Agent-based education evaluation method and system

Abstract

The invention discloses an Agent-based education evaluation method and system, and relates to the technical field of artificial intelligence and education evaluation, and the method comprises the steps: obtaining historical evaluation data and knowledge graph labeling data of students, constructing a student knowledge state matrix, calculating a mastering probability value and a forgetting attenuation coefficient of each knowledge point, and generating an evaluation question recommendation sequence; inputting the question text and the student answering text into a multi-Agent collaborative evaluation model to generate a multi-dimensional scoring result; dynamically updating the knowledge point mastering probability value of the student in the continuous answering process, calculating the mastering degree change rate of each knowledge point node, constructing an adaptive difficulty adjustment function, and generating a question difficulty parameter and a knowledge point coverage range parameter of the next round of evaluation; and taking the question difficulty parameter and the knowledge point coverage range parameter as constraint conditions, screening a candidate question set meeting the conditions from a question bank, and generating a self-adaptive evaluation path and a capability diagnosis report. The invention also discloses a method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and education evaluation technology, in particular to an education evaluation method and system based on Agent. BACKGROUND

[0002] With the deep integration of artificial intelligence and education informatization, intelligent education evaluation technology has gradually become a key component in personalized learning systems. Traditional online education evaluation systems mostly rely on fixed question banks and static test question generation strategies, and achieve student knowledge mastery evaluation through preset difficulty levels or manually configured question association rules. However, such methods generally have response lag and insufficient accuracy when facing complex situations such as learner knowledge structure differentiation, learning path dynamic changes, and nonlinear evolution of cognitive state. In recent years, evaluation methods based on knowledge graphs and intelligent recommendation algorithms have been widely studied, which establish the association between knowledge points to assist in question recommendation and knowledge tracking, but most methods still focus on static data fitting, making it difficult to realize real-time dynamic updating of student knowledge mastery state. In addition, the subjective question evaluation process in the evaluation process often relies on single model scoring mechanism, lacks multi-angle semantic alignment and cross-Agent collaborative evaluation mechanism, resulting in insufficient consistency and interpretability of the evaluation results, making it difficult to form a high-credibility intelligent evaluation system. Therefore, how to construct an education evaluation method that can integrate knowledge graph reasoning, dynamic cognitive tracking, and multi-Agent collaborative evaluation mechanism has become an important direction in the field of intelligent education research.

[0003] CN118132741B discloses an Agent-based autonomous learning method and system, which includes: text preprocessing and data cleaning of acquired teaching materials and supplementary data, and storing in a database; generating instructions according to user selection system and inputting related information, using a pre-trained intent recognizer to identify related information, obtaining data from the database and generating corresponding test papers; correcting user's answering situation, calculating scores; and combining inputted related information and test paper information to generate a user's this-time test learning situation report; according to the learning situation report and related information, extracting relevant teaching materials, supplementary original position corresponding content from the database through retrieval enhancement generation technology, and writing learning supplementary and exercises for weak items; generating a user's special review plan to assist the user in solving weak knowledge points. This method can realize personalized teaching recommendation for each user, but its core is still limited to the level of content retrieval and report generation, lacking dynamic quantification of student knowledge mastery state and multi-dimensional reasoning mechanism based on knowledge graph. At the same time, the Agent of an Agent-based autonomous learning method and system is mainly used for task distribution and data calling control, and the semantic interaction and collaborative evaluation among multiple Agents are not realized, making it difficult to intelligently identify and dynamically adjust student answering behavior at the semantic level in complex subjective evaluation scenarios.

[0004] In summary, the existing Agent-based education evaluation technology generally has problems such as lagging knowledge state updating mechanism, fixed evaluation question recommendation path, and insufficient reliability of subjective question evaluation results. The present application provides an Agent-based education evaluation method and system, which realizes high-precision quantification and intelligent adaptive adjustment of evaluation results. SUMMARY

[0005] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0006] In view of the problems of existing Agent-based education evaluation technology, such as lagging knowledge state updating, insufficient evaluation path adaptability, and low reliability of subjective question scoring results, the present application is proposed.

[0007] Therefore, the problem to be solved by the present application is how to realize dynamic updating of student knowledge mastery state and generation of adaptive evaluation path by using a multi-Agent collaborative evaluation mechanism and a Bayesian knowledge tracking algorithm, so as to build a high-precision, personalized and interpretable intelligent education evaluation system.

[0008] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the embodiments of the present application provide an Agent-based education evaluation method, which comprises, obtaining historical evaluation data and knowledge graph annotation data of students, constructing a student knowledge state matrix, and calculating mastery probability values and forgetting decay coefficients of each knowledge point by using a Bayesian knowledge tracking algorithm to generate a question recommendation sequence; inputting question texts and student answer texts in the question recommendation sequence into a multi-Agent collaborative evaluation model to generate multi-dimensional scoring results; based on the multi-dimensional scoring results, dynamically updating knowledge point mastery probability values in a continuous answering process, calculating a mastery degree change rate of each knowledge point node, constructing an adaptive difficulty adjustment function, and generating question difficulty parameters and knowledge point coverage range parameters for the next round of evaluation; using the question difficulty parameters and the knowledge point coverage range parameters as constraint conditions, screening a candidate question set meeting the conditions from a question bank, and using a reinforcement learning Agent to optimally select and sort the candidate question set according to a real-time state of the student knowledge state matrix to generate an adaptive evaluation path and a capability diagnosis report.

[0009] In a second aspect, the embodiments of the present application provide an agent-based education evaluation system, which comprises: a knowledge modeling module configured to acquire historical evaluation data and knowledge graph annotation data of a student, construct a knowledge state matrix of the student, calculate a mastery probability value and a forgetting decay coefficient of each knowledge point by using a Bayesian knowledge tracking algorithm, and generate a recommended sequence of evaluation questions; a collaborative evaluation module configured to input question texts in the recommended sequence of evaluation questions and student answer texts into a multi-agent collaborative evaluation model to generate a multi-dimensional score result; a dynamic updating module configured to dynamically update a knowledge point mastery probability value of the student in a continuous answering process based on the multi-dimensional score result, calculate a mastery degree change rate of each knowledge point node, construct an adaptive difficulty adjustment function, and generate a question difficulty parameter and a knowledge point coverage range parameter for a next round of evaluation; a path generating module configured to select a candidate question set meeting a condition from a question bank by taking the question difficulty parameter and the knowledge point coverage range parameter as constraint conditions, perform optimal selection and sorting on the candidate question set according to a real-time state of the knowledge state matrix of the student by using a reinforcement learning agent, and generate an adaptive evaluation path and a capability diagnosis report.

[0010] In a third aspect, the embodiments of the present application provide a computer device, which comprises a memory and a processor, and the memory stores a computer program, wherein the processor implements any step of the agent-based education evaluation method described above when executing the computer program.

[0011] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement any step of the agent-based education evaluation method described above.

[0012] Compared with the prior art, the present application has the beneficial effects that: by fusing historical evaluation data and knowledge graph annotation data to construct a student knowledge state matrix, and coupling a Bayesian knowledge tracking algorithm and a forgetting decay coefficient, the precise quantitative modeling and dynamic tracking of the student's knowledge mastery state are realized, and the blindness of traditional fixed item bank recommendation is avoided; by using a multi-Agent collaborative evaluation model combined with a knowledge point attention mechanism, the limitations of traditional single score evaluation are broken through, multi-dimensional score results including overall scores and subdivision indexes are generated, not only correct or wrong is judged, but also complex situations such as partial correctness and concept confusion are identified; based on the multi-dimensional score results, the knowledge point mastery probability is updated in real time and dynamically, the intelligent feedback and difficulty adjustment of the evaluation process are realized by calculating the mastery degree change rate and constructing an adaptive difficulty adjustment function, ensuring that the difficulty of the questions is always in the student's zone of proximal development, and through preposition dependency analysis to find the knowledge system fault, the evaluation process itself has a learning promotion effect; the difficulty parameter and the knowledge point coverage range are used as constraint conditions to screen the candidate question set, and a reinforcement learning Agent is used for global optimization sorting, not only the question difficulty is optimized, but also the knowledge point coverage integrity and the logical correlation before and after are considered, an adaptive evaluation path and an ability diagnosis report are generated; compared with the traditional standardized evaluation, the invalid question pushing is significantly reduced, the evaluation efficiency and the diagnosis accuracy are greatly improved, and the paradigm shift from selection evaluation to development evaluation is realized. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Figure 1 Flow chart of the Agent-based education evaluation method. DETAILED DESCRIPTION

[0014] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0015] Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0016] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0017] As described in the above background, the existing Agent-based education evaluation technology has the problems of lagging knowledge state updating, insufficient evaluation path self-adaptability and low reliability of subjective question scoring results. In view of the above problems, the present application provides an Agent-based education evaluation method.

[0018] Reference Figure 1 , Figure 1 is a flowchart of an Agent-based education evaluation method according to an embodiment of the present application. As shown in Figure 1 , in an Agent-based education evaluation method, it comprises: S1: obtaining historical evaluation data and knowledge graph annotation data of students, constructing a student knowledge state matrix, and calculating the mastery probability value and forgetting decay coefficient of each knowledge point by using a Bayesian knowledge tracking algorithm to generate an evaluation question recommendation sequence; S1.1: extracting the historical evaluation data of students from the education evaluation database, and obtaining the knowledge graph annotation data corresponding to the historical evaluation data; It should be noted that the historical evaluation data includes the answer record, answering time, score and wrong question distribution of the student in the past several evaluation periods; the knowledge graph annotation data includes knowledge point node identification, knowledge point hierarchical relationship and knowledge point dependency relationship; the historical evaluation data is sorted according to the time stamp to generate a time sequence evaluation record sequence.

[0019] S1.2: based on the time sequence evaluation record sequence, the historical answering times and correct answering times of the student for each knowledge point node are counted, and the initial mastery rate is calculated; It should be noted that the initial mastery rate is equal to the correct answering times divided by the historical answering times; each element in the student knowledge state matrix represents the initial mastery rate of the corresponding student for the corresponding knowledge point.

[0020] S1.3: based on the knowledge point node identification in the knowledge graph annotation data, a student knowledge state matrix of MxK dimensions is constructed, where M is the number of students and K is the total number of knowledge points; S1.4: traversing the row vector corresponding to the current student in the student knowledge state matrix, identifying and judging the initial mastery rate of each knowledge point node, and obtaining a weak knowledge point set; Specifically, when the initial mastery rate is less than the mastery rate threshold, the knowledge point node is marked as a weak knowledge point, and is added to the weak knowledge point set, and a high-priority evaluation identifier is set for the knowledge point node; when the initial mastery rate is greater than or equal to the mastery rate threshold, the knowledge point node is marked as a well-mastered knowledge point, and is added to the well-mastered knowledge point set, and a low-priority evaluation identifier is set for the knowledge point node; based on the knowledge point hierarchical relationship in the knowledge graph annotation data, the distribution quantity of the weak knowledge point set on each knowledge level is counted, and a weak knowledge point distribution vector is generated.

[0021] It should be noted that the mastery rate threshold is determined based on the standardization requirements in the education evaluation field and the course teaching objectives; the dimension of the weak knowledge point distribution vector is consistent with the number of knowledge levels.

[0022] S1.4: Based on the weak knowledge point set, a history answer sequence corresponding to the knowledge point node is extracted from the time sequence evaluation record sequence, and the mastery probability value is updated in time sequence; Specifically, the prior parameters of the Bayesian knowledge tracking algorithm are set, including the initial mastery probability, the learning gain rate, the guessing probability and the error probability; the initial mastery probability is assigned with the initial mastery rate of the corresponding element in the student knowledge state matrix.

[0023] Further, if the tth answer is correct, the mastery probability value at the t+1 moment is calculated by using the Bayesian update formula; if the tth answer is wrong, the mastery probability value is adjusted synchronously; by traversing the complete history answer sequence, the dynamic calculation and update of the mastery probability value of all weak knowledge point nodes are completed.

[0024] Further, the mastery probability value at the t+1 moment is specifically as follows: ; Wherein, is the knowledge point mastery probability value at the t+1 moment, is the mastery probability value at the t moment, is the error probability of the student in the mastery state, is the guessing probability of the student in the unmastered state, is a time decay adjustment coefficient, is the time interval between the tth answer and the t+1 answer, is a question difficulty influence function, is the difficulty coefficient of the current question, the value range of which is 0 to 1, is a cumulative answer number adjustment function, is the cumulative answer number of the knowledge point, is a learning gain weight coefficient, is a learning efficiency amplification factor, is the learning investment degree index when answering for the t-th time, which is calculated by the ratio of the answering time length and the standard time length, is the hyperbolic tangent function.

[0025] S1.5: Based on the time sequence evaluation record sequence, an Ebbinghaus forgetting curve model is used to construct a forgetting decay function, and a forgetting decay coefficient of the knowledge point node is obtained. It should be noted that the Ebbinghaus forgetting curve model should be based on the classical research results of memory psychology and the mathematical expression of forgetting law; the model is derived from Ebbinghaus's experimental findings on the forgetting law of human memory, which describes the trend of knowledge being forgotten over time through an exponential decay function; in the education evaluation system, the actual answering interval of the student and the characteristics of the knowledge point need to be adjusted to accurately reflect the natural decay of knowledge mastery over time, and to provide a scientific basis for the selection of evaluation time.

[0026] Specifically, based on the weak knowledge point set, the average time interval of adjacent answers of this knowledge point node in the time sequence evaluation record sequence is calculated, and the forgetting decay function is substituted to obtain the forgetting decay coefficient of this knowledge point node.

[0027] Further, the specific formula of the forgetting decay coefficient is as follows: ; wherein, is the forgetting decay coefficient of the k-th knowledge point, is the average time interval of adjacent answers of the k-th knowledge point, is the memory stability parameter of the k-th knowledge point, is the review reinforcement coefficient, is the historical review times of the knowledge point, is the knowledge point importance weight function, is the importance score of the k-th knowledge point in the knowledge graph, is the knowledge correlation adjustment parameter, is the sigmoid activation function, is the total number of other knowledge points associated with the k-th knowledge point, is the dependency weight of the j-th associated knowledge point, is the association strength between the k-th knowledge point and the j-th associated knowledge point, is the error answer penalty coefficient, is the historical error answer times of the knowledge point, is the penalty smoothing constant.

[0028] S1.6: Multiply the mastery probability value of each knowledge point node in the weak knowledge point set by the corresponding forgetting decay coefficient to obtain a comprehensive weakness score; It should be noted that the smaller the comprehensive weakness score value is, the more the knowledge point node needs to be prioritized.

[0029] S1.7: Topological sorting and weighted sorting are performed in combination with the knowledge point dependency relationship to generate a knowledge point priority sequence, and questions associated with the knowledge point nodes with high priority are retrieved from the question bank according to the knowledge point priority sequence to generate a test question recommendation sequence; Further, according to the knowledge point dependency relationship in the knowledge graph annotation data, the topological sorting of the weak knowledge point set is performed to ensure that the preceding dependent knowledge points are sorted before the subsequent knowledge points, and the weighted sorting is performed in combination with the comprehensive weakness score to generate a knowledge point priority sequence; based on the knowledge point priority sequence, questions associated with the knowledge point nodes with high priority are retrieved from the question bank, and for each knowledge point node, a question with a difficulty coefficient matching the mastery probability value of the knowledge point node is selected; the retrieved questions are arranged in order according to the knowledge point priority sequence to generate a test question recommendation sequence containing question identification, knowledge point annotation and difficulty coefficient.

[0030] It should be noted that the length of the test question recommendation sequence is determined according to the preset test duration and the average question answering time.

[0031] S2: Input the question text and student answer text in the test question recommendation sequence into the multi-Agent collaborative evaluation model to generate a multi-dimensional score result; S2.1: Extract the question text from the test question recommendation sequence, obtain the student answer text submitted for the question text, and perform preprocessing operations on the question text and the student answer text to generate question text word sequences and student answer text word sequences; It should be noted that the question text includes the stem content, question requirements and reference answers; the student answer text contains the student's textual answer content and the answer timestamp; the preprocessing operation includes removing special symbols, unifying character encoding and word segmentation processing.

[0032] Specifically, the multi-Agent collaborative evaluation model is constructed based on an improved BERT semantic understanding network; the improved BERT semantic understanding network adds a knowledge point perception layer based on the standard BERT model, and adds a knowledge point attention sublayer after the self-attention layer of the BERT encoder.

[0033] It should be noted that the knowledge point perception layer contains a plurality of knowledge point embedding vectors, and the knowledge point embedding vectors correspond one-to-one to the knowledge point nodes in the knowledge graph annotation data; the knowledge point attention sublayer calculates the correlation between the input text representation and the knowledge point embedding vector through a learnable weight matrix to generate a knowledge point perception weight distribution.

[0034] S2.2: Input the student's answer text word sequence into the multi-Agent collaborative evaluation model, calculate the attention score between the student's answer text word sequence and the knowledge point embedding vector through the knowledge point attention sub-layer, and extract the key knowledge point feature vector; Furthermore, positional encoding and paragraph encoding are added to the input layer, which is then processed by an L-layer Transformer encoder. The attention score between the student's answer text word sequence and the knowledge point embedding vector is calculated through the knowledge point attention sublayer to obtain the knowledge point activation matrix.

[0035] Furthermore, based on the knowledge point activation matrix, max pooling is performed on each knowledge point dimension to extract the maximum activation value of that knowledge point in the word sequence of the student's answer text. The maximum activation value of the knowledge point is then weighted and summed with the corresponding knowledge point embedding vector, with the weight coefficient being the normalized maximum activation value, to obtain a weighted knowledge point representation vector. This weighted knowledge point representation vector is then concatenated with the [CLS] label output vector of the last layer of BERT, and mapped to the space through a fully connected layer to generate the key knowledge point feature vector of the student's answer.

[0036] S2.3: Employ a multi-head attention mechanism to calculate the semantic similarity matrix between the feature vectors of key knowledge points and the feature vectors of standard answers; Specifically, the standard answer text corresponding to the question text is extracted from the question bank. The standard answer text is preprocessed to generate a standard answer text word sequence, which is then input into the improved BERT semantic understanding network to generate the knowledge point activation matrix and weighted knowledge point representation vector of the standard answer in sequence, and finally the feature vector of the standard answer is obtained.

[0037] It should be noted that the standard answer text is pre-annotated by domain experts and includes knowledge point coverage tags.

[0038] Furthermore, for each attention head, the query matrix, key matrix, and value matrix are calculated, and the attention score matrix is ​​calculated. The attention score matrices of multiple attention heads are concatenated to generate a semantic similarity matrix.

[0039] S2.4: Based on the semantic similarity matrix, slice the data according to the knowledge point dimension, and extract the submatrix of the corresponding knowledge point region from the semantic similarity matrix according to the knowledge point node identifier associated with the question text; S2.5: Calculate the row and column mean of the sub-matrix, and at the same time calculate the mean of the diagonal elements of the semantic similarity matrix as the overall similarity score to generate a multi-dimensional scoring feature vector; S2.6: input the multi-dimensional score feature vector into a score mapping network, wherein the score mapping network comprises two fully connected layers and an output layer; the output layer generates a score probability distribution of five dimensions, and according to the highest probability in the score probability distribution, an overall score level is obtained, while retaining each score in the multi-dimensional score feature vector as a subdivided score indicator, thereby generating a multi-dimensional score result; It should be noted that the row mean represents the coverage score of the student's answer on the knowledge point; the column mean represents the matching score of the student's answer and the standard answer on the knowledge point; the number of neurons in the first fully connected layer is 128, and the activation function adopts ReLU; the number of neurons in the second fully connected layer is 64; the softmax output layer generates a score probability distribution of five dimensions, corresponding to excellent, good, medium, passing and failing five levels respectively; the multi-dimensional score result is stored in association with the question identification of the question text.

[0040] S3: based on the multi-dimensional score result, dynamically updating the knowledge point mastery probability value of the student in the continuous answering process, calculating the mastery change rate of each knowledge point node, constructing an adaptive difficulty adjustment function, and generating the question difficulty parameter and knowledge point coverage range parameter of the next round of evaluation; S3.1: extracting the overall score level and subdivided score indicators from the multi-dimensional score result, and associating and mapping the overall score level with the knowledge point annotation of the corresponding question in the question recommendation sequence, to generate an answer result label sequence; Specifically, according to the coverage score and the matching score in the subdivided score indicators, the current mastery probability value of the corresponding knowledge point node in the student knowledge state matrix is located, and when the coverage score is greater than a preset coverage threshold and the matching score is greater than a preset matching threshold, it is marked as a successful answer, otherwise it is marked as a failed answer, to generate an answer result label sequence.

[0041] It should be noted that the preset matching threshold and the preset coverage threshold are determined based on the evaluation standard of natural language processing and semantic understanding; the matching threshold is used to judge whether the similarity of the student's answer content and the standard answer in the semantic level reaches an acceptable level, and the expert annotation data accumulated in the training process of the multi-Agent collaborative evaluation model is referred to, as well as the definition standard of the quality of answers of different levels; the coverage threshold focuses on whether the student's answer covers the core knowledge points required by the question, and is set based on the annotation of the importance of each knowledge point in the knowledge graph and the expectation of the completeness of the answer in the question design.

[0042] S3.2: based on the answer result label sequence, extracting the current mastery probability value of the corresponding knowledge point node in the student knowledge state matrix; Further, if the answer result is marked as a successful answer, the updated mastery probability value is calculated by using the Bayesian posterior update formula, and the specific formula is as follows: ; wherein, is the updated mastery probability value after a successful answer, is the current mastery probability value, is a joint regulation function of item quality and time, is a discrimination index of the item, t is a normalized ratio of the answer duration and the standard duration, is a historical performance weighting function, is the historical accuracy of the knowledge point, is a cognitive load regulation function, is a cognitive load index of the current assessment calculated through item complexity, is a learning process gain weight, is a time decay constant, is a time variable, is a learning intensity function at time u.

[0043] Further, if the answer result is marked as a failed answer, the updated mastery probability value is calculated as: ; wherein, is the updated mastery probability value after a failed answer, is the current mastery probability value, is a failure penalty intensity coefficient, is a difficulty regulation function, is a difficulty coefficient of the item, is a probability state regulation factor, is an error frequency penalty function, is the number of consecutive errors in recent times, is the total number of answers, is a matching degree penalty weight, is a matching degree sensitivity parameter, is the matching degree score of the actual answer, is a matching degree threshold value.

[0044] Preferably, the updated mastery probability value is written back to the corresponding element position of the student knowledge state matrix.

[0045] S3.3: Obtain the continuous answer records of the student in the current assessment period, apply the Bayesian posterior probability update operation in time sequence one by one, record the change amount of the mastery probability value before and after each update, and generate a knowledge point mastery probability tracking sequence; It should be noted that the continuous answer record is arranged in the order of answer timestamp, contains question identification, knowledge point annotation, answer duration and answer result marker sequence; the length of the knowledge point mastery probability tracking sequence is equal to the number of answers of the knowledge point node in the current evaluation period.

[0046] S3.4: Traverse all knowledge point nodes in the student knowledge state matrix, extract the initial mastery probability value and the final mastery probability value of the knowledge point from the knowledge point mastery probability tracking sequence, and obtain the mastery degree change rate; Specifically, the mastery probability value difference is calculated, the first answer timestamp and the last answer timestamp of the knowledge point in the current evaluation period are obtained, and the time span is calculated; the mastery probability value difference is divided by the time span to obtain the mastery degree change rate of the knowledge point node.

[0047] S3.5: Based on the smooth mastery degree change rate vector, classify the state of each knowledge point node in the knowledge graph annotation data, and generate a knowledge point state distribution statistical table; Further, set the sliding window length, extract the mastery degree change rate historical data of the student in the last several evaluation periods from the historical evaluation database; for each knowledge point node, construct a mastery degree change rate sequence, and use a weighted moving average algorithm to smooth the mastery degree change rate sequence, the weight coefficient is calculated using an exponential decay function, and a smooth mastery degree change rate vector is generated.

[0048] Further, if the mastery degree change rate is greater than the first change rate threshold, the knowledge point node is marked as a fast progress node; if the mastery degree change rate is less than or equal to the first change rate threshold and greater than the second change rate threshold, the knowledge point node is marked as a stable mastery node; if the mastery degree change rate is less than or equal to the second change rate threshold, the knowledge point node is marked as a regression risk node; the number of knowledge point nodes in each state category is counted, and a knowledge point state distribution statistical table is generated.

[0049] It should be noted that the first change rate threshold and the second change rate threshold are determined based on the judgment standard of learning progress monitoring and education intervention opportunity. The first change rate threshold is used to identify students with significant learning progress, and a critical value that can accurately capture positive learning effect is set by referring to the development law of normal learning curve and the characteristic performance of fast progress students in historical data; the second change rate threshold is used to warn learning regression risk, and a warning line that can timely find students with stagnant or declining mastery degree is set by combining the theory of forgetting curve and the critical point of knowledge decay; the knowledge point state distribution statistical table includes a fast progress node list, a stable mastery node list and a regression risk node list.

[0050] S3.6: Construct an adaptive difficulty adjustment function according to the knowledge point state distribution statistical table and the student knowledge state matrix, and generate a difficulty adjustment coefficient vector; Specifically, the weight coefficients are dynamically adjusted based on the status classification of knowledge point nodes. If a knowledge point node is a rapidly progressing node, a smaller weight is set to stabilize the increase in difficulty; if a knowledge point node is a regression risk node, a larger weight is set to increase the difficulty adjustment intensity. The output value of the adaptive difficulty adjustment function for all knowledge point nodes is calculated, and a difficulty adjustment coefficient vector is generated.

[0051] S3.7: Map the difficulty adjustment coefficient vector to the predefined question difficulty level system in the question bank to generate a question difficulty parameter sequence; It should be noted that the difficulty level system includes five difficulty levels, with the corresponding difficulty coefficient ranges as follows: [0, 0.2) is marked as easy, [0.2, 0.4) is marked as relatively easy, [0.4, 0.6) is marked as medium, [0.6, 0.8) is marked as relatively difficult, and [0.8, 1.0] is marked as very difficult. Specifically, for the k-th knowledge point node, the difficulty coefficient and difficulty level identifier of the corresponding question are determined based on the output value of the adaptive difficulty adjustment function; combined with the completeness score in the multi-dimensional scoring results, when the completeness score is lower than the preset completeness threshold, the difficulty coefficient of the corresponding knowledge point is lowered to reduce frustration, and a question difficulty parameter sequence is generated.

[0052] S3.8: Extract the list of regression risk nodes and the set of weak knowledge points from the knowledge point status distribution statistics table, merge the two lists and remove duplicate knowledge point node identifiers to generate a set of knowledge points to be prioritized for coverage; S3.9: Based on the knowledge point dependencies in the knowledge graph annotation data, perform a pre-dependency analysis on each knowledge point node in the priority coverage knowledge point set; S3.10: When the mastery probability value of the previous dependent knowledge point is lower than the mastery threshold, the previous dependent knowledge point is added to the priority coverage knowledge point set, and the hierarchical depth weight of each knowledge point node is calculated according to the hierarchical relationship of the knowledge points to generate the knowledge point coverage range parameter.

[0053] S4: Using the question difficulty parameter and knowledge point coverage parameter as constraints, select a set of candidate questions that meet the conditions from the question bank, and use the reinforcement learning agent to perform optimal selection and sorting of the candidate question set according to the real-time status of the student's knowledge state matrix, generating an adaptive assessment path and ability diagnosis report.

[0054] Specifically, the question difficulty coefficient and difficulty level identifier in the question difficulty parameter sequence are used as the first constraint, and the priority knowledge point set and examination weight coefficient in the knowledge point coverage parameter are used as the second constraint; a set of basic rules for question selection is set. It should be noted that the basic rule set includes that the knowledge point node associated with the question must be in the priority coverage knowledge point set, the difficulty coefficient of the question needs to match the difficulty coefficient of the corresponding knowledge point in the question difficulty parameter sequence and the matching error is not more than ±0.15, and the historical use frequency of the question is lower than the upper frequency threshold to avoid question repetition; according to the available time length of the current evaluation and the average answer time of the question, the upper limit of the number of questions is calculated, and the upper limit of the number of questions is taken as the third constraint condition to generate a multi-dimensional constraint condition vector.

[0055] S4.1: Connect the question database interface, execute query retrieval of all questions meeting the knowledge point coverage requirement and difficulty matching requirement according to the first constraint condition and the second constraint condition in the multi-dimensional constraint condition vector, and generate a candidate question set; Further, for each question in the retrieval result, verify whether the historical use frequency meets the frequency limit in the basic rule set, and eliminate the questions whose use frequency exceeds the upper frequency threshold; count the knowledge point distribution associated with the remaining questions, calculate the knowledge point coverage rate, and when the knowledge point coverage rate is lower than the minimum coverage rate threshold, relax the difficulty coefficient matching error to ±0.25 and re-retrieve to generate a candidate question set.

[0056] Preferably, the specific formula of the knowledge point coverage rate is as follows: ; Wherein, is the knowledge point coverage rate, is the number of covered knowledge points, is the total number of knowledge points in the priority coverage knowledge point set, is the examination weight coefficient of the i-th knowledge point, is the coverage sensitivity parameter of the i-th knowledge point, is the current mastery probability value of the i-th knowledge point, is the hierarchical importance function, is the hierarchical depth of the i-th knowledge point in the knowledge graph, is the number of knowledge points involving dependency relationship, is the mastery probability increment of the k-th dependent knowledge point, is the dependency strength coefficient of the k-th knowledge point.

[0057] S4.2: Traverse each question in the candidate question set to extract the multi-dimensional feature attributes of the question; It should be noted that the multi-dimensional feature attributes include the question difficulty coefficient, the knowledge point node identifier list associated with the question, the question discrimination index, the question reliability index, the question average answer time and the question type identifier; the dimension of the candidate question feature vector is equal to the number of feature attributes plus 1.

[0058] Furthermore, for each question, based on the list of knowledge point node identifiers associated with it, the corresponding examination weight coefficient is found from the knowledge point coverage parameter, and the comprehensive weight of the question is calculated; the multi-dimensional feature attributes are normalized, and the normalized feature value range is uniformly mapped to the [0, 1] interval, and concatenated to form the candidate question feature vector.

[0059] S4.3: Construct the state space representation of the reinforcement learning agent, generate the state vector, and define the action space of the reinforcement learning agent as the operation of selecting a question from the candidate question set; It should be noted that the state space representation consists of three parts: the first part is the row vector corresponding to the current student in the student knowledge state matrix, the second part is the smooth mastery change rate vector, and the third part is the knowledge point coverage vector of the currently selected question; when the k-th knowledge point has been covered by the selected question, the corresponding element has a value of 1, otherwise it is 0; the above three parts are concatenated to generate the state vector; the state vector represents the environmental state of the reinforcement learning agent at time t.

[0060] Specifically, the size of the action space is equal to the number of questions in the candidate question set; for the i-th question in the candidate question set, an action is defined to select the question to be added to the current assessment path; to avoid repeatedly selecting already selected questions, an action mask vector is maintained; when the i-th question has been selected, the i-th element in the action mask vector is set to 0, otherwise it is set to 1, where 0 indicates that the action is unavailable and 1 indicates that the action is available.

[0061] S4.4: Design the reward function for the reinforcement learning agent; Furthermore, the reward function comprehensively considers the benefits of four dimensions: the first dimension is the benefit of knowledge point coverage, the second dimension is the benefit of difficulty matching, the third dimension is the benefit of question quality, and the fourth dimension is the benefit of path smoothness.

[0062] Preferably, the specific formula for the reward function is as follows: ; in, Let be the total reward value at time t. Let be the weight coefficient of the i-th dimension, and let the four dimensions correspond to knowledge point coverage, difficulty matching, question quality, and path smoothness, respectively. For the i-th dimension in the state Execute action The subsequent immediate return function, Let be the variance penalty coefficient for the i-th dimension. For the historical average return of the i-th dimension, For the number of newly covered knowledge points, the gain weight of the jth new covered knowledge point, the coverage increment of the jth knowledge point, the quality adjustment parameter of the jth knowledge point, the comprehensive quality score of the jth knowledge point corresponding question, the number of penalty terms in the path, the penalty intensity of the mth penalty term, the mth penalty condition the penalty function.

[0063] S4.5: Construct a deep Q network as a value function approximator of the reinforcement learning agent, wherein the deep Q network comprises an input layer, multiple layers of fully connected hidden layers, and an output layer; Further, the input layer receives the state vector, the first hidden layer has 512 neurons, and the activation function uses ReLU; the second hidden layer has 256 neurons, and the activation function uses ReLU; the third hidden layer has 128 neurons, and the activation function uses ReLU; the output layer has the same number of neurons as the size of the action space, and each output neuron corresponds to a Q value estimate of a candidate question; a target network is also constructed to stabilize the training process.

[0064] In an optional embodiment, an experience replay buffer is created to store the interaction experience of the reinforcement learning agent, wherein the capacity of the experience replay buffer is set to 10,000 experience samples; each experience sample is represented as a five-tuple; after the question is selected, the immediate reward is calculated according to the reward function, the knowledge point coverage vector is updated to generate the next state vector, and the experience sample is stored in the experience replay buffer; when the number of selected questions reaches the upper limit of the number of questions or the knowledge points in the priority coverage knowledge point set are all covered, the termination flag is set to 1 to indicate that the selection process is terminated.

[0065] S4.6: Use an exploration strategy to select questions, set the initial value of the exploration rate and gradually decay it with the training progress; Specifically, at time t, a random number in the interval [0, 1] is generated, and when the random number is less than the exploration rate, a random action is selected from the available actions; when the random number is greater than or equal to the exploration rate, the current state vector is input into the deep Q network to obtain the Q value estimation vector of all candidate questions, the Q value corresponding to the position with a value of 0 in the action mask vector is set to negative infinity, and the action with the maximum Q value is selected. The action is executed to add the corresponding question to the current assessment path, and the action mask vector and the knowledge point coverage vector are updated.

[0066] S4.7: When the number of experience samples stored in the experience replay buffer exceeds the batch size, experience samples are randomly sampled from the experience replay buffer; Further, for each experience sample in the batch, the target network is used to calculate the target Q value, the deep Q network is used to calculate the current Q value prediction, the time difference error is calculated, and the mean square error loss function is constructed; the optimizer is used to update the network parameters of the deep Q network by gradient descent, and the parameters of the deep Q network are periodically copied to the target network to maintain the stability of the training.

[0067] S4.8: repeatedly perform the question selection process, select questions to join the evaluation path, update the state, until the termination condition is met, and generate a question selection sequence; Further, the question identifier and the selection time of each selection are recorded, and the question selection sequence is generated in the order of the selection time; the question selection sequence is post-processed, the knowledge point dependency relationship in the knowledge graph annotation data is checked, and whether the pre-knowledge point is not covered between adjacent questions is checked, if the dependency relationship is detected to be missing, the question covering the missing pre-knowledge point is supplemented in the candidate question set and inserted into the corresponding position, and the final question selection sequence is generated.

[0068] It should be noted that the termination condition: the termination flag is 1; the question selection sequence contains a plurality of question identifiers.

[0069] S4.9: based on the question selection sequence, extract the complete content of the question from the question bank, and construct an adaptive evaluation path; It should be noted that the complete content includes the stem, the option or the answer requirement, the reference answer and the scoring standard; the adaptive evaluation path includes the question presentation order, the expected answer time of each question and the jump logic between questions.

[0070] Preferably, a dynamic feedback mechanism is configured for each question in the adaptive evaluation path, and after the student submits the answer, whether a supplementary question for a weak knowledge point needs to be added in the subsequent path is determined according to the evaluation result of the multi-agent collaborative evaluation model; the path generation timestamp, the target knowledge point coverage list and the expected total evaluation time are recorded in the metadata of the adaptive evaluation path.

[0071] S4.10: after the student completes all the questions in the adaptive evaluation path, complete evaluation data is collected, and a knowledge point answer efficiency vector is generated; Specifically, the correct answer rate of each knowledge point in the priority coverage knowledge point set is counted, the ratio of the average answer time of each knowledge point to the standard answer time is calculated, and a knowledge point answer efficiency vector is generated.

[0072] S4.11: based on the final state of the student knowledge state matrix, according to the knowledge point hierarchical relationship in the knowledge graph annotation data, the knowledge point nodes are aggregated into a plurality of ability dimensions; Further, for each ability dimension, the mastery probability values of all knowledge point nodes contained in the dimension in the student knowledge state matrix are extracted, a weighted average ability score is calculated, the weighted average ability score is mapped to a percentage score, and a quantitative evaluation score of each ability dimension is generated.

[0073] It should be noted that the evaluation data includes the answer content, the answer duration, the multi-dimensional score results and the final updated student knowledge state matrix of each question; the ability dimensions correspond to the core ability requirements of the course such as understanding ability, application ability and comprehensive analysis ability.

[0074] S4.12: From the student knowledge state matrix, knowledge point nodes with a mastery probability value below the mastery threshold are screened, and in combination with the smooth mastery degree change rate vector, knowledge point nodes with a low mastery probability value and a negative mastery degree change rate are identified, and these knowledge point nodes are marked as key weak links; Further, according to the knowledge point answer efficiency vector, knowledge point nodes with a larger answer efficiency ratio are marked as efficiency weak links, and based on the coverage score and the matching score in the multi-dimensional score results, the average score of each knowledge point is calculated, when the average coverage score is low, it is marked as insufficient coverage, and when the average matching score is low, it is marked as understanding deviation, and a weak link diagnosis list is generated.

[0075] S4.13: For each weak knowledge point in the weak link diagnosis list, corresponding learning resource recommendations are retrieved from the knowledge resource library, and personalized learning suggestion texts are generated; Specifically, according to the setback risk node list in the knowledge point state distribution statistical table, key review suggestions are generated; based on the mastery degree change rate vector, knowledge point nodes with a positive mastery degree change rate and a larger value are identified, an advantage knowledge point list and an ability maintenance suggestion are generated; learning resource recommendations, key review suggestions and ability maintenance suggestions are summarized, and personalized learning suggestion texts are generated.

[0076] Further, an ability diagnosis report document is created, the ability diagnosis report document is exported in PDF format, and the report data is stored in JSON format to the student archive database.

[0077] It should be noted that the learning resource recommendation includes a teaching video link, an exercise set identifier and a knowledge point explanation document; the intensive review suggestion clearly indicates the knowledge points that need to be strengthened and the recommended review order; the ability diagnosis report document contains five main parts: the first part is a test overview, recording the test time, the total number of test questions, the total test duration and the overall completion quality rating; the second part is the ability dimension evaluation, which displays the quantitative evaluation score of each ability dimension and the radar chart visualization result; the third part is the knowledge point mastery analysis, which lists the current mastery probability value and the smooth mastery degree change rate of each knowledge point in the student knowledge state matrix, and uses a heat map to display the knowledge point mastery state distribution; the fourth part is the weak link diagnosis, which details each weak link in the weak link diagnosis list and its diagnosis basis; and the fifth part is the learning suggestion, which presents the personalized learning suggestion text and the subsequent learning path planning.

[0078] In summary, the present application constructs a student knowledge state matrix by fusing historical test data and knowledge graph annotation data, and couples a Bayesian knowledge tracking algorithm with a forgetting decay coefficient, thereby realizing precise quantitative modeling and dynamic tracking of the student's knowledge mastery state, and avoiding the blindness of traditional fixed item bank recommendation; a multi-Agent collaborative evaluation model is used in combination with a knowledge point attention mechanism to break through the limitations of traditional single score evaluation, generate a multi-dimensional score result including an overall score and a subdivided index, and not only determine right or wrong, but also identify complex situations such as partial correctness and concept confusion; the knowledge point mastery probability is updated in real time based on the multi-dimensional score result, the mastery degree change rate is calculated, and an adaptive difficulty adjustment function is constructed, thereby realizing intelligent feedback and difficulty adjustment of the test process, ensuring that the difficulty of the questions is always in the student's zone of proximal development, and discovering knowledge system faults through preposition dependency analysis, so that the test process itself has a learning promotion effect; the difficulty parameter and the knowledge point coverage range are used as constraint conditions to screen a candidate question set, and a reinforcement learning Agent is used for global optimization sorting, not only optimizing the question difficulty, but also taking into account the completeness of knowledge point coverage and the logical correlation before and after, to generate an adaptive test path and an ability diagnosis report; compared with traditional standardized testing, the number of invalid questions is significantly reduced, the test efficiency and diagnosis accuracy are greatly improved, and a paradigm shift from selection evaluation to development evaluation is realized.

[0079] Under the guidance of the above embodiments, another aspect of the embodiments of the present application also proposes an Agent-based education test system, comprising: a knowledge modeling module, configured to acquire historical test data and knowledge graph annotation data of a student, construct a student knowledge state matrix, and calculate a mastery probability value and a forgetting decay coefficient of each knowledge point by using a Bayesian knowledge tracking algorithm, to generate a test question recommendation sequence; The cooperative evaluation module is configured to input the question texts in the recommended sequence of the evaluation questions and the student answer texts into a multi-agent cooperative evaluation model to generate multi-dimensional scoring results. The dynamic updating module is configured to dynamically update the knowledge point mastery probability values of the student in a continuous answering process based on the multi-dimensional scoring results, calculate the mastery degree change rates of the knowledge point nodes, construct an adaptive difficulty adjustment function, and generate question difficulty parameters and knowledge point coverage range parameters for the next round of evaluation. The path generation module is configured to select a candidate question set meeting the conditions from a question bank by taking the question difficulty parameters and the knowledge point coverage range parameters as constraint conditions, and perform optimal selection and sorting on the candidate question set by using a reinforcement learning agent according to the real-time state of the student knowledge state matrix to generate an adaptive evaluation path and an ability diagnosis report.

[0080] The embodiment also provides a computer device suitable for the agent-based educational evaluation method, including a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the agent-based educational evaluation method proposed in the above embodiment.

[0081] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is configured to perform wired or wireless communication with external terminals. The wireless communication can be achieved through WIFI, operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. The input device of the computer device can also be an external keyboard, touchpad or mouse, etc.

[0082] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to implement the agent-based educational evaluation method proposed in the above embodiment.

[0083] The storage medium proposed in the embodiment belongs to the same inventive concept as the data storage method proposed in the above embodiment. The technical details not described in detail in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0084] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An agent-based educational assessment method, characterized in that: include, Acquire students' historical assessment data and knowledge graph annotation data, construct a student knowledge state matrix, and use the Bayesian knowledge tracing algorithm to calculate the mastery probability value and forgetting decay coefficient of each knowledge point, and generate a recommended sequence of assessment questions; Input the question texts and student answer texts from the recommended sequence of assessment questions into the multi-Agent collaborative evaluation model to generate multi-dimensional scoring results; Based on the multi-dimensional scoring results, the probability values ​​of students' mastery of knowledge points during continuous answering are dynamically updated, the rate of change of mastery of each knowledge point node is calculated, an adaptive difficulty adjustment function is constructed, and the question difficulty parameters and knowledge point coverage parameters for the next round of assessment are generated. Using the question difficulty parameter and knowledge point coverage parameter as constraints, a set of candidate questions that meet the conditions is selected from the question bank. Then, a reinforcement learning agent is used to optimally select and sort the candidate question set according to the real-time status of the student's knowledge state matrix, generating an adaptive assessment path and ability diagnosis report.

2. The agent-based educational assessment method as described in claim 1, characterized in that: The method for generating the question difficulty parameters and knowledge point coverage parameters for the next round of assessment is as follows: Extract the overall rating level and detailed rating indicators from the multi-dimensional scoring results, and associate the overall rating level with the knowledge point annotations of the corresponding questions in the recommended sequence of assessment questions to generate a sequence of answer result labels. Based on the answer result label sequence, extract the current mastery probability value of the corresponding knowledge point node in the student knowledge state matrix; Obtain students' continuous answer records within the current assessment period, apply Bayesian posterior probability update operations sequentially according to time order, record the change in mastery probability value before and after each update, and generate a knowledge point mastery probability tracking sequence. Traverse all knowledge point nodes in the student knowledge state matrix, extract the initial mastery probability value and the final mastery probability value of the knowledge point from the mastery probability tracking sequence, and obtain the mastery change rate. Based on the smooth mastery rate of change vector, the state of each knowledge point node in the knowledge graph annotation data is classified, and a knowledge point state distribution statistics table is generated. Based on the knowledge point status distribution statistics table and the student knowledge status matrix, an adaptive difficulty adjustment function is constructed to generate a difficulty adjustment coefficient vector.

3. The agent-based educational assessment method as described in claim 2, characterized in that: It also includes, The difficulty adjustment coefficient vector is mapped to a predefined question difficulty level system in the question bank to generate a question difficulty parameter sequence. Extract the list of regression risk nodes and the set of weak knowledge points from the knowledge point status distribution statistics table, merge the two lists and remove duplicate knowledge point node identifiers to generate a set of priority coverage knowledge points. Based on the knowledge point dependencies in the knowledge graph annotation data, a pre-dependency analysis is performed on each knowledge point node in the priority coverage knowledge point set. If the mastery probability value of a current dependent knowledge point is lower than the mastery threshold, the current dependent knowledge point is added to the set of priority coverage knowledge points, and the hierarchical depth weight of each knowledge point node is calculated according to the hierarchical relationship of knowledge points to generate knowledge point coverage parameters.

4. The agent-based educational assessment method as described in claim 2, characterized in that: The method for generating the multi-dimensional scoring results is as follows: The student's answer text word sequence is input into the multi-Agent collaborative evaluation model. The attention score between the student's answer text word sequence and the knowledge point embedding vector is calculated through the knowledge point attention sub-layer, and the key knowledge point feature vector is extracted. A multi-head attention mechanism is used to calculate the semantic similarity matrix between the feature vectors of the key knowledge points and the feature vectors of the standard answers; Based on the semantic similarity matrix, a slicing operation is performed according to the knowledge point dimension. According to the knowledge point node identifiers associated with the question text, a sub-matrix corresponding to the knowledge point region is extracted from the semantic similarity matrix. The row mean and column mean of the submatrix are calculated, and the mean of the diagonal elements of the semantic similarity matrix is ​​calculated as the overall similarity score to generate a multi-dimensional scoring feature vector. The multi-dimensional rating feature vector is input into the rating mapping network, which includes two fully connected layers and an output layer. The output layer generates a five-dimensional rating probability distribution, and the level with the highest probability in the rating probability distribution is used as the overall rating level. At the same time, the scores of each item in the multi-dimensional rating feature vector are retained as subdivided rating indicators to generate a multi-dimensional rating result.

5. The agent-based educational assessment method as described in claim 4, characterized in that: The multi-agent collaborative evaluation model is constructed based on an improved BERT semantic understanding network. The improved BERT semantic understanding network adds a knowledge point perception layer to the standard BERT model, and adds a knowledge point attention sub-layer after the self-attention layer of the BERT encoder.

6. The agent-based educational assessment method as described in claim 3, characterized in that: The method for obtaining the mastery probability value is as follows: Traverse the row vectors corresponding to the current student in the student knowledge state matrix, identify and judge the initial mastery rate of each knowledge point node, and obtain the set of weak knowledge points; Based on the set of weak knowledge points, the historical answer sequence corresponding to the knowledge point node is extracted from the time-series assessment record sequence, and the mastery probability value is updated sequentially in chronological order. If the t-th answer is correct, the mastery probability value at time t+1 is calculated using the Bayesian update formula; if the t-th answer is incorrect, the mastery probability value is adjusted accordingly. By traversing the complete historical answer sequence, the probability of mastering all weak knowledge points is dynamically calculated and updated.

7. The agent-based educational assessment method as described in claim 6, characterized in that: The method for generating the recommended sequence of assessment questions is as follows: Based on the time-series evaluation record sequence, the forgetting decay function is constructed using the Ebbinghaus forgetting curve model to obtain the forgetting decay coefficient of knowledge point nodes; The overall weakness score is obtained by multiplying the mastery probability value of each knowledge point node in the set of weak knowledge points with the corresponding forgetting decay coefficient. The system performs topological sorting and weighted sorting based on the knowledge point dependencies to generate a knowledge point priority sequence. Based on this priority sequence, it retrieves questions from the question bank that are associated with the top-ranked knowledge point nodes, thus generating a recommended sequence of assessment questions.

8. An agent-based educational assessment system, based on the agent-based educational assessment method according to any one of claims 1 to 7, characterized in that: include, The knowledge modeling module is used to acquire students' historical assessment data and knowledge graph annotation data, construct students' knowledge state matrix, and use the Bayesian knowledge tracing algorithm to calculate the mastery probability value and forgetting decay coefficient of each knowledge point, and generate a recommended sequence of assessment questions. The collaborative evaluation module is used to input the question text and student answer text from the recommended sequence of evaluation questions into the multi-Agent collaborative evaluation model to generate multi-dimensional scoring results. The dynamic update module, based on the multi-dimensional scoring results, dynamically updates the probability values ​​of students' mastery of knowledge points during continuous answering, calculates the rate of change of mastery of each knowledge point node, constructs an adaptive difficulty adjustment function, and generates the question difficulty parameters and knowledge point coverage parameters for the next round of assessment. The path generation module is used to select a set of candidate questions that meet the conditions from the question bank, using the question difficulty parameter and knowledge point coverage parameter as constraints. Then, it uses a reinforcement learning agent to perform optimal selection and sorting of the candidate question set according to the real-time status of the student's knowledge state matrix, and generates an adaptive assessment path and ability diagnosis report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the agent-based educational assessment method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the agent-based educational assessment method according to any one of claims 1 to 7.

Images