Intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph

Abstract

The invention discloses an intelligent question bank retrieval and recommendation system based on an artificial intelligence knowledge graph, and relates to the technical field of artificial intelligence education. Comprising a knowledge graph construction module which is used for processing original education data and constructing a neighborhood knowledge graph comprising a hard preposition relation and a soft incidence relation; the user knowledge state graph construction module is used for constructing a user personal knowledge state graph isomorphic to the domain knowledge graph, and dynamically calculating a mastery index of each knowledge node through a deep knowledge tracking model based on user historical answer data; according to the method, by constructing the domain knowledge graph containing the hard preposition relation and the soft incidence relation, discrete knowledge points are organized into the structured network conforming to the cognitive law, so that the system can understand and follow the internal logic between knowledge, and a learning path which is clear in organization, efficient and coherent is generated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education technology, specifically to an intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graphs. Background Technology

[0002] With the popularization of online education, massive question banks, while bringing abundant resources, also present two core challenges: accurate retrieval and personalized recommendations. Existing systems largely rely on keyword matching and simple tags, failing to understand users' deeper semantic intent or capture the complex relationships between knowledge points, resulting in unsatisfactory search results. In terms of recommendations, they generally rely on superficial indicators such as answer accuracy, lacking fine-grained modeling and dynamic tracking of users' knowledge states, making it difficult to achieve truly personalized learning path planning and deep, fine-grained dynamic modeling of knowledge states.

[0003] For example, patent publication number CN114117252A discloses a method and system for recommending intelligent, personalized question banks for students. Specifically, it includes: creating a target student profile based on the basic information of the registered target student and obtaining the target student's incorrect questions; establishing an intelligent recommendation model and associating the target student profile with the intelligent recommendation model; wherein the intelligent recommendation model includes calculating the similarity of the target student's incorrect questions using a collaborative filtering algorithm, and combining a clustering algorithm to categorize the incorrect questions based on the knowledge points of the questions to obtain a target knowledge base; generating a recommended question bank based on the target student profile and the target knowledge base; and forming a set of incorrect questions for the target student in the recommended question bank. This invention precisely helps students strengthen their weak knowledge points according to their actual situation, thereby quickly improving their academic performance.

[0004] However, the above and similar technical solutions still have the following shortcomings: treating knowledge points as independent labels and only performing rough classification through clustering lacks structured modeling of the internal logic and dependencies of knowledge, resulting in the system's inability to understand the knowledge structure and the chaotic and inefficient recommended learning paths; relying on static basic information of users and collaborative filtering, it fails to establish a dynamically evolving knowledge state model, causing recommendations to remain at the level of superficial similarity and unable to achieve accurate root cause diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graphs to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graphs, comprising: The knowledge graph construction module is used to process raw educational materials and construct a neighborhood knowledge graph that includes hard preconditions and soft associations; and to construct a user's personal knowledge state graph that is isomorphic to the domain knowledge graph, and dynamically calculate the mastery index of each knowledge node based on the user's historical answer data through a deep knowledge tracking model. Intelligent semantic retrieval module: It is used to perform semantic parsing and expansion of the user's natural language questions through a pre-trained language model, map them to the corresponding knowledge nodes in the neighborhood knowledge graph, traverse the question nodes that have semantic association with the knowledge node, and return the retrieval results; Diagnosis and Recommendation Module: In response to the user's answer results, it combines the user's personal knowledge state graph and performs source tracing analysis in the neighborhood knowledge graph based on the dependency chain tracing model based on the mastery probability to locate the fundamental knowledge weaknesses and generate a suitable learning path and question sequence. Interaction module: used to receive natural language questions from users; and to present users with their knowledge mastery status, learning path, and progress.

[0007] Furthermore, the method for constructing the domain knowledge graph includes: The original educational materials are analyzed and cleaned to identify and extract the smallest knowledge units that can be independently examined according to the subject education standards. These units are defined as knowledge nodes, and a unique identifier and standardized name are generated for each knowledge node. Each question in the question bank is parsed into structured data and defined as a question node; each question node is labeled with the core knowledge node it directly tests, and a connection is established with the corresponding knowledge node through examination relationship edges; Based on the official teaching syllabus and textbook catalog, extract the explicit sequence of knowledge nodes; analyze the knowledge definitions and problem-solving principles to determine the prerequisite knowledge points that must be relied upon to master specific knowledge points; and unify and solidify the above dependencies into hard prerequisite relationship edges to form a directed hard prerequisite relationship network. Based on the co-occurrence frequency of different knowledge nodes in the same question, the implicit connections mined from users' historical answer data, and the semantic similarity between knowledge nodes, a weighted fusion is used to assign association strength weights to knowledge node pairs, and soft association edges are established for node pairs that exceed a preset threshold.

[0008] Furthermore, the relationships in the domain knowledge graph also include confusion relationships, which are used to identify knowledge node pairs that are likely to generate cognitive errors in association; these relationships are established based on the association analysis of user groups' incorrect response behavior on corresponding questions.

[0009] Furthermore, the method of dynamically calculating the mastery index of each knowledge node based on the user's historical answer data through a deep knowledge tracing model specifically includes: By fusing user historical answer data, question metadata obtained from domain knowledge graphs, and user behavior context data, a temporal composite interaction sequence is constructed as the input source for the deep knowledge tracing model. Based on the topological relationships in the neighborhood knowledge graph, a knowledge node semantic vector is generated for each knowledge node using graph embedding technology. Then, the semantic vectors of each knowledge node involved in each sequence unit in the composite interaction sequence are weighted and aggregated through an attention mechanism to form an interaction representation vector. The interaction representation vector is simultaneously input into a dual-path architecture that runs parallel to both a temporal model and a graph neural network. This architecture captures the evolution of the user's knowledge state over time and its dependencies on the domain knowledge graph structure, and then fuses them to generate an enhanced knowledge state vector. Based on the enhanced knowledge state vector, the deep knowledge tracing model synchronously outputs the dynamic mastery probability of all knowledge nodes and the uncertainty quantification of each mastery estimate; By introducing graph logic constraints, we ensure that the prediction results conform to cognitive laws; and we update the output results to the user's personal knowledge state graph in real time.

[0010] Furthermore, the intelligent semantic retrieval module specifically includes: The general pre-trained language model is continuously trained using corpus in this field, and the model is tuned based on constructed paired samples; The user's natural language question is input into the optimized pre-trained language model to generate a question semantic vector. In the same semantic vector space, the similarity between the question semantic vector and the semantic vectors of all knowledge nodes is calculated, and the core knowledge node corresponding to the user's question intent is located from the domain knowledge graph. Starting with the core knowledge node, traverse along the soft association edges in the domain knowledge graph to collect all related knowledge nodes and form a weighted set of query knowledge nodes. The system iterates through the knowledge node set, collects related question nodes by examining relational edges, and forms a candidate question set. It then integrates semantic matching degree, knowledge point association strength, and question quality to perform multi-dimensional weighted sorting and filtering of the candidate questions. The sorted question nodes and their metadata are organized into a structured result, and explanatory information based on core and related knowledge points is added. Finally, the search results are returned to the user.

[0011] Furthermore, the dependency chain tracing model based on mastery probability specifically includes: When it is detected that the user's recent answer accuracy rate on the current knowledge node is lower than the first preset threshold, the source analysis of the node is triggered; the current node is used as the starting point of the source chain and marked as the initial question knowledge node; Based on the fusion analysis of the instantaneous state of mastery of each node in the user's personal knowledge state graph and its temporal change trend, the depth and direction of the tracing path are dynamically determined in the hard pre-relationship network of the domain knowledge graph, and the fundamental knowledge weakness of this tracing is marked. After the source analysis of all paths is terminated, all knowledge nodes marked as fundamental knowledge weaknesses are collected to form the final set of fundamental knowledge weaknesses; at the same time, all dependency chains from the initial problem knowledge node to each fundamental knowledge weakness are recorded.

[0012] Furthermore, the fusion analysis includes determining whether the parent node simultaneously satisfies the condition that its mastery level is lower than a second preset threshold and that its mastery level change trend satisfies a decreasing condition; the decreasing condition is calculated by determining whether the linear regression slope of the mastery level probability within the sliding window is positive or negative.

[0013] Furthermore, the generation of the adapted learning path and question sequence specifically includes: The initial problem knowledge node that triggered this source tracing analysis is set as the short-term learning objective, and constraints are defined. Extract all hard prerequisite relationship paths connecting the set of fundamental knowledge weaknesses and the target knowledge nodes from the domain knowledge graph to form a target dependency path set; For each path in the target dependency path set, calculate its path comprehensive weakness, and sort all paths to obtain the priority learning path set; From all paths in the priority learning path set, extract the starting knowledge node of the learning path; calculate the weighted probability of the user's mastery of the starting knowledge node with its centrality in the domain knowledge graph to obtain the knowledge repair priority score, and sort them from low to high to form the first learning path; For each knowledge node in the first learning path, an adaptive question sequence is dynamically generated; the adaptive question sequence includes diagnostic questions and reinforcement questions.

[0014] Furthermore, the constraints include knowledge constraints, mastery constraints, and cognitive load constraints; the dynamically generated adaptive question sequence introduces a question exposure control mechanism when recommending questions to prevent the repeated recommendation of similar questions to users.

[0015] Furthermore, the diagnostic recommendation module is also configured to dynamically switch between different recommendation strategy modes based on the user's current learning stage and set learning goals; the recommendation strategy modes include remedial mode, advanced mode, and leap mode.

[0016] Compared with the prior art, the beneficial effects of the present invention are: An intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graphs upgrades the system from mechanical keyword matching to intelligent understanding of user intent and semantic associations through semantic retrieval and expansion based on domain knowledge graphs. This effectively solves the problems of biased understanding of retrieval intent and one-sided results. Through dynamic switching of multi-mode recommendation strategies, the system can adapt to users with different learning stages and goals, achieving full coverage from remedial to advanced learning, making teaching intervention more precise and intelligent.

[0017] Meanwhile, by constructing a domain knowledge graph containing hard prerequisite relationships and soft association relationships, discrete knowledge points are organized into a structured network that conforms to cognitive laws, enabling the system to understand and follow the inherent logic between knowledge points, thereby generating a clear, efficient and coherent learning path. By constructing a user's personal knowledge state graph that is isomorphic to the domain knowledge graph, and using a deep knowledge tracking model to dynamically calculate the mastery of each node, a knowledge state model that can evolve in real time is established, enabling the system to accurately locate the user's fundamental knowledge weaknesses and achieve accurate diagnosis and question recommendation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the intelligent question bank retrieval and recommendation system of the present invention; Figure 2 This is a schematic diagram of the domain knowledge graph construction method of the present invention; Figure 3 This is a schematic diagram of the method for dynamically calculating the mastery index of each knowledge node in this invention; Figure 4 This is a schematic diagram of the intelligent semantic retrieval method of the present invention; Figure 5 This is a schematic diagram of the method for tracing and analyzing the source of knowledge to locate fundamental knowledge weaknesses in this invention. Figure 6 This is a schematic diagram of the learning path and question sequence generation method of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figure 1 As shown, the present invention provides a technical solution: an intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graphs, comprising: The knowledge graph construction module is used to process raw educational materials and construct a neighborhood knowledge graph that includes hard prerequisite relationships and soft association relationships; and to construct a user's personal knowledge state graph that is isomorphic to the domain knowledge graph, and dynamically calculate the mastery index of each knowledge node based on the user's historical answer data through a deep knowledge tracking model.

[0021] like Figure 2 As shown, this invention provides a method for constructing a domain knowledge graph; Specifically: The original educational materials are analyzed and cleaned to identify and extract the smallest knowledge units that can be independently examined according to the subject education standards. These units are defined as knowledge nodes, and a unique identifier and standardized name are generated for each knowledge node. Each question in the question bank is parsed into structured data and defined as a question node; each question node is labeled with the core knowledge node it directly tests, and a connection is established with the corresponding knowledge node through examination relationship edges; Based on the official teaching syllabus and textbook catalog, extract the explicit sequence of knowledge nodes; analyze the knowledge definitions and problem-solving principles to determine the prerequisite knowledge points that must be relied upon to master specific knowledge points; and unify and solidify the above dependencies into hard prerequisite relationship edges to form a directed hard prerequisite relationship network. Based on the co-occurrence frequency of different knowledge nodes in the same question, the implicit connections mined from users' historical answer data, and the semantic similarity between knowledge nodes, a weighted fusion is used to assign association strength weights to knowledge node pairs, and soft association edges are established for node pairs that exceed a preset threshold.

[0022] It should be noted that the original educational materials include, but are not limited to, official teaching syllabi, electronic textbooks, lesson plans, and question banks. Text is extracted from the original educational materials using a PDF parser or named entity recognition technology, and preprocessed with word segmentation and part-of-speech tagging. A language model pre-trained on the educational text (such as the RoBERTa model) is fine-tuned into a sequence labeling model (using the BIOES standard system) to automatically identify and extract the smallest knowledge units from the text. A knowledge node is created for each knowledge unit, with node attributes including knowledge node ID, canonical name, and definition.

[0023] Each question in the question bank is parsed and stored as a question node. Node attributes include question ID, question text, answer choices, standard answer, explanation text, and question metadata. Question metadata includes question difficulty, question type, question score, and ability dimension. For questions containing special formats (such as formulas or graphs), appropriate parsing and standardization processing is performed. Each question node is labeled with its directly tested core knowledge points. The implementation methods include: expert annotation, where domain experts manually associate knowledge points using a built-in annotation tool; and automatic classification, using a text classification model (such as a BERT-based text classifier) ​​to predict the knowledge nodes tested, taking the question content and explanation text as input. The training data for this classifier comes from a set of questions with expert-annotated knowledge nodes, thus constructing a supervised learning model from question text to knowledge node IDs. In the domain knowledge graph, examination relationship edges are created from question nodes to knowledge nodes. These edges can be attached with examination intensity attributes, such as core examination or comprehensive examination.

[0024] The official course syllabus and textbook catalog are analyzed, and their explicit sequential order is transformed into directed connections. A hybrid approach combining rule-based and graph reasoning is employed to analyze knowledge definitions and problem-solving principles: Logical rule mining extracts the following key patterns from the definitions of knowledge nodes and problem-solving paradigms: definition dependency: if the definition of node A explicitly mentions node B, a hard precondition relationship from B to A is established; formula / theorem dependency: if the derivation of a formula or proof of a theorem for node A requires the conclusion of node B as a premise, a hard precondition relationship from B to A is established. This is achieved by analyzing keywords in the derivation steps (such as "substitution," "according to...theorem," "from...it can be obtained"). The extracted relationships are compared, verified, and supplemented with a pre-defined core hard precondition relationship rule base defined by domain experts. For relationships that cannot be determined by the above methods or that conflict, the expert rules prevail. As a supplement to the expert knowledge base, the system can employ a graph embedding-based reasoning method. Based on the relationships constructed above, graph embedding algorithms (such as TransE) learn node vectors and predict and complete missing hard prerequisite relationships through vector operations. Before solidifying the directed connections, the constructed hard prerequisite relationship network is validated to ensure its acyclicity, in accordance with the irreversibility of knowledge accumulation. Finally, directed hard prerequisite relationship edges are established from prerequisite knowledge points to target knowledge points.

[0025] The weighted fusion specifically includes: co-occurrence frequency, which counts the number of times two knowledge nodes appear together in all questions and calculates their standardized score, such as using the Jaccard coefficient; implicit connection, which uses association rule mining algorithms (such as Apriori) to calculate rules from the historical answer data of user groups, such as "users who have mastered knowledge point a have a high probability of also mastering knowledge point b", and calculates the lift of the rule as the implicit connection strength score; and semantic similarity, which uses the name and definition text of the knowledge nodes to generate text vectors through a pre-trained semantic model (such as Sentence-BERT) and calculates the cosine similarity as the semantic similarity score. Finally, the comprehensive association strength weight of the knowledge node pair can be calculated by linear weighted combination, such as comprehensive association strength weight = 0.5 × co-occurrence frequency score + 0.3 × implicit connection strength score + 0.2 × semantic similarity score. The 85th percentile of the comprehensive association strength weight value of all node pairs is selected as the threshold, and undirected soft association edges are established for node pairs exceeding the threshold, and the weights are stored.

[0026] Optionally, the relationships in the domain knowledge graph also include confusion relationships, used to identify knowledge node pairs that are cognitively prone to association errors. The establishment method includes: All records of incorrect answers were filtered out. For each incorrect question, the knowledge point tested (the knowledge point corresponding to the correct answer, denoted as A) and the related knowledge point (the knowledge point incorrectly applied by the student, derived from the analysis of the incorrect answer, denoted as B) were recorded. The criteria for determining knowledge point B included: the incorrect formula or concept name explicitly written by the student in the incorrect answer; the teacher's classification and annotation of typical error patterns; and automatic inference by analyzing the semantic similarity between the incorrect answer text and the knowledge node definition.

[0027] Count the frequency of occurrence of all knowledge point pairs (A, B) in the above error record. A,B For each knowledge point pair (A, B), construct a 2×2 contingency table containing O A,B And their marginal frequencies (the sum of rows and columns in the contingency table). Calculate their chi-square value and determine whether it rejects the null hypothesis that the confusion between A and B is random at a significance level (e.g., α = 0.01). For a knowledge point pair that passes the significance test, its confusion strength is defined as: O A,B The total number of incorrect answers related to knowledge point A. Establish an undirected confusion relationship edge between knowledge points A and B, and store the confusion strength attribute. The domain knowledge graph is stored in a graph database (such as Neo4j).

[0028] The system creates a personal knowledge state graph for each user. This graph is isomorphic to the domain knowledge graph, meaning it replicates all knowledge nodes, question nodes, and all relationships between them (hard prerequisites, soft associations, and examination relationships). The key difference is that each knowledge node in the user's personal knowledge state graph is appended with the user's personal, real-time updated mastery index; this graph does not store global, inherent question metadata, but it does store records of user interactions with questions.

[0029] like Figure 3 As shown, this invention provides a method for dynamically calculating the mastery index of each knowledge node; Specifically: By fusing user historical answer data, question metadata obtained from domain knowledge graphs, and user behavior context data, a temporal composite interaction sequence is constructed as the input source for the deep knowledge tracing model. Based on the topological relationships in the neighborhood knowledge graph, a knowledge node semantic vector is generated for each knowledge node using graph embedding technology. Then, the semantic vectors of each knowledge node involved in each sequence unit in the composite interaction sequence are weighted and aggregated through an attention mechanism to form an interaction representation vector. The interaction representation vector is simultaneously input into a dual-path architecture that runs parallel to both a temporal model and a graph neural network. This architecture captures the evolution of the user's knowledge state over time and its dependencies on the domain knowledge graph structure, and then fuses them to generate an enhanced knowledge state vector. Based on the enhanced knowledge state vector, the deep knowledge tracing model synchronously outputs the dynamic mastery probability of all knowledge nodes and the uncertainty quantification of each mastery estimate; By introducing graph logic constraints, we ensure that the prediction results conform to cognitive laws; and we update the output results to the user's personal knowledge state graph in real time.

[0030] It is important to note that the user's historical answer data is extracted from system logs and includes, but is not limited to, user ID, question ID, answer result (binarized, 1 for correct, 0 for incorrect), and timestamp; question metadata is obtained by associating with question nodes in the domain knowledge graph; user behavior context data includes, but is not limited to, normalized solution time, number of attempts (retrying the same question), and interaction type (e.g., first time learning, review, quiz). All user answer records are sorted in ascending order according to timestamps to form the original interaction sequence. For the t-th interaction in the interaction sequence, a composite interaction unit is constructed. This unit contains a question-knowledge node association vector, answer result, normalized difficulty coefficient, and behavior context feature vector. Finally, a temporally sequenced composite interaction sequence is formed as the input source of the model. The question-knowledge node association vector is a multi-hot encoded vector with a dimension equal to the total number of knowledge nodes, and the position of the examined node is 1.

[0031] Training is performed on a constructed domain knowledge graph using graph embedding techniques (such as TransR). Training follows a standard knowledge graph embedding process, employing maximum margin loss and negative sampling to generate a low-dimensional semantic vector containing graph structure information for each knowledge node. For each interaction, the semantic vectors of all knowledge nodes tested in the question are obtained. An attention mechanism is used to assign weights to these semantic vectors. The attention score depends on the vector itself and the model's hidden state at the previous time step (initially a zero vector), implemented through a single-layer feedforward network and a Softmax function. The weighted semantic vectors are summed to obtain the interaction representation vector for this interaction. This vector comprehensively reflects the semantics and relative importance of the knowledge points involved in this test.

[0032] The dual-path architecture generates enhanced knowledge state vectors. Specific implementation methods include: a temporal path, employing a Long Short-Term Memory (LSTM) network as the temporal model, sequentially inputting the serialized interaction representation vectors into the LSTM to capture the long-term dependencies and evolution patterns of the user's knowledge state over time, outputting the temporal hidden state; and a structural path, using a Graph Attention Network (GAT) as the graph neural network. The GAT uses the relatively stable topological relationships of the domain knowledge graph as a fixed adjacency matrix. These topological relationships include hard preconditions and soft associations, and are considered static structures within a model training or inference cycle, but the system periodically updates and optimizes them offline based on newly collected group data. At each time step t, the initial features of each knowledge node in the GAT are formed by concatenating the user's mastery probability of the node at the previous time step (t-1) with the node's static semantic vector, and then fusing them through a multilayer perceptron. The GAT propagates information on this initial graph, using graph relationships to smooth and infer the user's step-by-step knowledge state, outputting the enhanced node representation. For each knowledge node, its global state in the temporal path and its personalized state in the structural path are concatenated. The concatenated vector is then input into a fusion multilayer perceptron, which outputs the enhanced knowledge state vector for that knowledge node.

[0033] The enhanced knowledge state vector of each knowledge node is input into an independent Sigmoid output layer. The output of this layer is the mastery probability of that node at the current time, a continuous value between 0 and 1, representing the estimated probability that the user has mastered that knowledge node. The uncertainty of the mastery estimate can be quantified using Monte Carlo Dropout. During model prediction, the Dropout layer is kept on, and multiple (e.g., 50) random forward propagations are performed to obtain multiple samples of the mastery probability for each knowledge node. The standard deviation of these samples is calculated as a quantification of the uncertainty of the mastery estimate. The larger the standard deviation, the more uncertain the model's prediction.

[0034] A hard precondition logical constraint term is introduced into the loss function during model training to ensure that the prediction conforms to cognitive patterns. This constraint term is calculated for each hard precondition relationship: when the model predicts that the user's mastery of a sub-knowledge point is significantly higher than (exceeding a preset tolerance interval, such as 0.1) the mastery of the parent knowledge point, a penalty is imposed. The model's total loss function is composed of a weighted sum of the standard binary cross-entropy loss and the logical constraint loss. Each time a user completes a question, the system immediately triggers the aforementioned calculation process. The calculated probability of mastery of all knowledge nodes and the uncertainty quantification index are then updated in real time to the corresponding user's personal knowledge state graph via a graph database interface.

[0035] Intelligent semantic retrieval module: It is used to perform semantic parsing and expansion of the user's natural language questions through a pre-trained language model, map them to the corresponding knowledge nodes in the neighborhood knowledge graph, traverse the question nodes that have semantic association with the knowledge node, and return the retrieval results.

[0036] like Figure 4 As shown, the present invention provides an intelligent semantic retrieval method; Specifically: The general pre-trained language model is continuously trained using corpus in this field, and the model is tuned based on constructed paired samples; The user's natural language question is input into the optimized pre-trained language model to generate a question semantic vector. In the same semantic vector space, the similarity between the question semantic vector and the semantic vectors of all knowledge nodes is calculated, and the core knowledge node corresponding to the user's question intent is located from the domain knowledge graph. Starting with the core knowledge node, traverse along the soft association edges in the domain knowledge graph to collect all related knowledge nodes and form a weighted set of query knowledge nodes. The system iterates through the knowledge node set, collects related question nodes by examining relational edges, and forms a candidate question set. It then integrates semantic matching degree, knowledge point association strength, and question quality to perform multi-dimensional weighted sorting and filtering of the candidate questions. The sorted question nodes and their metadata are organized into a structured result, and explanatory information based on core and related knowledge points is added. Finally, the search results are returned to the user.

[0037] It is important to note that a large-scale corpus of local knowledge is collected, including but not limited to textbooks, lesson plans, academic papers, high-quality question bank parsing texts, and existing knowledge node standardized names and relationships in domain knowledge graphs. Using a general pre-trained language model (such as ERNIE) as a foundation, the model is continuously trained using the aforementioned corpus on a masked language model task, enabling the model to deeply integrate domain knowledge. The paired samples are pre-constructed (question, core knowledge point) paired datasets by domain experts. A dual-tower encoder structure is adopted, using multi-negative sample ranking loss for contrastive learning. The fine-tuned model is then used as a unified domain semantic encoder.

[0038] The user's question is input into the optimized domain semantic encoder to generate an initial question semantic vector. A dedicated projection network (a two-layer MLP) projects this initial question semantic vector from the original semantic space of the pre-trained language model to the same vector space as the neighborhood knowledge graph, resulting in an aligned question semantic vector. This projection network is trained under supervision on the previously constructed (question, core knowledge point) pairing dataset. Its training objective is to minimize the mean squared error between the projected question semantic vector and the TransR semantic vector of the core knowledge node paired with the question. The correct answer is the core knowledge point that the question most intends to examine, pre-annotated by neighborhood experts. In the unified vector space, the similarity (e.g., cosine similarity) between the aligned question semantic vector and the TransR semantic vectors of all knowledge nodes in the neighborhood knowledge graph is calculated. The knowledge node with the highest similarity is selected as the core knowledge node for this query. To ensure retrieval quality, a similarity threshold (e.g., θ=0.7) is set for quality filtering. If the highest similarity is below this threshold, the user's question intent is deemed ambiguous or beyond the current knowledge scope, and the system will return a prompt requesting clarification from the user.

[0039] Starting with the core knowledge point, perform a two-step breadth-first traversal along the soft association edges to collect related knowledge points. Assign a weight of 1.0 to the core knowledge point, and the weight of the related node is the maximum product of the weights of the soft associations on the path.

[0040] In the domain knowledge graph, all question nodes associated with the set of queried knowledge nodes are collected by examining relational edges. A comprehensive scoring formula is used: S = ɑ × S 语义 +β×S 知识点 +γ×S 质量 , where: S 语义 It is the cosine similarity between the question semantic vector and the question vector. The question vector is generated by concatenating the question stem and the parsed text and then passing it through a neighborhood semantic encoder; S 知识点 The question tests the knowledge of the maximum weight of a concept in a lookup set; S 质量The normalized quality score is calculated based on a combination of factors including question difficulty and historical accuracy; the weighting coefficients are set to α=0.5, β=0.3, and γ=0.2. The top ten question nodes are selected in descending order based on the overall score.

[0041] The system retrieves complete metadata for each question node after final sorting from the graph database. This data is then organized into a structured list (e.g., a JSON array). Each question must include the following information: question ID, complete text content of the question, standard answer and detailed explanation, question difficulty level, and semantic match score (S) between the question and the user's question. 语义 The system automatically generates an explanatory text explaining the logical basis of the search, including: "For your question, we have identified the core knowledge point: Core Knowledge Node Name," "and expanded to the following related knowledge points: The top three related knowledge node names sorted in descending order of weight," and "The following are recommended questions for you." This structured question list and explanatory information are then encapsulated and presented to the user through the system's interactive module. In the front-end interface, the explanatory information serves as guiding text, with a clear question list displayed below. Each question in the list fully displays its content, standard answer, explanation, and overall score.

[0042] Diagnosis and Recommendation Module: In response to the user's answer results, it combines the user's personal knowledge state graph and uses a dependency chain tracing model based on mastery probability to perform tracing analysis in the neighborhood knowledge graph to locate the fundamental knowledge weaknesses and generate suitable learning paths and question sequences.

[0043] like Figure 5 As shown, this invention provides a method for tracing and analyzing the source to locate fundamental knowledge gaps; Specifically: When it is detected that the user's recent answer accuracy rate on the current knowledge node is lower than the first preset threshold, the source analysis of the node is triggered; the current node is used as the starting point of the source chain and marked as the initial question knowledge node; Based on the fusion analysis of the instantaneous state of mastery of each node in the user's personal knowledge state graph and its temporal change trend, the depth and direction of the tracing path are dynamically determined in the hard pre-relationship network of the domain knowledge graph, and the fundamental knowledge weakness of this tracing is marked. After the source analysis of all paths is terminated, all knowledge nodes marked as fundamental knowledge weaknesses are collected to form the final set of fundamental knowledge weaknesses; at the same time, all dependency chains from the initial problem knowledge node to each fundamental knowledge weakness are recorded.

[0044] It is important to note that the system monitors the user's recent answer accuracy rate for each knowledge node in real time. This accuracy rate is calculated based on the user's answers to questions directly testing that knowledge node in their last five attempts. When the accuracy rate is detected to be below a first preset threshold (e.g., 60%), a source tracing analysis of that knowledge node is immediately triggered. The knowledge node of the current question is established as the starting point of this source tracing chain and marked as the initial question knowledge node. A source tracing queue is created, employing a first-in, first-out (FIFO) strategy, and the initial question knowledge node is added to the queue. This queue is used to systematically manage all paths to be analyzed.

[0045] When the tracing queue is not empty, the following steps are executed cyclically: a. Retrieve the current node: Retrieve the knowledge node to be analyzed and its path information from the initial question knowledge node from the head of the tracing queue; b. Obtain the parent node: In the neighborhood knowledge graph, find all direct parent nodes of the current node, i.e., its prerequisite knowledge points, along the reverse of the hard prerequisite relationship edge; c. Analyze each parent node: Perform the following fusion analysis on each parent node: Status query: Obtain the key indicators of the parent node from the user's personal knowledge status graph. The key indicators include: Instantaneous mastery: The latest mastery probability value of the node; Temporal change trend: Calculate its linear regression slope based on the mastery probability values ​​of the node when it was last updated ten times. If the slope is negative, the trend is determined to be downward; if the slope is zero or positive, the trend is determined to be stable or upward. Decision and marking: If the parent node simultaneously satisfies that the instantaneous mastery is lower than the second preset threshold (e.g., the second preset threshold is 70%) and its change trend is downward, then mark the parent node as a fundamental knowledge weakness and record the complete tracing path. The tracing of this path will not continue upwards; if the parent node only meets the condition that the instantaneous mastery is lower than the second preset threshold (e.g., the second preset threshold is 70%) but its trend is not downward, then the parent node will be added back to the tail of the tracing queue, and its own parent node will be further analyzed; if the instantaneous mastery of the parent node is higher than or equal to the second preset threshold, then the node is considered not a weak point, and the tracing path upwards from the node will stop.

[0046] When the above loop ends and the tracing queue is empty, all possible tracing paths have been analyzed. The system collects all knowledge nodes marked as fundamental knowledge weaknesses, forming the final set of fundamental knowledge weaknesses. The system also saves the complete dependency path from the initial problem knowledge node to each fundamental knowledge weakness in the set, clearly demonstrating the transmission path of knowledge deficiencies. Finally, the system outputs the set of fundamental knowledge weaknesses and the corresponding set of dependency chains.

[0047] like Figure 6 As shown, this invention provides a method for generating learning paths and question sequences; Specifically: The initial problem knowledge node that triggered this source tracing analysis is set as the short-term learning objective, and constraints are defined. Extract all hard prerequisite relationship paths connecting the set of fundamental knowledge weaknesses and the target knowledge nodes from the domain knowledge graph to form a target dependency path set; For each path in the target dependency path set, calculate its path comprehensive weakness, and sort all paths to obtain the priority learning path set; From all paths in the priority learning path set, extract the starting knowledge node of the learning path; calculate the weighted probability of the user's mastery of the starting knowledge node with its centrality in the domain knowledge graph to obtain the knowledge repair priority score, and sort them from low to high to form the first learning path; For each knowledge node in the first learning path, an adaptive question sequence is dynamically generated; the adaptive question sequence includes diagnostic questions and reinforcement questions.

[0048] It is important to note that the constraints include: knowledge constraints, the learning path must strictly follow the hard prerequisite relationships in the domain knowledge graph to ensure that the path is logically feasible; mastery constraints, the target mastery probability of each knowledge node on the path needs to reach or exceed the mastery threshold (e.g., 85%); and cognitive load constraints, the number of recommended new knowledge nodes in a single learning session does not exceed the cognitive load limit (e.g., 5) to avoid information overload for users.

[0049] Starting with each node in the set of fundamental knowledge weaknesses and ending with a short-term learning objective, the algorithm searches for all existing directed paths for each pair of start and end points within the hard pre-relationship network of the neighborhood knowledge graph. This can be achieved using a depth-first search algorithm. To ensure efficiency and practicality, the path search should have a maximum path length limit (e.g., 6 nodes) and automatically ignore paths containing cycles. All found paths are then aggregated to form a target dependency path set. This set contains all knowledge dependency chains that need to be repaired.

[0050] For each path in the target-dependent path set, obtain the mastery probability values ​​of all fundamental knowledge weaknesses on that path from the user's personal knowledge graph. Calculate the overall weakness score for that path, which is the minimum mastery probability of all fundamental knowledge weaknesses on that path. This ensures that paths containing the weakest links are prioritized for improvement, addressing the biggest obstacles to achieving short-term learning goals. Based on the calculated overall weakness scores, all paths are sorted in ascending order, with the weakest paths listed first. The sorted set of paths is designated as the priority learning path set.

[0051] From all paths in the prioritized learning path set, extract the starting knowledge node for each path, which represents the fundamental knowledge weakness corresponding to that path, forming a set of starting nodes to be repaired. For each starting node K in the set... s : Obtain its current mastery probability p s ; Calculate the centrality c of this node in the entire domain knowledge graph. s The specific metrics for centrality include degree centrality, proximity centrality, betweenness centrality, and eigenvector centrality. Preferably, in this example, we use in-degree centrality as the metric. This is because a higher in-degree of a piece of knowledge indicates that it is more fundamental and core, and repairing it will have a greater positive impact on the learning of subsequent knowledge. In-degree centrality refers to the number of hard predecessor edges pointing to the node. To compare different nodes, normalization is usually required, i.e., dividing the node's in-degree by the maximum possible in-degree in the domain knowledge graph (usually the total number of nodes minus one). The knowledge repair priority score is obtained through weighted calculation, using the formula: S(K s )=ω p ×(1-p s )+ω c (1-c s ), where ω p ω c Weighting coefficients (e.g., ω) p It is 0.7, ω c (0.3), (1-p) s Emphasizing the degree of weakness, (1-c) s Nodes with high centrality are assigned higher priority. The lower the knowledge repair priority score, the higher the repair priority. Based on the knowledge repair priority score, the starting nodes are sorted in ascending order, and the starting node with the lowest score is selected. The complete path from the set of priority learning paths containing this starting node is defined as the first learning path. If this starting node exists in multiple priority learning paths, the path with the lowest overall weakness is selected as the first learning path. The first learning path is the optimal path from the starting point that most needs repair to the short-term learning goal.

[0052] For each knowledge node in the first learning path: if the knowledge node represents a fundamental knowledge weakness, diagnostic questions are configured for that node to accurately detect and reinforce that weakness; if the knowledge node is an intermediate node on the path, reinforcement questions are configured for that node to ensure the path remains clear. For the short-term learning target node itself, comprehensive application questions are configured to test the final learning effect. The system maintains a user question exposure record table, recording the questions the user has recently answered. When recommending questions for each knowledge node in the first learning path, the system prioritizes questions with clear examination relationships that have not been answered by the user recently (e.g., questions not included in the user's last twenty answers). By calculating the semantic similarity between the recommended questions and the user's set of completed questions (using the aforementioned optimized domain semantic encoder), questions that are highly similar to completed questions (e.g., similarity exceeding 0.9) are filtered out to ensure the diversity and coverage of the recommended questions and avoid redundant training. The question sequence on the first learning path is generated sequentially according to the order of the knowledge nodes. Users need to complete the questions in the current node and reach the mastery requirement (such as answering three questions correctly in a row) before the system will unlock the questions in the next node, in order to ensure that learning is gradual and the knowledge structure is solid.

[0053] Optionally, the diagnostic recommendation module is further configured to dynamically switch between different recommendation strategy modes based on the user's current learning stage and set learning objectives, thereby achieving precise teaching intervention. The recommendation strategy modes include: remedial mode, advanced mode, and leapfrog mode.

[0054] The system achieves automatic mode switching through a rule-based state machine. The decision-making process includes: Data input layer: user's personal knowledge status graph, learning stage identifiers (such as synchronous learning, review, preview), learning goals (such as interest tags), and recent interaction data (such as answer accuracy rate, answer time, and search history).

[0055] Mode switching condition judgment: The remedial mode is triggered when the recent answer accuracy rate of a knowledge node is below 60% and the dependency chain tracing model successfully locates a non-empty set of fundamental knowledge weaknesses. The advanced mode is triggered when the mastery probability of the current core knowledge node continuously exceeds the advanced threshold (e.g., 85%) for three update cycles, and the current core knowledge node is automatically set by the system according to the official teaching schedule, and the remedial mode trigger condition is met. The current core knowledge node is automatically set by the system according to the official teaching schedule, and its specific determination method includes: the system maintains a knowledge node teaching sequence based on the official course outline and textbook catalog; the system traverses this sequence, and automatically sets the direct predecessor node of the first knowledge node whose mastery probability is below the advanced threshold as the user's current core knowledge node; if the user has mastered all nodes in the sequence, then that node is set as the last node in the sequence. The trigger conditions for the transition mode are: within a time period (e.g., within 24 hours), in a non-test state, the number of times you actively search for or click on knowledge nodes that are not related to the current learning path exceeds a preset threshold (e.g., 3 times), or the number of questions you continuously complete in the advanced mode exceeds a preset threshold (e.g., 50 questions) without triggering other modes or meeting the trigger conditions for remedial and advanced modes.

[0056] Execution flow for each mode: Remedial mode: Performs a complete dependency chain tracing analysis, generates a set of fundamental knowledge weaknesses and dependency chains, and outputs a remediation plan according to the learning path generation method. Advanced mode: Locates the current core knowledge node, searches for direct child nodes with a mastery level below 85% along hard prerequisite relationship edges, and selects the optimal child node using a weighted formula, such as: Comprehensive score = node in-degree centrality × 0.6 + cosine similarity between the knowledge node's standardized name and the user's interest tag × 0.4, generating the shortest learning path and question sequence to the new target. Leap mode: Selects the knowledge node with a mastery level exceeding 90% and the highest interest matching degree as the starting point, traverses along soft relationship edges, filters knowledge nodes with high association strength weights, preset difficulty matching the user's historical answer accuracy, organizes them into extended reading topics, and provides accompanying recommended thinking questions and application questions.

[0057] Interaction Module: This module receives user input via an application programming interface (API), which can be natural language text or text converted from speech recognition. It presents the user's knowledge mastery status, including knowledge nodes and their mastery levels displayed in a visual knowledge graph, as well as the learning path and progress shown on a timeline. When the user interacts with a node in the visual knowledge graph, the system presents recommended questions related to that knowledge point.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. An intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph, characterized in that, Comprise: A knowledge graph construction module: for processing raw educational materials, constructing a neighborhood knowledge graph including hard prerequisite relationships and soft association relationships; And for constructing a user's personal knowledge state graph isomorphic with the domain knowledge graph, and based on the user's historical answering data, dynamically calculating the mastery degree index of each knowledge node through a deep knowledge tracking model; An intelligent semantic retrieval module: for mapping the user's natural language question to the corresponding knowledge node in the neighborhood knowledge graph through semantic analysis and expansion by a pre-trained language model, and traversing the question nodes having semantic association with the knowledge node, and returning the retrieval result; A diagnosis recommendation module: for responding to the user's answering result, combining the user's personal knowledge state graph, and according to a dependency chain tracing model based on mastery degree probability, tracing and analyzing in the neighborhood knowledge graph to locate the root knowledge weakness, and generating an adaptive learning path and question sequence; An interaction module: for receiving the user's natural language question; and for presenting the user's knowledge mastery state and the learning path and completion progress to the user. 2.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The construction method of the domain knowledge graph comprises: S1, analyzing and cleaning the raw educational materials, identifying and extracting the smallest knowledge unit that can be independently examined according to the subject education standard, defining it as a knowledge node, and generating a unique identifier and a standard name for each knowledge node; S2, parsing each question in the question bank into structured data, and defining it as a question node; labeling the core knowledge node directly examined by each question node, and establishing a connection with the corresponding knowledge node through an examination relationship edge; S3, extracting the explicit sequence between knowledge nodes according to the official teaching syllabus and textbook directory; analyzing the knowledge definition and problem solving principle to determine the prerequisite knowledge points that must be relied on to master a specific knowledge point; unify the above dependency association into a hard prerequisite relationship edge to form a directed hard prerequisite relationship network; S4, based on the co-occurrence frequency of different knowledge nodes in the same question, the implicit association mined from the user's historical answering data, and the semantic similarity between knowledge nodes, assign an association strength weight to the knowledge node pair through weighted fusion, and establish a soft association relationship edge for node pairs exceeding a preset threshold. 3.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 2, characterized in that: The relationship in the domain knowledge graph also includes a confusion relationship, which is used to identify knowledge node pairs that are easily associated with errors in cognition; this relationship is established based on the association analysis of the error response behavior of the user group on the corresponding question. 4.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The deep knowledge tracking model dynamically calculates the mastery degree index of each knowledge node based on the user's historical answering data, specifically including: M1, fuse the user's historical answering data, the question metadata obtained from the domain knowledge graph, and the user behavior context data to construct a time-series composite interaction sequence as the input source of the deep knowledge tracking model; M2, based on the topological relationship in the neighborhood knowledge graph, generate a knowledge node semantic vector for each knowledge node using graph embedding technology, and through an attention mechanism, weight and aggregate the knowledge node semantic vectors involved in each sequence unit in the composite interaction sequence to form an interaction representation vector; M3, a double-path architecture that simultaneously inputs the interaction representation vector into the time sequence model and the graph neural network in parallel, respectively captures the evolution law of the user's knowledge state in time and the dependency relationship on the domain knowledge graph structure, and fuses to generate an enhanced knowledge state vector; M4, based on the enhanced knowledge state vector, the deep knowledge tracking model synchronously outputs the dynamic mastery probability of all knowledge nodes and the uncertainty quantification of each mastery estimation; M5, introduce graph logic constraints to ensure that the prediction results conform to the cognitive rules; and update the output results to the user's personal knowledge state graph in real time. 5.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The intelligent semantic retrieval module specifically comprises: N1, continuously train the general pre-trained language model using the corpus in the field, and adjust the model based on the constructed paired samples; N2, input the user's natural language question into the optimized pre-trained language model to generate a question semantic vector; in the same semantic vector space, calculate the similarity between the question semantic vector and all knowledge node semantic vectors, and locate the core knowledge node corresponding to the user's problem intention from the domain knowledge graph; N3, take the core knowledge node as the starting point, traverse along the soft association relationship edge in the domain knowledge graph, collect all associated knowledge nodes, and form a weighted query knowledge node set; N4, traverse the query knowledge node set, collect associated question nodes through the relationship edge, form a candidate question set, and fuse the semantic matching degree, knowledge point association strength and question quality to perform multi-dimensional weighted sorting and screening on the candidate questions; N5, organize the sorted question nodes and their metadata into a structured result, and attach explanatory information based on the core knowledge point and associated knowledge point, and finally return the retrieval result to the user. 6.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The dependency chain tracing model based on mastery probability specifically comprises: P1, when it is detected that the user's recent answer accuracy on the current knowledge node is lower than a first preset threshold, trigger the tracing analysis of the node; take the current node as the starting point of the tracing chain and mark it as the initial problem knowledge node; P2, based on the fusion analysis of the instantaneous state and the time sequence change trend of the mastery degree of each node in the user's personal knowledge state graph, dynamically decide the depth and direction of the tracing path in the hard prerequisite relationship network of the domain knowledge graph, and mark the root knowledge weakness point of this tracing; P3, after the tracing analysis of all paths is terminated, collect all knowledge nodes marked as root knowledge weakness points to form a final root knowledge weakness point set; at the same time, record all dependency chains from the initial problem knowledge node to each root knowledge weakness point. 7.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 6, characterized in that: The fusion analysis includes judging whether the parent node simultaneously satisfies the conditions of mastery degree being lower than a second preset threshold and its mastery degree change trend satisfying the decline condition; the calculation method of the decline condition includes determining whether the linear regression slope of the mastery probability in the sliding window is positive or negative. 8.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The generation of the adaptive learning path and question sequence specifically comprises: R1, set the initial problem knowledge node that triggers this tracing analysis as the short-term learning goal, and define the constraint condition; R2, from the field knowledge graph, extract all hard prerequisite relationship paths connecting the set of basic knowledge weak points and the target knowledge node to form a set of target dependency paths; R3, for each path in the set of target dependency paths, calculate its path comprehensive weakness degree, and sort all paths to obtain a set of priority learning paths; R4, from all paths in the set of priority learning paths, extract the starting knowledge node of the learning path; calculate the mastery probability of the starting knowledge node by the user and its centrality in the field knowledge graph to obtain a knowledge repair priority score, and sort from low to high to form a first learning path; R5, for each knowledge node in the first learning path, dynamically generate an adaptive question sequence; the adaptive question sequence includes diagnostic questions and consolidation questions. 9.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 8, characterized in that: The constraint conditions include knowledge constraints, mastery constraints, and cognitive load constraints; the dynamically generated adaptive question sequence introduces a question exposure control mechanism when recommending questions to prevent similar questions from being repeatedly recommended to users. 10.The intelligent question bank retrieval and recommendation system based on artificial intelligence knowledge graph of claim 1, characterized in that: The diagnostic recommendation module is also configured to dynamically switch between different recommendation strategy modes according to the user's current learning stage and the set learning goal; the recommendation strategy modes include remediation mode, advancement mode, and transition mode.

Images