Personalized learning guiding system based on generative agent model

Abstract

The invention relates to the technical field of intelligent education, in particular to a personalized learning guiding system based on a generative agent model, and the system comprises a generative agent role module which is used for simulating learner roles under different subject backgrounds according to a large language model; the interdisciplinary dialogue simulation module is used for simulating interactive communication between the learner roles under a preset task or theme condition, and obtaining simulated dialogue content generated by the learner roles; the evaluation and feedback module is used for carrying out quantitative analysis on the simulated dialogue content and outputting an evaluation index to optimize a personalized learning path; and the memory mechanism module is used for storing and managing historical simulated dialogue content, acquiring key information and inputting the key information into the large language model to optimize the simulated dialogue content generated by the learner role. According to the invention, low efficiency of communication between subjects and obstacles of knowledge transmission in a traditional education mode can be solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to a personalized learning guidance system based on a generative agent model. Background Technology

[0002] With the development of intelligent education technology, personalized learning and interdisciplinary knowledge integration have become important means to improve educational effectiveness. The application of Generative Agent Models (GAMs) and Large Language Models (LLMs) in intelligent education systems has gradually gained widespread attention, especially in the fields of personalized guidance and knowledge transfer. However, existing personalized learning platforms have certain limitations in simulating interdisciplinary communication and knowledge transfer, failing to fully meet the interactive needs of students from different disciplines. Particularly in learning scenarios requiring multidisciplinary knowledge integration, existing educational platforms are often limited to a single subject area, failing to effectively address the differences in language styles and barriers to knowledge transfer between disciplines. Traditional teaching methods typically rely on static curriculum design and teaching resources, unable to dynamically adjust according to students' backgrounds and needs in real time. Therefore, a new method is needed to simulate communication and interaction between students from different disciplinary backgrounds and optimize their personalized learning paths. Summary of the Invention

[0003] The purpose of this invention is to provide a personalized learning guidance system based on a generative agent model. This system optimizes learning paths and improves interdisciplinary communication and knowledge sharing efficiency by accurately simulating the dialogue styles of learners in different subject backgrounds. It aims to solve the problems of inefficient interdisciplinary communication and knowledge transfer barriers in traditional education models.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A personalized learning system based on generative agent models includes: a generative agent role-playing module, an interdisciplinary dialogue simulation module, an evaluation and feedback module, and a memory mechanism module.

[0006] The generative agent role-playing module is used to simulate learner roles in different disciplinary backgrounds based on a large language model;

[0007] The interdisciplinary dialogue simulation module is used to simulate the interaction and communication between the learner roles under preset task or theme conditions, and to obtain the simulated dialogue content generated by the learner roles.

[0008] The evaluation and feedback module is used to perform quantitative analysis on the simulated dialogue content and output evaluation indicators to optimize the personalized learning path.

[0009] The memory mechanism module is used to store and manage historical simulated dialogue content, obtain key information, and input the key information into the large language model to optimize the simulated dialogue content generated by the learner role.

[0010] Optionally, simulating learner roles in different subject backgrounds based on the large language model includes: obtaining preset prompt words, inputting the preset prompt words into the large language model, and simulating learner roles in different subject backgrounds.

[0011] Optionally, obtaining the simulated dialogue content generated by the learner role includes:

[0012] The learner role setting input information, dialogue context input information, and knowledge retrieval input information are input into the large language model to obtain the simulated dialogue content generated by the learner role. The learner role setting input information includes subject background, language style, expression characteristics, and personalized parameters. The dialogue context input information includes the speech content of the current round, historical interaction records, and context state. The knowledge retrieval input information is obtained from the knowledge base.

[0013] Optionally, obtaining the knowledge retrieval input information from the knowledge base includes: retrieving matching information from the knowledge base using RAG technology to obtain the knowledge retrieval input information.

[0014] Optionally, the interdisciplinary dialogue simulation module also controls the speaking order, interaction rounds, and behavioral boundaries of learner roles through AgentScope.

[0015] Optionally, the simulated dialogue content may be quantitatively analyzed, including by using information redundancy, compression rate, sentence similarity, lexical abstraction, TF-IDF value, and information entropy to perform quantitative analysis on the simulated dialogue content.

[0016] Optionally, optimizing personalized learning paths includes:

[0017] The learner's role is evaluated based on the evaluation indicators to construct a personalized competency profile;

[0018] The interaction relationships between the learner roles are analyzed based on the role interaction data and the personalized ability profiles to identify the optimal dialogue pairings.

[0019] Based on the optimal dialogue pairing combination, the role scheduling strategy in the dialogue rounds is dynamically optimized to achieve optimization of the personalized learning path.

[0020] Optionally, the historical simulated dialogue content is stored and managed, and key information is obtained, including:

[0021] A timestamp is set for the historical simulated dialogue content, wherein the timestamp marks the time when the historical simulated dialogue content was generated and the latest retrieval time;

[0022] A memory decay mechanism is used to calculate the memory value of the historical simulated dialogue content to obtain the updated memory value;

[0023] The key information is obtained based on the timestamp and the updated memory value.

[0024] Optionally, obtaining the key information based on the timestamp and the updated memory value includes:

[0025] When the updated memory value is lower than a preset threshold, the historical simulated dialogue content is marked as low priority and deleted.

[0026] When the simulated dialogue content of the history is recalled, the timestamp and initial memory strength are updated, and the simulated dialogue content of the history is retained for an extended period.

[0027] The beneficial effects of this invention are as follows: The system of this invention can significantly improve the effectiveness of interdisciplinary communication, reduce information redundancy, enhance language adaptability, and effectively promote knowledge transfer and integration. This system provides an innovative solution for interdisciplinary collaboration and communication in intelligent education and points to a new direction for the development of future education systems. By continuously optimizing interdisciplinary dialogue strategies and learning paths, this invention provides solid theoretical support and technical guarantees for promoting the development of intelligent education systems, especially in interdisciplinary collaboration and multidisciplinary integration. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a framework diagram of a personalized learning system based on a generative intelligent agent model according to an embodiment of the present invention;

[0030] Figure 2This is a comparison chart of communication indicators under different disciplinary perspectives and memory modes in embodiments of the present invention. Among them, (a) is a comparison chart of communication indicators under the humanities-humanities mode, (b) is a comparison chart of indicators under the science and engineering mode, (c) is a comparison chart of communication indicators under the humanities-science and engineering mode with the humanities as the analytical perspective, (d) is a comparison chart of communication indicators under the humanities-science and engineering mode with the science and engineering as the analytical perspective, (e) is a comparison chart of communication indicators under the humanities-science and engineering mode with memory mechanism with the humanities as the analytical perspective, and (f) is a comparison chart of communication indicators under the humanities-science and engineering mode with memory mechanism with the science and engineering as the analytical perspective. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Example 1:

[0034] This embodiment provides a personalized learning system based on a generative agent model, including: a generative agent role-playing module, an interdisciplinary dialogue simulation module, an evaluation and feedback module, and a memory mechanism module;

[0035] The generative agent role-playing module is used to simulate learner roles in different disciplinary contexts based on a large language model;

[0036] The interdisciplinary dialogue simulation module is used to simulate the interaction and communication between learner roles under preset task or theme conditions, and to obtain the simulated dialogue content generated by the learner roles.

[0037] The evaluation and feedback module is used to quantitatively analyze the simulated dialogue content and output evaluation indicators to optimize personalized learning paths.

[0038] The memory mechanism module is used to store and manage historical simulated dialogue content, retrieve key information, and input the key information into the large language model to optimize the simulated dialogue content generated by the learner role.

[0039] Specifically, in existing systems, significant differences exist between disciplines in terms of language style, thought processes, and knowledge organization, often limiting effective interdisciplinary communication and knowledge transfer. This leads to the problems of disciplinary fragmentation and low collaborative efficiency in traditional education models. To address this, this embodiment constructs an interdisciplinary dialogue intelligent agent system based on the Qwen-Max Large Language Model (LLM). By simulating multi-round interactions between humanities and science learners, it explores new paths to improve the efficiency of interdisciplinary communication and cognitive transformation. This system can accurately reproduce the language expression characteristics and thinking habits of different disciplinary backgrounds, automatically generate personalized dialogue content, and achieve intelligent scheduling of multiple agents through dynamic path planning, thereby improving the quality of interdisciplinary communication and the efficiency of knowledge sharing.

[0040] Furthermore, to achieve a highly personalized and controllable learning process, the system relies on the assessment and feedback module to quantitatively analyze the dialogue content generated by the interdisciplinary dialogue simulation module in each round, covering six core language behavior indicators: compression rate, information redundancy, sentence similarity, lexical abstraction, TF-IDF value, and information entropy. Based on the above data, the system can comprehensively evaluate the specific performance of each generative agent, i.e., learner role, in terms of language expression refinement, knowledge integration ability, and semantic diversity, and thereby construct a personalized ability profile including subject comprehension preference, expression complexity, knowledge retrieval depth, and language style characteristics.

[0041] Building upon this foundation, the system further analyzes the interaction relationships between agents by combining the role interaction data and personalized ability profiles generated during the simulation process. This allows for the identification of optimal dialogue pairings. When the system faces new learning tasks or communication goals, it can prioritize the use of these "optimal pairings" in the task simulation, significantly improving the quality of dialogue expression and cognitive output. This strategy not only effectively avoids the repeated consumption of resources among inefficient agent combinations but also improves the cost-effectiveness of agent calls and task completion efficiency, achieving optimal scheduling of generated resources and maximizing cognitive benefits. Specifically, the system dynamically compares the collaborative performance of agents with different disciplinary backgrounds in multiple rounds of simulated interactions, comprehensively evaluating their matching degree in dimensions such as information gain, semantic connectivity, and complementary expression styles. From this, it selects "high-efficiency learning partners" with high collaborative potential under specific learning tasks. This combination information will serve as key input feedback in subsequent learning path planning. When the system faces new learning tasks or communication goals, it will prioritize the use of the previously identified optimal agent combinations in simulated dialogue. This strategy can not only significantly improve the expressive quality and cognitive output of dialogue generation, but also effectively avoid the repeated consumption of computing resources in inefficient combinations, thereby improving the overall efficiency of agent invocation and maximizing the optimized scheduling of generation resources and cognitive benefits.

[0042] To further enhance the adaptability and accuracy of task execution, the system simultaneously introduces a role scheduling strategy. This strategy dynamically selects the most suitable combination of agents based on the current task type, agent capability profiles, and historical interaction performance, guiding them to participate in the dialogue. This strategy achieves dual optimization of semantic complementarity and task adaptation between roles, ensuring that the generated content achieves superior levels in language expression, knowledge integration, and thought transfer. Specifically, the role scheduling strategy refers to the system dynamically selecting agent combinations that highly match the requirements of the current task to participate in simulated dialogues when facing different types of learning tasks or communication goals, based on the learner's agent capability profiles and existing interaction history data. This strategy aims to maximize information synergy and optimize semantic complementarity between roles, thereby comprehensively improving the quality of generated dialogue content, the depth of knowledge integration, and the efficiency of interdisciplinary transfer.

[0043] Furthermore, simulating learner roles in different disciplinary contexts based on the large language model includes: obtaining preset prompt words, inputting the preset prompt words into the large language model, and simulating learner roles in different disciplinary contexts.

[0044] Furthermore, the simulated dialogue content generated by the learner role includes:

[0045] Input the learner role setting information, dialogue context information, and knowledge retrieval information into the large language model to obtain the simulated dialogue content generated by the learner role. The learner role setting information includes subject background, language style, expression characteristics, and personalized parameters. The dialogue context information includes the speech content of the current round, historical interaction records, and context state. The knowledge retrieval information is obtained from the knowledge base.

[0046] Specifically, the interdisciplinary dialogue simulation module simulates communication through four different dialogue modes combining different disciplines (such as learners from humanities and STEM backgrounds). By simulating the differences in thinking and language styles among learners from diverse disciplinary backgrounds, the system studies the challenges and solutions in interdisciplinary communication, particularly how to promote knowledge sharing and improve language adaptability during dialogue. Through simulating interactions between different disciplines, the system helps students overcome interdisciplinary language barriers during learning, thereby enhancing their understanding and mastery of complex subject content in actual learning.

[0047] Furthermore, obtaining knowledge retrieval input information from the knowledge base includes: retrieving matching information from the knowledge base using RAG technology to obtain knowledge retrieval input information.

[0048] Furthermore, the interdisciplinary dialogue simulation module also uses AgentScope to control the speaking order, interaction rounds, and behavioral boundaries of learner roles.

[0049] Furthermore, quantitative analysis is conducted on the simulated dialogue content, including: quantitative analysis of the simulated dialogue content through information redundancy, compression rate, sentence similarity, lexical abstraction, TF-IDF value, and information entropy.

[0050] Specifically, the assessment and feedback module uses six core indicators to quantitatively analyze and evaluate the effectiveness of interdisciplinary communication: N-gram information redundancy, compression rate, sentence similarity, lexical abstraction, TF-IDF value, and information entropy. These indicators explore the quality of dialogues between learners from different disciplinary backgrounds. The results of this analysis provide a reference for subsequent personalized learning paths, helping to develop more precise personalized strategies. By providing regular feedback on learners' performance, the system can adjust the dialogue content in real time to improve learners' language communication skills and interdisciplinary knowledge integration abilities. In this way, the system can not only accurately assess the effectiveness of interdisciplinary communication but also promptly identify problems encountered by students during the learning process, thereby providing personalized improvement suggestions and ensuring continuous optimization of the learning path.

[0051] Furthermore, optimizing personalized learning paths includes:

[0052] The learner's role is evaluated based on the evaluation indicators to construct a personalized competency profile;

[0053] The interaction relationships between the learner roles are analyzed based on the role interaction data and the personalized ability profiles to identify the optimal dialogue pairings.

[0054] Based on the optimal dialogue pairing combination, the role scheduling strategy in the dialogue rounds is dynamically optimized to achieve optimization of the personalized learning path.

[0055] Furthermore, the historical simulated dialogue content is stored and managed to obtain key information, including:

[0056] Set timestamps for historical simulated dialogue content, where the timestamps mark the time when the historical simulated dialogue content was generated and the latest retrieval time;

[0057] A memory decay mechanism is used to calculate the memory value of historical simulated dialogue content and obtain the updated memory value.

[0058] Retrieve key information based on timestamps and updated memory values.

[0059] Furthermore, based on the timestamp and the updated memory value, key information can be obtained, including:

[0060] When the updated memory value is lower than a preset threshold, the historical simulated dialogue content is marked as low priority and deleted.

[0061] When historical simulated dialogue content is recalled, the timestamp and initial memory strength are updated, and the historical simulated dialogue content is retained for an extended period.

[0062] Specifically, the memory mechanism module records and retains key information during agent dialogue to help the agent maintain contextual consistency and improve dialogue coherence and knowledge integration capabilities. By storing historical dialogue content in the short term, the memory mechanism module allows the agent to refer to previous dialogue content in subsequent rounds of communication, thereby optimizing information transmission logic and the fluency of language expression. Experiments show that introducing the memory mechanism can effectively reduce information redundancy in dialogue, improve the efficiency of interdisciplinary communication, and enhance knowledge transfer and integration capabilities. The memory mechanism module supports information accumulation in multi-round dialogues, enabling the system to continuously optimize language strategies and dynamically adjust knowledge expression methods, thus facilitating learners to more smoothly achieve knowledge transfer and cognitive expansion in multidisciplinary environments.

[0063] Example 2:

[0064] like Figure 1 The image shows a personalized learning system based on a generative agent model. The system includes a generative agent role-playing module, an interdisciplinary dialogue simulation module, an evaluation and feedback module, and a memory mechanism module. It can build an intelligent interdisciplinary dialogue environment and support the dynamic optimization of personalized learning paths.

[0065] The generative agent role-playing module is based on the Qwen-Max large language model and constructs a generative agent role-playing modeling framework for interdisciplinary communication scenarios. It aims to endow agents with clear subject identities and language characteristics, enabling them to have controllable expression styles, knowledge structures and cognitive paths, and realistically reproduce the interactive behaviors of learners from different disciplines.

[0066] This module employs a role-based engineering approach, using carefully designed Prompt instructions to transform a general language model into generative intelligent agents with specific role personalities. Each agent receives a unique Prompt during the initialization phase, clarifying its discipline, professional direction, language preferences, expression style, and thinking habits, thus exhibiting disciplinary consistency and behavioral stability during language generation.

[0067] When the model is connected, a role-setting prompt is first input to ensure that Qwen-Max can accurately simulate the role's language characteristics. In subsequent text generation, the agent always organizes language and outputs knowledge based on the set role. For example, humanities roles tend to use conceptual and argumentative language, emphasizing rhetoric and semantic depth; while science and engineering roles focus on structured expression and data support, with a more precise and logical language style.

[0068] The system constructs a pool of 20 virtual learners, divided into the H series (H1-H10) and the S series (S1-S10), representing learners with backgrounds in the humanities and STEM fields, respectively. Each agent possesses unique role settings and language behavior characteristics. For example, H1 focuses on cognitive psychology, with language leaning towards theoretical analysis and case citations; H2 comes from classical literature and excels in philosophical discourse; S1 has a computer science background, expresses itself clearly and concisely, and can analyze topics based on algorithms and information structures.

[0069] The interdisciplinary dialogue simulation module aims to construct a multi-role-driven, task-oriented language interaction environment for high-fidelity simulation of communication behavior between learners from different disciplinary backgrounds. Building upon the previously constructed generative agent role modeling foundation, this module drives various agents to conduct multiple rounds of interaction around common topics under preset thematic conditions, thereby reproducing the differences in language styles, knowledge structures, and thought processes between disciplines. By simulating the dynamic evolution of language expression, this module not only enhances the system's adaptability to interdisciplinary communication but also provides structured dialogue data support for subsequent communication quality analysis and learning strategy optimization.

[0070] In terms of technical implementation, this module is based on the Qwen-Max large language model and integrates key components such as LangChain, AgentScope, and RAG. The system uses LangChain to dynamically invoke Prompts and manage role instructions, ensuring the model maintains consistency in subject-specific expression characteristics throughout the generation process. AgentScope is responsible for scheduling and managing multiple agents, controlling their speaking order, interaction rounds, and behavioral boundaries to maintain the logical structure and multi-turn consistency of the dialogue. Behavioral boundaries refer to the system's restrictions on the acceptable information sources for each agent, clarifying which other roles they can interact with in the dialogue to ensure targeted communication. RAG technology provides knowledge enhancement support for the generation process. The system dynamically retrieves information from external knowledge bases based on the dialogue content to supplement the agent's knowledge reserves, ensuring stronger semantic coverage and response depth in cross-disciplinary scenarios. The collaborative operation of these modules constitutes a configurable, highly scalable, and realistic cross-disciplinary language generation system.

[0071] During operation, the system receives three core inputs: first, role setting input, which clarifies the subject identity, language style, and behavioral characteristics of each agent; second, dialogue context input, including the content of the current turn's speech, historical interaction records, and contextual state; and third, knowledge retrieval input, where the system extracts matching information from an external knowledge base based on the current dialogue topic, supporting accurate responses to complex questions and the development of viewpoints. The model combines these three inputs to generate dialogue text, producing outputs with high role consistency, diverse language styles, and logical coherence. Simultaneously, it utilizes a memory mechanism to maintain multi-turn contextual consistency, strengthening the stability of interdisciplinary language expression and the ability to cross-integrate subject information.

[0072] In practical applications, the system can conduct experimental dialogue simulations for specific task themes and agent role combinations. For example, under the theme of "Student Learning Motivation and Social Group Influence," the system sets up H1 (with a psychology background) and S1 (with a computer science background) to conduct 20 rounds of multi-round interactive dialogue. Role H1 focuses on the construction of psychological theories and empirical research, and its expression style tends to develop viewpoints based on theoretical models and experimental data, often citing classic behavioral models such as self-determination theory and social identity mechanisms; while S1 takes system modeling and data-driven analysis as its main perspective, and is good at using natural language processing technology to analyze the evolution of language structure, and explores the manifestation of learning motivation in dialogue corpora and its modeling methods by calculating indicators such as language redundancy and information entropy.

[0073] The evaluation and feedback module is used to quantitatively analyze and assess the quality of interdisciplinary dialogue content, providing fundamental data support for system optimization and dynamic adjustment of personalized learning paths. This module is built upon a multi-dimensional language and information evaluation model, combining six core evaluation indicators and using mathematical formulas to calculate dialogue quality, comprehensively measuring the agent's language style, information efficiency, and knowledge integration capabilities during the dialogue process.

[0074] The specific evaluation indicators and their calculation methods are as follows:

[0075] Information redundancy measures the proportion of repetitive information in a dialogue, reflecting the conciseness of the content and the efficiency of information transmission. Frequent occurrences of the same or similar phrases or sentence structures indicate high information redundancy, which may reduce the effectiveness and quality of the dialogue. This metric is calculated by statistically analyzing N-grams, where an N-gram is a phrase consisting of several consecutive words in the text. For example, a 2-gram consists of two adjacent words, and a 3-gram consists of three adjacent words. In this system, 2-grams are used to measure repetition in the dialogue. The formula for calculating information redundancy R is as follows:

[0076]

[0077] Where, N rep Let N be the number of repeated N-grams in the dialogue. total R represents the total number of N-grams in the dialogue. A higher R value indicates that there is more repetition in the dialogue, which may lead to information redundancy and thus affect the efficiency and quality of communication; while a lower R value indicates that the dialogue content is more concise and the information is transmitted more efficiently.

[0078] Compression ratio measures the conciseness of dialogue content and the efficiency of language expression. A lower compression ratio means that the text is more concise while conveying the same amount of information, which helps improve the effectiveness of the dialogue and the accuracy of information delivery. This metric is calculated by compressing the original dialogue content and comparing the text length before and after compression. The formula is as follows:

[0079]

[0080] Among them, L comp This indicates the length of the compressed text, such as through automatic summarization or redundancy removal; L raw This indicates the length of the original, uncompressed dialogue text. Both can be counted by the number of characters or words. The lower the compression ratio C, the more concise the text is, reducing redundancy and repetition without losing key information, thus improving the logical density and reading efficiency of the dialogue.

[0081] Sentence similarity (S) measures the semantic overlap between adjacent sentences in a dialogue and is an important indicator for assessing the fluency and accuracy of information delivery. Higher sentence similarity may indicate information repetition or a lack of variation in language structure, while lower sentence similarity suggests more diverse expressions, contributing to greater depth and breadth of thought. This indicator is calculated by measuring the cosine similarity between adjacent sentences, using the following formula:

[0082]

[0083] in, and Let represent the word vectors of the i-th and (i+1)-th sentences in the dialogue, respectively, where n is the total number of sentences in the dialogue. Cosine similarity measures the angle between two sentences in the vector space, with a value ranging from -1 to 1, where 1 indicates complete similarity and 0 indicates complete dissimilarity.

[0084] Lower sentence similarity indicates more novel expressions and perspective shifts in the dialogue, making it suitable for assessing the openness and innovation of interdisciplinary conversations; while higher similarity may suggest repetitive expressions and slow information progression. This indicator helps determine the coherence and diversity of expressions in a dialogue, further optimizing generation strategies and ensuring that interdisciplinary communication is both accurate and intellectually dynamic.

[0085] Lexical abstraction level (A) measures the degree of abstract vocabulary used in a dialogue, reflecting the conceptual and theoretical depth of the conversation. A higher lexical abstraction level indicates the use of more abstract and specialized vocabulary, enhancing interdisciplinary collaboration and knowledge sharing. The formula for calculating this indicator is as follows:

[0086]

[0087] Where, N abs This represents the number of highly abstract and conceptual words in the dialogue, where M is the total number of words appearing in the dialogue. Abstract words are typically those that are theoretical, conceptual, and have a high academic level, such as "cognitive development" and "cultural heritage." A higher word abstraction value means that the dialogue involves more professional terms and academic concepts, which helps to promote deep integration between disciplines and knowledge transfer.

[0088] The lexical abstraction level index helps assess the disciplinary depth reflected in a dialogue, especially in interdisciplinary conversations, where it effectively utilizes abstract vocabulary for high-level academic discussion and knowledge sharing. This index is crucial for interdisciplinary communication, promoting the understanding and integration of knowledge from different fields.

[0089] Information entropy measures the complexity and uncertainty of dialogue content and is an important indicator for assessing dialogue diversity and information density. A higher information entropy value indicates that the dialogue contains more diverse information, which helps to promote the exchange and collision of diverse perspectives and enhances the depth and breadth of the dialogue. The formula for calculating information entropy H is as follows:

[0090]

[0091] Among them, P(w c Let represent the probability of word c appearing in the dialogue, k be the total number of words in the dialogue, and H represent the total information entropy of the dialogue. Information entropy, by calculating the probability distribution of each word in the dialogue, reflects the diversity and uncertainty of the dialogue content. A higher entropy value indicates a more uniform distribution of words in the dialogue, containing more diverse information and ideas, which helps to promote the exchange and collision of ideas among diverse perspectives, thereby increasing the complexity and information content of the dialogue.

[0092] Information entropy is a useful indicator for assessing the diversity and richness of knowledge in a conversation, and is particularly relevant to interdisciplinary communication. A higher information entropy value means that agents from different disciplinary backgrounds can contribute more diverse perspectives and theories, making the conversation deeper and broader, and promoting the integration and sharing of interdisciplinary knowledge.

[0093] TF-IDF scores are used to assess the richness of vocabulary and the weighting of key information in a conversation. A higher TF-IDF score indicates a higher frequency of academic terminology used in the conversation, and that these terms better reflect the subject matter and professional content, thus ensuring the accuracy and academic rigor of the information conveyed. The calculation formula is as follows:

[0094] TF-IDF(C,d)=TF(c,d)×IDF(C);

[0095] Where: TF is the term frequency, IDF(c,d) is the inverse document frequency of word c in the current dialogue d, and TF(c,d) represents the frequency of word c in the current dialogue d, reflecting the importance of the word. The calculation formula is:

[0096]

[0097] IDF(c) is the inverse document frequency of word c in the entire corpus, calculated using the following formula:

[0098]

[0099] Where K represents the total number of documents in the corpus, and DF(c) represents the number of documents containing word c. Inverse Document Frequency (IDF)(c) indicates the rarity of word c in the entire corpus; frequently occurring words have lower IDF values, while rare words have higher IDF values. A higher TF-IDF value indicates that word c is more important in the current conversation and has stronger disciplinary relevance. This means that the word is crucial for conveying core information and academic content. The TF-IDF metric is used to evaluate the use of academic terminology in conversations, especially in interdisciplinary dialogues, ensuring the accuracy and relevance of technical terms. This metric helps the system identify key academic vocabulary in conversations, thereby optimizing the professionalism and precision of expression, and promoting effective interdisciplinary knowledge exchange.

[0100] The evaluation and feedback module systematically collects the communication content generated in the interdisciplinary dialogue simulation module and summarizes, cleans, and organizes the dialogue characteristics across different groups and disciplinary backgrounds. Subsequently, based on the characteristics of interdisciplinary communication, six core indicators (compression rate, sentence similarity, N-gram redundancy, information entropy, TF-IDF, and lexical abstraction) are used to quantitatively analyze the dialogue content. Specifically, firstly, the dialogue data is classified, cleaned, and organized according to different roles and their disciplinary backgrounds to construct a disciplinary interaction matrix, clarifying the performance characteristics of each disciplinary combination in the dialogue. Next, for each group of dialogue data, the values ​​of the six indicators are calculated one by one: compression rate measures the conciseness and information density of the dialogue; sentence similarity reflects the consistency and diversity of expression; N-gram redundancy assesses the repetitiveness of expression; information entropy reveals the richness of content; TF-IDF analyzes the density of disciplinary terminology; and lexical abstraction measures the concreteness and abstractness of expression. Through the comprehensive analysis of these indicators, the expressive characteristics and information structure of dialogues across different disciplines can be quantified, further exploring the expressive changes and knowledge flow characteristics brought about by disciplinary differences in interdisciplinary communication. This indicator-based quantitative analysis helps to gain a deeper understanding of the specific impact of disciplinary background on dialogue, providing data support for optimizing interdisciplinary dialogue models.

[0101] The memory mechanism module enhances the agent's contextual consistency and information integration capabilities during dialogue. By recording and managing historical dialogue content, the system enables the agent to retain key information across multiple rounds of dialogue, thus providing a more coherent dialogue experience. This module not only supports short-term memory but also simulates long-term memory, optimizing dialogue coherence and information transmission efficiency through a memory decay mechanism. Specifically, before generating a new round of simulated dialogue content, the system retrieves historical key information highly relevant to the current context from the memory mechanism module, based on the current dialogue topic, context state, and learner role settings. This retrieval is based on semantic similarity matching algorithms (such as BERT vector representation and cosine similarity calculation) to perform semantic retrieval of historical dialogue content and selects several memory entries with a matching degree higher than a preset threshold as contextual enhancement information. Subsequently, the system injects the selected key information into the generation prompt of the large language model in the form of a structured prompt template, forming an enhanced input sequence. The enhanced Prompt template includes, but is not limited to: “Based on the points you mentioned in the previous round…”, “Please elaborate further based on your previous analysis of the XXX topic,” and “Continuing from your expression in the last dialogue, discuss the following questions,” thereby guiding the large language model to generate simulated dialogue content with stylistic consistency, semantic coherence, and knowledge continuity. Through the above mechanism, the memory module not only improves the dialogue stability and contextual adaptability in multi-turn interactions of the agent, but also enhances the system's expressive coherence and knowledge fusion capabilities in the construction of personalized learning paths.

[0102] The memory decay mechanism simulates the forgetting process in the human brain, gradually reducing reliance on irrelevant or repetitive information. This ensures that new information is processed first, preventing outdated information from interfering with the current conversation. To ensure the effectiveness and simplicity of memory management, the system assigns a timestamp to each generated dialogue message, marking its creation time and latest retrieval time. When a message is recalled or referenced in a conversation, its timestamp is updated to show its latest usage, thus tracking the frequency and timeliness of dialogue content in real time. In the memory decay mechanism, the system uses an exponential decay formula to update and discard historical dialogue content:

[0103] M(t) = M0·e -λt ;

[0104] Where M(t) represents the memory value after time t, i.e., the updated memory value, M0 is the initial memory strength, λ is the decay coefficient, and t is the time interval. According to this formula, the system can dynamically calculate the validity of the dialogue content. When M(t) is lower than the set threshold, the system will determine that the timeliness and importance of the content have decreased, and will directly delete it from the memory bank to free up storage space and avoid the long-term accumulation of inefficient information.

[0105] In storage management, the system sets up a structured storage sequence for each memory record, including: the associated role, dialogue time, dialogue content, and current memory value M(t). This structure provides the system with the basic information to determine the priority of memory retention. Simultaneously, when a memory is recalled in a subsequent dialogue, its timestamp and initial memory strength M0 are refreshed, thereby extending its retention period and improving its reusability and contextual consistency in multi-turn dialogues.

[0106] This mechanism achieves automatic filtering and dynamic updating through a time decay function, ensuring that the system retains key, frequently used information while promptly eliminating redundant content. This maintains the lightweight nature and responsiveness of the memory module, providing solid support for the context construction of generative agents during continuous dialogue. For example, in an interdisciplinary dialogue between S1 (computer science background) and H4 (sociology background), the memory mechanism module records key information in each round of dialogue in real time, including: the identities of the currently interacting agents (e.g., S1 and H4), the corresponding dialogue content, the timestamp of the last call, and the memory value M(t) dynamically calculated based on the time decay function. For instance, if H4 mentions "the structural shaping mechanism of learning motivation by social groups" in a certain round of dialogue, this content will be stored as a memory record by the system, along with a timestamp and initial memory strength M0. As time progresses, the system continuously updates its memory value according to a preset decay function.

[0107] When content related to the topic reappears in subsequent conversations, the system will automatically retrieve the memory entry and determine whether its current memory value M(t) is still higher than the set threshold. If the value is not lower than the threshold, it indicates that the content still has value, and the system will execute a callback operation, restoring the strength value of the memory to M(t) = 1, and refreshing its timestamp to the current call time t, thereby extending its retention period in the memory bank.

[0108] Through this mechanism, the system can dynamically identify, update, and prioritize the retention of important information that is frequently referenced in multi-turn dialogues, while effectively removing content that has not been called for a long time and whose value is gradually diminishing. This ensures that the memory database always remains efficient, concise, and highly relevant, thereby enhancing the agent's ability to maintain context and respond stably in complex situations.

[0109] like Figure 2 (a)- Figure 2 As shown in (f), this embodiment demonstrates the distribution of interaction paths and language behavior characteristics of different agent roles under a specific task theme during the interdisciplinary generative agent dialogue simulation process. Figure 2 (a)- Figure 2 The dialogue theme shown in (f) is: "You are now in an academic exchange dialogue scenario with the aim of learning from each other and improving yourselves. You will promote knowledge expansion and intellectual exchange through 20 rounds of in-depth exchanges with professional characteristics. Each member needs to present their disciplinary background, share core viewpoints, and explore interdisciplinary knowledge and ideas with other members. The exchange tasks include: 1) explaining important theories in your own field; 2) responding to the viewpoints of others and proposing interdisciplinary connections; 3) raising integrative questions and exploring solutions."

[0110] In this simulation, the system schedules agent agents from different disciplinary backgrounds to conduct multiple rounds of dialogue around a set task, demonstrating the evolution of language style and the trajectory of knowledge interaction in a multidisciplinary context, thereby verifying the effectiveness of the system in simulating real academic communication, maintaining the differences in disciplinary expression, and promoting knowledge integration.

[0111] Among them, "H1-H10" represent roles in the humanities and social sciences, corresponding to disciplines such as cognitive psychology, education, philosophy, history, and linguistics. These roles exhibit strong abstract and conceptual characteristics in their language style, often employing theoretical construction and academic argumentation, and tending to cite classic literature, academic frameworks, and historical experience to support their viewpoints.

[0112] "S1-S10" represent five agent roles with STEM backgrounds, covering technical fields such as computer science, artificial intelligence, data science, information engineering, and systems modeling. These agents are more structured and precise in their language, tending to use logical deduction, modeling frameworks, and quantitative analysis methods in their discussions, and emphasizing factual evidence and feasibility studies of technical solutions.

[0113] In this embodiment, to comprehensively evaluate the differences in language behavior and expressive characteristics in intradisciplinary and interdisciplinary dialogues, the system designed three types of dialogue strategies: the first type is intradisciplinary communication within the humanities, where “H1-H5” are paired with “H6-H10” one-to-one within the discipline; the second type is intradisciplinary communication within the science and engineering disciplines, where “S1-S5” are paired with “S6-S10” to simulate dialogue within the same discipline; and the third type is interdisciplinary communication, where the system sets “H1-H5” to be paired with “S6-S10” across disciplines to simulate cross-disciplinary communication scenarios in different disciplinary backgrounds.

[0114] To facilitate comparison and analysis, the system assigns concise labels to the three types of dialogue strategies: "HH" for internal communication within the humanities, "SS" for internal communication within the science and engineering disciplines, and "HS" for interdisciplinary communication. Furthermore, a dialogue strategy that incorporates a memory mechanism module based on interdisciplinary dialogue is named "HS-M," used to explore the impact of long-term memory on the stability of language behavior and the consistency of knowledge citation.

[0115] Through this strategy design, the system can structurally compare and analyze the differences in language expression, contextual adaptability, and information organization characteristics of different subject combinations, while providing clear and controllable grouping criteria for subsequent multi-dimensional quantitative assessment.

[0116] In each dialogue set, the system sets a fixed task topic and conducts multiple rounds of interaction. The dialogue content generated from each pairing is then recorded and stored in a structured manner. To improve data stability and result reliability, each dialogue strategy is tested three times, and the results of each round are averaged. Finally, the system quantitatively evaluates the dialogue samples under different strategies based on six core language metrics: compression rate, sentence similarity, N-gram redundancy, information entropy, TF-IDF, and lexical abstraction.

[0117] Furthermore, to more intuitively demonstrate the differences in language expression characteristics among agents from different disciplines, representative role groups H1-H5 (humanities and social sciences) and S6-S10 (science and engineering) were selected, and radar charts for six language indicators were created to achieve multi-dimensional visual comparison. These six indicators include: Compression Ratio, Cosine Similarity, N-gram Redundancy, Entropy, TF-IDF value, and Lexical Abstraction. This visualization analysis clearly reflects the performance differences of each agent group in terms of language structure, information density, and expressive diversity.

[0118] Building upon this, a statistical analysis was further conducted on the performance of the humanities group (H1-H5) and the science and engineering group (S6-S10) across six language indicators. The mean and standard deviation for each indicator were calculated for each group to characterize the overall performance trends and internal consistency of language styles across different disciplinary groups. This analysis helps reveal the systematic differences among disciplinary groups in terms of expression density, information redundancy, language diversity, and level of abstraction. Table 1 shows the statistical analysis of the mean and standard deviation of the six core language indicators for humanities agents under different communication modes, and Table 2 shows the statistical analysis of the six core language indicators for science and engineering agents under different communication modes.

[0119] Table 1

[0120] index HH HS HS-M Compression ratio 0.40±0.07 0.47±0.02 0.54±0.06 Similarity 0.61±0.06 0.57±0.07 0.57±0.10 Redundancy 0.51±0.06 0.50±0.05 0.54±0.08 Information entropy 0.56±0.07 0.50±0.07 0.50±0.05 TF-IDF value 0.47±0.07 0.48±0.04 0.39±0.08 Abstraction 0.45±0.07 0.51±0.04 0.45±0.04

[0121] Table 2

[0122] index SS HS HS-M Compression ratio 0.49±0.06 0.50±0.05 0.57±0.02 Similarity 0.62±0.04 0.60±0.07 0.63±0.03 Redundancy 0.50±0.05 0.47±0.09 0.50±0.04 Information entropy 0.50±0.05 0.54±0.05 0.49±0.04 TF-IDF value 0.44±0.02 0.43±0.07 0.40±0.03 Abstraction 0.46±0.08 0.44±0.04 0.45±0.05

[0123] Experimental results show that, compared to intradisciplinary dialogue, interdisciplinary communication performs better in terms of language compression rate and adaptability, with lower information redundancy, demonstrating higher information density and language transfer ability. After introducing a memory mechanism, the agent can optimize language expression while maintaining dialogue coherence, enhancing the consistency of knowledge retrieval and the stability of expression. Agents from different disciplines also exhibit differentiated characteristics in their response to the memory mechanism; humanities agents are more prone to a decrease in abstractness, while STEM agents are better able to maintain or enhance their abstract expression capabilities.

[0124] In summary, this invention not only verifies the feasibility and effectiveness of generative agents in simulating interdisciplinary dialogue, but also provides a quantifiable and scalable evaluation and optimization method, which is of great significance for promoting interdisciplinary collaboration and personalized learning path recommendation in intelligent education.

[0125] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A personalized learning guidance system based on a generative agent model, characterized in that, include: Generative intelligent agent role-playing module, interdisciplinary dialogue simulation module, evaluation and feedback module, and memory mechanism module; The generative agent role-playing module is used to simulate learner roles in different disciplinary backgrounds based on a large language model; The interdisciplinary dialogue simulation module is used to simulate the interaction and communication between the learner roles under preset task or theme conditions, and to obtain the simulated dialogue content generated by the learner roles. The evaluation and feedback module is used to perform quantitative analysis on the simulated dialogue content and output evaluation indicators to optimize the personalized learning path. The memory mechanism module is used to store and manage historical simulated dialogue content, obtain key information, and input the key information into the large language model to optimize the simulated dialogue content generated by the learner role.

2. The personalized learning system based on a generative agent model according to claim 1, characterized in that, Simulating learner roles in different subject backgrounds based on a large language model includes: obtaining preset prompt words, inputting the preset prompt words into the large language model, and simulating learner roles in different subject backgrounds.

3. The personalized learning system based on a generative agent model according to claim 2, characterized in that, The simulated dialogue content generated by the learner role includes: The learner role setting input information, dialogue context input information, and knowledge retrieval input information are input into the large language model to obtain the simulated dialogue content generated by the learner role. The learner role setting input information includes subject background, language style, expression characteristics, and personalized parameters. The dialogue context input information includes the speech content of the current round, historical interaction records, and context state. The knowledge retrieval input information is obtained from the knowledge base.

4. The personalized learning system based on a generative agent model according to claim 3, characterized in that, Obtaining the knowledge retrieval input information from the knowledge base includes: retrieving matching information from the knowledge base using RAG technology to obtain the knowledge retrieval input information.

5. The personalized learning system based on a generative agent model according to claim 1, characterized in that, The interdisciplinary dialogue simulation module also uses AgentScope to control the speaking order, interaction rounds, and behavioral boundaries of learner roles.

6. The personalized learning system based on a generative agent model according to claim 1, characterized in that, The simulated dialogue content is quantitatively analyzed, including by using information redundancy, compression rate, sentence similarity, lexical abstraction, TF-IDF value, and information entropy to analyze the simulated dialogue content.

7. The personalized learning system based on a generative agent model according to claim 1, characterized in that, Optimizing personalized learning paths includes: The learner's role is evaluated based on the evaluation indicators to construct a personalized competency profile; The interaction relationships between the learner roles are analyzed based on the role interaction data and the personalized ability profiles to identify the optimal dialogue pairings. Based on the optimal dialogue pairing combination, the role scheduling strategy in the dialogue rounds is dynamically optimized to achieve optimization of the personalized learning path.

8. The personalized learning system based on a generative agent model according to claim 1, characterized in that, The storage and management of the historical simulated dialogue content, and the acquisition of key information, include: A timestamp is set for the historical simulated dialogue content, wherein the timestamp marks the time when the historical simulated dialogue content was generated and the latest retrieval time; A memory decay mechanism is used to calculate the memory value of the historical simulated dialogue content to obtain the updated memory value; The key information is obtained based on the timestamp and the updated memory value.

9. The personalized learning system based on a generative agent model according to claim 8, characterized in that, Obtaining the key information based on the timestamp and the updated memory value includes: When the updated memory value is lower than a preset threshold, the historical simulated dialogue content is marked as low priority and deleted. When the simulated dialogue content of the history is recalled, the timestamp and initial memory strength are updated, and the simulated dialogue content of the history is retained for an extended period.

Images