as the subject to be revoked at University XX, the system will perform a current status analysis based on a large model and call appropriate plotting functions to visualize the data. For example... Figure 13 As shown in (b), to help users select the subjects to be merged into for a subject to be revoked, the system performs correlation analysis from both global and local perspectives. The global correlation analysis displays and describes the correlation distribution of all subjects across the university, while the local correlation analysis provides the top-5 most relevant subjects for the user-input subject to be revoked, along with explanations from a research direction perspective. In this case, the subjects most relevant to "Philosophy" at XX University are "Marxist Theory," "Education," "Political Science," "Art," and "Chinese History." It should be noted that the correlation calculation results for the same subject at different universities will vary significantly due to differences in the research directions of researchers. For example... Figure 13 As shown in (c), the user selects potential disciplines to be merged from the system's list, and the large model analyzes and supports the relevant decisions, providing a basis for the relevant merger work. Figure 13 As shown in (d), the system merges the data of "Philosophy" from XX University and "Marxist Theory" from XX University, predicts the changes under different indicators after the merger, and further provides relevant analysis to help users understand the effects of the implementation of the plan.

[0106] exist Figure 13 Based on this, this implementation method is Figure 14 The document showcases the intermediate outputs of each module in the subject refinement—coarse-grained merging case study, to further demonstrate the specific functions of the built system.

[0107] This invention establishes a subject structure optimization method based on a collaborative approach driven by size models, such as... Figure 15 The system for optimizing the subject structure is shown, and the video attached demonstrates how to use the system.

[0108] Furthermore, this implementation method evaluated the time and cost overhead of the system during actual operation, and summarized the results through 10 operational cases. Figure 16 and Figure 17 Here, LLM indicates that the entire process of optimizing the subject structure is completed using only a large model. Specifically, regarding time overhead, compared to LLM, the proposed large-scale model collaborative framework improves system running speed by 25.61%–36.77%, significantly reducing the waiting time for users to interact with the system each time. It should be noted that... Figure 16 The mid-case study refers to the overall generation time of a multimodal report that includes user interaction. Regarding cost, since closed-source large models are billed based on tokens, in... Figure 17The number of tokens in the input and output of the closed-source large model was summarized. Compared with LLM, the system proposed in this invention reduces the cost of 10 run cases by 47.11% to 50.17%. This result shows that, compared with the time cost, the system reduces the running cost more effectively by using a large and small model collaboration approach.

[0109] In summary, to address the technical problems existing in the aforementioned traditional discipline governance systems, this invention proposes a system based on a collaborative deployment approach using both large and small models. This system leverages the respective advantages of both models to form a complementary and collaborative mechanism. The specific modules and their characteristics are as follows:

[0110] 1) Large Model Module. This module is used to handle complex interdisciplinary tasks, such as semantic understanding and generation of long texts, extraction and compression of research directions, and interdisciplinary correlation analysis and logical reasoning. Leveraging its extensive prior knowledge and logical decision-making capabilities, the large model can solve problems that are difficult for smaller models to cover, such as the explanation of specialized terms, the correlation of multi-dimensional indicators, and the analysis of complex logical links, thereby ensuring the overall performance of the system in terms of knowledge coverage and reasoning accuracy.

[0111] 2) Small Model Module. Designed for high-frequency, localized, and structured sub-tasks, such as fine-grained matching of researchers with subject areas, recall and screening of research data, local similarity calculation, and redundancy detection. Small models offer better fitting results for specific tasks and data distributions, effectively alleviating the inefficiency of large models in handling long contexts, and achieving performance close to or even surpassing that of large models on some tasks, thereby improving overall processing speed and reducing the system's computational and hardware resource consumption.

[0112] 3) Scheduling and Optimization. As the central hub for collaboration between large and small models, this module uses task allocation and scheduling algorithms to assign global and complex tasks to the large model, while delegating local and repetitive tasks to the small model. This differentiated scheduling not only ensures that the large model can concentrate its computing power on core and challenging problems but also avoids wasting computing power on high-frequency subtasks. Simultaneously, this module can dynamically optimize the interaction between the large and small models. For example, the small model can first retrieve candidate results, and then the large model can perform in-depth analysis and interpretation, ultimately achieving a collaborative optimization effect that balances efficiency and accuracy.

[0113] 4) Explainability and Visualization Support. In practical applications of discipline structure optimization, decision-making results need to be understood and accepted by researchers and managers. To this end, this invention introduces large-model-driven explainable decision-making and small-model-assisted visualization analysis to automatically generate explanatory text and chart reports based on research data, research directions, and discipline indicators, thereby improving the transparency and acceptability of system results.

Claims

1. A subject structure optimization system based on the collaboration of large and small models, characterized in that, This includes the system interface, database, large model, and small model; The system interface is used to allow users to select functions and display the output results of large and small models accordingly; The database is used to store subject-related data and research information of researchers; The large model is used to generate the data required for the small model calculation based on the user-selected functions and the data in the database, and to perform in-depth analysis, logical reasoning and interpretation based on the user-selected functions and the results output by the small model; Small models are used to analyze the correlation between disciplines based on comprehensive descriptions of different disciplines to determine a list of strongly correlated disciplines for a given discipline, or to determine a list of strongly correlated disciplines for a given researcher or a list of strongly correlated researchers for a given discipline based on the researcher's research information.

2. The subject structure optimization system based on the collaboration of large and small models according to claim 1, characterized in that, When the user selects the function of subject refinement to reduce the redundancy of subject settings, the data required for the large model to generate the small model includes a comprehensive subject description. The process of generating the comprehensive subject description includes: using a research direction extraction agent built based on the prior knowledge of the large model to read and process the subject-related data stored in the database to generate a list of research directions for the subject; and then using an association analysis agent built based on the logical analysis capability of the large model to semantically combine the list of research directions for the subject to generate a comprehensive subject description.

3. The subject structure optimization system based on the collaboration of large and small models according to claim 1, characterized in that, The process of analyzing the correlation between disciplines using a small model includes: projecting the comprehensive description of each discipline into the latent space using a pre-trained embedding model, and calculating the similarity between disciplines based on the vectors of the comprehensive descriptions of each discipline in the latent space.

4. The subject structure optimization system based on the collaboration of large and small models according to claim 1, characterized in that, The model used in the small model to determine the list of strongly related disciplines for a given researcher or the list of researchers strongly related to a given discipline is a BERT-based classification model. The BERT-based classification model includes a BERT encoder, an embedding space processing unit, a linear layer, and a Softmax layer connected in sequence.

5. The subject structure optimization system based on the collaboration of large and small models according to claim 1, characterized in that, When the user selects the coarse-grained merging function under subject refinement, the small model is used to analyze the correlation between the subject to be revoked and other subjects based on the comprehensive subject description of the subject to be revoked and the comprehensive subject description of other sciences input by the user to determine the list of strongly related subjects of the subject to be revoked. When the user selects the fine-grained sub-distribution under subject refinement, the small model is used to determine the list of strongly related subjects for a specified researcher based on the researcher's research information. Its input is the researcher's research information, and its output is the predicted list of strongly related subjects. When the user selects the discipline leap function, the small model is used to determine the list of strongly related researchers for the specified discipline to be strengthened based on the researchers' research information. Its input is the relevant data of the specified discipline, and its output is the predicted list of strongly related researchers. Discipline refinement is used to reduce redundancy in discipline settings, while discipline leapfrogging is used to adjust the disciplines to which researchers belong; coarse-grained merging is used to incorporate abolished disciplines into new disciplines, while fine-grained diversion is used to assign researchers from disciplines to be abolished to new disciplines.

6. The subject structure optimization system based on the collaboration of large and small models according to claim 5, characterized in that, When the user selects the coarse-grained merging function under subject refinement, the large model provides the following decision-making and analysis process: using a decision-making agent built based on the large model to make a decision interpretation on the final affiliation of the proposed revoked subject according to the list of subjects to be revoked and their strongly related subjects; When the user selects the fine-grained diversion under the subject refinement function, the big model provides the following decision and analysis process: using the decision-making agent built based on the big model to make a fine-grained analysis and decision on the final affiliation of the researchers under the proposed revoked subject based on the information of the researchers of the proposed revoked subject and the list of strongly related subjects of the researchers of the proposed revoked subject. When the user selects the disciplinary leap function, the big model provides the following decision-making and analysis process: using a decision-making agent built on the big model to make resource reallocation decisions based on the specified discipline to be strengthened, its strongly related list of researchers, and the level to be achieved.

7. The subject structure optimization system based on the collaboration of large and small models according to claim 2, characterized in that, The subject-related data input to the research direction extraction agent is the subject-related data after preliminary processing, which includes noise removal.

8. The subject structure optimization system based on the collaboration of large and small models according to claim 3, characterized in that, The similarity is cosine similarity.

9. The subject structure optimization system based on the collaboration of large and small models according to claim 6, characterized in that, The large model is also used to predict the effects of various subject structure optimization schemes proposed by the decision-making agent, in order to help users decide whether to adopt the relevant schemes.

10. The subject structure optimization system based on the collaboration of large and small models according to any one of claims 1 to 9, characterized in that, The system interface is used to display results by calling relevant functions using a visual intelligent agent. The visual intelligent agent is built based on the tool calling capabilities of a large model. The relevant functions include functions for creating chord diagrams, word clouds, radar charts, bar charts, and heatmaps.

Images