Multidisciplinary automatic question generating method and system based on knowledge graph

Abstract

This invention discloses a method and system for automatic question generation across multiple disciplines using an AI algorithm based on knowledge graphs, relating to the field of natural language processing. The method includes: integrating textbook and test question databases from multiple disciplines; constructing a multi-disciplinary knowledge graph containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes through knowledge extraction and fusion; receiving a question generation instruction containing discipline type, target knowledge point, question type, and difficulty requirements; extracting relevant subgraphs from the discipline knowledge graph corresponding to the discipline type; encoding the relevant subgraphs and the question type requirements in the question generation instruction and inputting them into a conditional text generation model to automatically generate a test paper template, including question text, answers, and explanations; performing difficulty prediction and logical consistency detection on the test paper template, and optimizing it based on the detection results. This application solves the technical problem of insufficient accuracy in question matching in existing multi-disciplinary automatic question generation methods, achieving the technical effect of improving the accuracy of question matching.

Description

Technical Field

[0001] This application relates to the field of natural language processing, and in particular to a method and system for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs. Background Technology

[0002] Multidisciplinary automated question generation technology is a core supporting component in the field of intelligent education. Its application can significantly reduce the manpower cost of test paper compilation, improve the accuracy of matching questions with knowledge points, and has irreplaceable industrial value for large-scale teaching assessment and personalized learning program development. Currently, the mainstream technical methods for solving multidisciplinary question generation problems are mostly based on question bank retrieval and simple rule matching. Specifically, this involves pre-building a subject-specific question database, filtering questions related to target knowledge points through keyword matching, and then generating test papers by combining them according to a preset ratio of question types and difficulty. Existing methods lack the ability to structurally model the subject knowledge system, relying only on superficial keyword associations rather than the deep semantic logic between knowledge points. This results in generated test papers with problems such as one-sided knowledge point coverage, unbalanced difficulty gradients, and insufficient design of comprehensive cross-knowledge point questions.

[0003] At present, among the relevant technologies, the automatic question generation for multiple disciplines suffers from insufficient accuracy in question matching. Summary of the Invention

[0004] This application provides a method and system for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs. It integrates multidisciplinary textbooks and test question databases, constructs a multidisciplinary knowledge graph containing knowledge point entities, semantic relationships, and difficulty and cognitive level attributes through knowledge extraction and fusion, receives question generation instructions containing subject type, target knowledge point, question type, and difficulty requirements, extracts relevant subgraphs from the corresponding subject knowledge graph, encodes the relevant subgraphs and question type requirements, and inputs them into a conditional text generation model. This automatically generates a test paper template containing question text, answers, and explanations. The test paper template undergoes difficulty prediction and logical consistency checks, and the test paper is optimized based on the results. These technical means address the technical problem of insufficient accuracy in question matching in existing multidisciplinary automatic question generation systems, achieving a significant improvement in the accuracy of question matching.

[0005] This application provides a multi-disciplinary automatic question generation method based on knowledge graph AI algorithms, including: integrating textbook and test question databases from multiple disciplines; constructing a multi-disciplinary knowledge graph containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes through knowledge extraction and fusion; receiving a question generation instruction containing subject type, target knowledge point, question type, and difficulty requirements; extracting relevant subgraphs from the subject knowledge graph corresponding to the subject type; encoding the relevant subgraphs and the question type requirements in the question generation instruction and inputting them into a conditional text generation model to automatically generate a test paper template, including test question text, answers, and explanations; performing difficulty prediction and logical consistency detection on the test paper template, and optimizing it based on the detection results.

[0006] In one possible implementation, textbook and test question databases from multiple disciplines are integrated. Through knowledge extraction and fusion, a multi-disciplinary knowledge graph is constructed, containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. The following processing is performed: knowledge extraction is performed based on the first textbook data of the first discipline to generate a first set of knowledge point entities; a set of questions associated with the first set of knowledge point entities is identified and extracted from the first test question database corresponding to the first discipline; semantic relationships, difficulty, and cognitive level attributes between different knowledge point entities are established by analyzing single-knowledge-point questions and multi-knowledge-point questions to establish a first-discipline knowledge graph corresponding to the first discipline; and the first-discipline knowledge graph is added to the multi-disciplinary knowledge graph.

[0007] In a possible implementation, the semantic relationships, difficulty, and cognitive level attributes between entities of different knowledge points are established by analyzing single-knowledge-point questions and multi-knowledge-point questions, thereby establishing a first-subject knowledge graph corresponding to the first subject. The following processing is then performed: extracting single-knowledge-point question sets and multi-knowledge-point question sets from the question set; identifying the difficulty and cognitive level attributes of the single-knowledge-point question sets based on expert-labeled tags to establish a single-knowledge-point graph; performing semantic relationship analysis between related entities and identifying the difficulty and cognitive level attributes based on expert-labeled tags on the multi-knowledge-point question sets to generate a multi-knowledge-point fused knowledge graph; and concatenating the single-knowledge-point graph and the multi-knowledge-point fused knowledge graph to generate the first-subject knowledge graph.

[0008] In a possible implementation, the relevant subgraphs and the question type requirements in the question generation instructions are encoded and input into a conditional text generation model to automatically generate a test paper template. The following processes are then performed: parsing the question type requirements to generate the distribution characteristics of question type knowledge points and the distribution characteristics of question type quantity; obtaining a reference subgraph for each question type from the relevant subgraphs based on the distribution characteristics of question type knowledge points; obtaining the quantity requirements for each question type based on the quantity distribution characteristics of question type; inputting the encoded reference subgraphs and quantity requirements for each question type into the conditional text generation model to generate a question combination for each question type, including the question, answer, and explanation; and combining the question combinations for each question type according to a preset test paper layout rule to generate the test paper template.

[0009] In a possible implementation, the reference subgraph and quantity requirements for each question type are encoded and input into the conditional text generation model to generate a question combination for each question type. The following processing is then performed: for the reference subgraph of each question type, a cognitive trap map representing cognitive traps is established; the cognitive trap map, together with the reference subgraph and quantity requirements for each question type, are encoded and input into the conditional text generation model to generate the corresponding question combination for each question type.

[0010] In a possible implementation, the following processing is performed: the cognitive trap map includes a first trap map for linking any knowledge point with common learner misunderstandings and misinterpretations, a second trap map for linking a knowledge point with other easily confused knowledge points, and a third trap map for linking a knowledge point with the prerequisites or implicit assumptions that must be relied upon for correct understanding and application.

[0011] In a possible implementation, the following processing is performed: the conditional text generation model is a large language model, and different question types correspond to different large language models; the corresponding large language model is called according to the question type, and the reference subgraph and quantity requirements of each question type are analyzed to generate intermediate question combinations; the large language model is used to deploy cognitive defects in the intermediate question combinations based on the cognitive trap map to generate the final question combination.

[0012] In a possible implementation, the following processing is performed: when setting cognitive deficits, cognitive deficits are set for the options based on the cognitive trap map for multiple-choice questions, and cognitive deficits are set for the question stems based on the cognitive trap map for fill-in-the-blank and short-answer questions.

[0013] In a possible implementation, the test paper template is subjected to difficulty prediction and logical consistency detection. Based on the detection results, optimization is performed, and the following processing is executed: Independent difficulty features and overall structural features are extracted from the test paper template. Independent difficulty features include the difficulty of knowledge points associated with the question text and text complexity features. Overall structural features include the total number of questions, question type distribution, and knowledge point distribution dispersion. The independent difficulty features and overall structural features are input into a trained test paper difficulty prediction model, and the overall predicted difficulty value of the test paper template is output. For all questions, answers, and explanations in the test paper template, the core concepts, theorems, and factual relationships are verified to conform to a preset knowledge reasoning graph, obtaining a logical consistency detection result. If the overall predicted difficulty value does not conform to the difficulty corresponding to the question-generating instruction and / or the logical consistency detection result is inconsistent, the abnormal questions are regenerated.

[0014] This application also provides a multi-disciplinary automatic question generation system based on knowledge graph AI algorithms, including: a subject knowledge graph construction module, used to integrate textbook and test question databases of multiple subjects, and construct multiple subject knowledge graphs containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes through knowledge extraction and fusion; a question generation instruction receiving module, used to receive question generation instructions containing subject type, target knowledge point, question type, and difficulty requirements, and extract relevant subgraphs from the subject knowledge graph corresponding to the subject type; a conditional text generation module, used to encode the relevant subgraphs and the question type requirements in the question generation instructions and input them into a conditional text generation model to automatically generate a test paper template, including test question text, answers, and explanations; and an optimization module, used to perform difficulty prediction and logical consistency detection on the test paper template, and optimize it according to the detection results.

[0015] This application proposes a knowledge graph-based AI algorithm for multi-disciplinary automatic question generation. First, it integrates textbook and test question databases from multiple disciplines. Through knowledge extraction and fusion, it constructs a multi-disciplinary knowledge graph containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. Next, it receives a question generation instruction containing the discipline type, target knowledge point, question type, and difficulty requirements. Relevant subgraphs are extracted from the discipline knowledge graph corresponding to the discipline type. Then, the relevant subgraphs and the question type requirements in the question generation instruction are encoded and input into a conditional text generation model to automatically generate a test paper template, including question text, answers, and explanations. Finally, the test paper template undergoes difficulty prediction and logical consistency checks, and is optimized based on the results. Through this process, the proposed method and system achieve the technical effect of improving the accuracy of question matching. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating the multi-disciplinary automatic question generation method based on knowledge graph AI algorithms provided in this application embodiment.

[0018] Figure 2 This is a schematic diagram of the structure of a knowledge graph-based AI algorithm-based multi-disciplinary automatic question generation system provided in an embodiment of this application.

[0019] Figure labeling: Subject knowledge graph construction module 10, question generation instruction receiving module 20, conditional text generation module 30, optimization module 40. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This application provides an embodiment of a multi-disciplinary automatic question generation method based on knowledge graph AI algorithms, such as... Figure 1 As shown, the method includes:

[0022] Step S100: Integrate textbook and test question databases from multiple disciplines, and construct a multi-disciplinary knowledge graph that includes knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes through knowledge extraction and fusion.

[0023] Specifically, this involves establishing structured knowledge networks for different disciplines, transforming scattered textbook knowledge points and test question information into machine-understandable graph data. Here, knowledge point entities refer to specific concepts in a discipline, such as quadratic equations in mathematics and Newton's second law in physics. Semantic relationships between entities refer to the logical connections between knowledge points, such as the quadratic equation formula being a derived knowledge point of quadratic equations. Difficulty attributes refer to the level of difficulty of the questions corresponding to the knowledge points, such as easy, medium, and difficult. Cognitive level attributes refer to the cognitive requirements corresponding to the knowledge points, such as memorization, understanding, application, analysis, evaluation, and creation.

[0024] This study employs a knowledge extraction framework combining rule-based and deep learning. First, it uses regular expressions to match keywords related to knowledge points in textbook texts, such as matching definitions, theorems, and properties as prefixes to identify knowledge point entities. Then, it utilizes a combined bidirectional long short-term memory network and conditional random field model to extract knowledge point entities and preliminary relationships between them from textbook paragraphs and test question stems. For difficulty and cognitive level attributes, a labeling dictionary is constructed, including a difficulty label dictionary and a Bloom cognitive level label dictionary. Next, a text classification model is used to label the knowledge points corresponding to the test questions. Finally, a knowledge fusion tool is used to align the same knowledge points from different sources, such as textbooks and test question banks. For example, uniformly accelerated linear motion in textbooks and uniformly accelerated linear motion calculations in test question banks are merged into a single entity, eliminating redundant data and forming a unified multidisciplinary knowledge graph.

[0025] In one possible implementation, textbook and test question databases from multiple disciplines are integrated. Through knowledge extraction and fusion, a multi-disciplinary knowledge graph is constructed, containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. Step S100 further includes step S110, which involves knowledge extraction based on the first textbook data of the first discipline to generate a first knowledge point entity set. Specifically, core knowledge point concepts are extracted from the textbooks of a single discipline. Here, the first discipline refers to any target discipline, such as high school chemistry; the first textbook data refers to the standardized textbook text data of that discipline, such as the People's Education Press version of high school chemistry compulsory course 1; and the first knowledge point entity set refers to the set of all knowledge point concepts in the textbooks of that discipline. The textbook text is preprocessed, including text cleaning and word segmentation, and then a named entity recognition model is used for knowledge point extraction. The named entity recognition model adopts a combination model of bidirectional long short-term memory network and conditional random field. The input of the model is the word-segmented textbook text sequence, and the output is a text fragment labeled with knowledge point tags. After deduplication of the labeled text fragments, the first knowledge point entity set is generated. For example, knowledge points such as amount of substance, molar mass, and molar volume of gases can be extracted from the compulsory first-year high school chemistry textbook. Each knowledge point entity is uniquely identified and coded in the format of subject code-chapter number-knowledge point number to ensure the uniqueness of the entity.

[0026] Step S120: Identify and extract the set of questions associated with the first knowledge point entity set from the first test question database corresponding to the first subject. Analyze single-knowledge-point questions and multi-knowledge-point questions to establish semantic relationships, difficulty, and cognitive level attributes between different knowledge point entities, thus establishing a first subject knowledge graph corresponding to the first subject. Specifically, combine test question data to improve the relationship and attribute information of the subject knowledge graph, associate knowledge points with test questions, assign difficulty and cognitive level attributes to knowledge points, and establish semantic relationships between knowledge points. Construct a test question-knowledge-point matching model. The model uses a text similarity calculation algorithm to calculate the cosine similarity between the test question stem text and the knowledge point entity name and definition text. Set the similarity threshold to 0.7, and filter out test questions with similarity higher than the threshold to form a set of questions associated with the first knowledge point entity set. The question set is divided into two categories: single-knowledge-point questions and multi-knowledge-point questions. Single-knowledge-point questions test only one knowledge point, such as calculating the mass of 1 mole of water, which only tests molar mass. Multi-knowledge-point questions test two or more knowledge points, such as calculating the volume and mass of 1 mole of carbon dioxide under standard conditions, which tests both molar volume and molar mass of a gas. For single-knowledge-point questions, their difficulty and cognitive level attributes are determined through manual annotation or text classification models, and these attributes are assigned to the corresponding knowledge point entities. For multi-knowledge-point questions, association rule mining algorithms, such as prior algorithms, are used to mine the co-occurrence relationships of knowledge points in the questions. For example, if the co-occurrence frequency of molar volume and molar mass of a gas is found to be 30%, a semantic relationship between the two knowledge points is established. Finally, the knowledge point entities, the semantic relationships between entities, and the attribute information are integrated and stored in a graph database to form a first-subject knowledge graph.

[0027] Step S130: Add the first subject knowledge graph to the multiple subject knowledge graphs. Specifically, using the merging function of the graph database, import the entity and relation data of the first subject knowledge graph into the database of the multi-subject knowledge graph. During the import process, perform cross-subject entity alignment checks. For example, check whether there is a relationship between function knowledge points in mathematics and function graph analysis knowledge points in physics. If so, establish cross-subject semantic relationships. After the import is completed, update the metadata information of the multi-subject knowledge graph, including the number of subjects, the total number of entities, and the total number of relations. Finally, optimize the index of the merged knowledge graph by establishing a multi-dimensional index based on knowledge point name, subject type, and difficulty attribute to improve query efficiency.

[0028] In one possible implementation, semantic relationships, difficulty, and cognitive level attributes between different knowledge point entities are established by analyzing single-knowledge-point questions and multi-knowledge-point questions, thus constructing a first-subject knowledge graph corresponding to the first subject. Step S120 further includes step S121, extracting a single-knowledge-point question set and a multi-knowledge-point question set from the question set. Specifically, a question knowledge point quantity detection model is constructed. The model input is the text data of the question stem and answer explanation, and the output is the number of knowledge points tested in the question. In specific implementation, the question stem and explanation text are first matched with the knowledge point entity set, and the number of matched knowledge points is counted. Matching rules are set: if the question matches only one knowledge point entity and the explanation does not involve other knowledge points, it is determined to be a single-knowledge-point question; if the question matches two or more knowledge point entities and the explanation contains multiple related applications of knowledge points, it is determined to be a multi-knowledge-point question. For example, solving the equation x... 2 -5x+6=0 only matches the knowledge point of solving quadratic equations, belonging to a single knowledge point question; calculating the acceleration and displacement of an object using Newton's second law and kinematic equations matches two knowledge points: Newton's second law and the formulas for uniformly accelerated linear motion, belonging to a multi-knowledge point question. The question set is divided into two subsets: a single-knowledge point question set and a multi-knowledge point question set.

[0029] Step S122: The single-knowledge-point question set is analyzed based on expert-labeled difficulty and cognitive level attributes to establish a single-knowledge-point graph. Specifically, an expert-labeled dataset is collected, containing single-knowledge-point questions and their corresponding expert-labeled difficulty and cognitive level tags. A text classification model is trained based on this dataset, using a convolutional neural network architecture. The input layer is the word vector sequence of the question text, the convolutional layers use three different sized convolutional kernels, and the pooling layer uses max pooling. The single-knowledge-point question set is input into the trained classification model, which outputs the difficulty and cognitive level attributes for each question. These attributes are then assigned to the corresponding knowledge-point entities. Finally, a single-knowledge-point graph is constructed using knowledge-point entities as nodes and attribute information as node attributes. The single-knowledge-point graph does not contain relationships between entities; it only contains the attribute information of a single knowledge point.

[0030] Step S123 involves performing semantic relationship analysis between related entities and identifying difficulty and cognitive level attributes based on expert-labeled tags on the multi-knowledge-point question set to generate a multi-knowledge-point fusion knowledge graph. Specifically, for each question in the multi-knowledge-point question set, the knowledge point entity pairs it examines are extracted. For example, if a question examines Newton's second law and the formula for uniformly accelerated linear motion, then an entity pair (Newton's second law, formula for uniformly accelerated linear motion) is formed. A relation extraction model is used to identify the semantic relationship between entity pairs. The relation extraction model is fine-tuned using a pre-trained language model, such as a bidirectional transformer model. The fine-tuning dataset is a multi-knowledge-point question relation dataset annotated by experts, and the relation types include derivation relations, dependency relations, and parallel relations. Simultaneously, based on expert-labeled tags, a text classification model is used to identify the overall difficulty and cognitive level attributes corresponding to the question, and the attributes are assigned to the association relations of the knowledge point entity pairs. For example, the association relation of the entity pair (Newton's second law, formula for uniformly accelerated linear motion) is a derivation relation, the difficulty attribute is difficult, and the cognitive level attribute is analysis. Finally, using knowledge point entities as nodes, identified semantic relationships as edges, and relationship attributes as edge attributes, a multi-knowledge point fusion knowledge graph is constructed.

[0031] Step S124: The single-knowledge-point graph is concatenated with the multi-knowledge-point fused knowledge graph to generate the first subject knowledge graph. Specifically, a graph concatenation algorithm is used to merge nodes in the single-knowledge-point graph with nodes of the same name in the multi-knowledge-point fused knowledge graph, retaining all attribute information of the nodes during the merging process. Edges from the multi-knowledge-point fused knowledge graph are added between the merged nodes. For example, a node for solving quadratic equations in a single-knowledge-point graph is merged with a node for solving quadratic equations in a multi-knowledge-point fused knowledge graph, retaining its medium difficulty and application-level cognitive attributes, and adding a derivation relationship edge between this node and the discriminant node of the root. After concatenation, redundant relationships are removed and attribute consistency is checked in the graph, for example, duplicate edges are removed, and different attribute descriptions of the same node are unified. Finally, the concatenated graph is exported as a standard graph data format, completing the construction of the first subject knowledge graph.

[0032] Step S200: Receive a question-generating instruction containing the subject type, target knowledge points, question type, and difficulty requirements, and extract relevant subgraphs from the subject knowledge graph corresponding to the subject type.

[0033] Specifically, based on the user's question-generating requirements, a local knowledge network matching the requirements is selected from the complete subject knowledge graph, providing a precise knowledge scope for question generation. A typical example of a question-generating instruction is: "Middle School Math + Quadratic Equation + Multiple Choice + Medium Difficulty." A related subgraph refers to a subset of the graph containing the target knowledge point, related knowledge points, corresponding difficulty levels, and cognitive level attributes. A question-generating instruction parsing module is constructed, using natural language processing technology to convert the user's input into structured query statements. For example, a "Middle School Math Quadratic Equation Multiple Choice Question (Medium Difficulty)" is transformed into: Subject = Middle School Math, Knowledge Point = Quadratic Equation, Question Type = Multiple Choice, Difficulty = Medium. Based on graph database query language, a query is executed in the subject knowledge graph to filter out knowledge point relationships that contain the target knowledge point entity, corresponding difficulty attribute, cognitive level attribute, and are suitable for the question type. The filtered entities and relationships are exported as a subgraph data structure, containing attribute information for both entities and relationships.

[0034] Step S300: Encode the relevant subgraphs and the question type requirements in the question generation instructions and input them into the conditional text generation model to automatically generate a test paper template, including the question text, answers and explanations.

[0035] Specifically, the structured knowledge subgraphs and question type requirements are transformed into complete test paper content. Here, encoding refers to converting graph structure data and question type parameters into vector forms that the model can accept. The conditional text generation model refers to an artificial intelligence model that generates text based on specific input conditions. The test paper template refers to a standardized test paper format that includes questions, answers, and explanations.

[0036] A graph neural network is used to encode relevant subgraphs, transforming each knowledge point entity into a high-dimensional vector, while the relationships between entities are transformed into connection weights between vectors. For question type requirements, such as multiple choice and fill-in-the-blank questions, parameters such as question type, number of questions, and score are transformed into one-dimensional feature vectors. The subgraph vectors output by the graph neural network are concatenated with the question type feature vectors and used as input to the conditional text generation model. The conditional text generation model adopts an encoder-decoder architecture, with a bidirectional transformer model for the encoder and a unidirectional transformer model for the decoder. The model is fine-tuned using a subject-specific test question dataset during training. When outputting, the model first generates the question text, and then generates the answer and explanation based on the logical relationships between knowledge points in the knowledge graph. For example, when generating a multiple choice question about a quadratic equation, the answer must match the correct solution of the knowledge point, and the explanation must include the application steps of the knowledge point.

[0037] In one possible implementation, the relevant subgraphs and the question type requirements in the question-generating instructions are encoded and input into a conditional text generation model to automatically generate a test paper template. Step S300 further includes step S310, parsing the question type requirements and generating question type knowledge point distribution characteristics and question type quantity distribution characteristics. Specifically, the question type knowledge point distribution characteristics refer to the distribution of knowledge point cognitive levels and difficulty levels adapted to different question types, and the question type quantity distribution characteristics refer to the distribution of the number of questions and scores for different question types. A question type requirement parsing rule base is constructed, which contains adaptation rules for different question types, such as multiple-choice questions adapting to knowledge points at the memory and comprehension levels, and calculation questions adapting to knowledge points at the application and analysis levels. The user inputs their question type requirements, such as 10 multiple-choice questions, 5 fill-in-the-blank questions, and 3 calculation questions. The parsing module then generates knowledge point distribution characteristics for each question type based on a rule base. For example, the knowledge point distribution characteristics for multiple-choice questions are {cognitive level: memory, understanding; difficulty: easy, medium}, and for calculation questions, {cognitive level: application, analysis; difficulty: medium, difficult}. Simultaneously, the module calculates the number of questions and the percentage of marks for each question type, generating a question type quantity distribution characteristic. For example, 10 multiple-choice questions, 3 marks each, total 30 marks; 5 fill-in-the-blank questions, 4 marks each, total 20 marks; 3 calculation questions, 10 marks each, total 30 marks. Finally, both types of features are converted into numerical feature vectors for subgraph filtering and question generation.

[0038] Step S320: Based on the distribution characteristics of knowledge points for each question type, obtain a reference subgraph for each question type from the relevant subgraph. Specifically, for the distribution characteristics of knowledge points for each question type, set filtering conditions. For example, the filtering conditions for multiple-choice questions are: cognitive level ∈ {memory, understanding} and difficulty ∈ {easy, medium}; the filtering conditions for calculation questions are: cognitive level ∈ {application, analysis} and difficulty ∈ {medium, difficult}. A subgraph filtering algorithm is used to perform condition filtering in the relevant subgraph, extracting knowledge point entities and relationships that meet the filtering conditions for each question type. The filtered entities and relationships are combined to generate a reference subgraph for each question type. The reference subgraph contains all knowledge point information suitable for the question type.

[0039] Step S330: Obtain the quantity requirement for each question type based on the question type quantity distribution characteristics. Specifically, parse the question quantity information in the question type quantity distribution characteristics and convert it into quantity requirement parameters for each question type. The parameter format is: Question Type - Number of Questions - Score per Question. For example, extract the quantity requirements from the question type quantity distribution characteristics as follows: Multiple Choice - 10 - 3, Fill-in-the-Blank - 5 - 4, Calculation - 3 - 10. Verify the quantity requirements to ensure that the sum of the number of questions for each question type matches the total number of questions requested by the user, and that the sum of the scores matches the total score. If there is a discrepancy, output a prompt message requesting the user to adjust the question type requirements. After successful verification, store the quantity requirement parameters as key-value pairs for subsequent model input.

[0040] Step S340: Input the reference subgraph and quantity requirement encoding for each question type into the conditional text generation model to generate a question combination for each question type, including the question, answer, and explanation. Specifically, the reference subgraph for each question type is encoded using a graph neural network to generate a subgraph vector. The quantity requirement parameters are numerically encoded to generate a one-dimensional feature vector. The subgraph vector and the quantity requirement vector are concatenated and used as input to the conditional text generation model. Different model decoding strategies are used for different question types. For example, when decoding multiple-choice questions, the model needs to generate the question stem, four options (one correct and three distractors), when decoding fill-in-the-blank questions, the model needs to generate the question stem and the answer to the blank, and when decoding calculation questions, the model needs to generate the question stem, solution steps, and final answer. The question combination output by the model includes the complete question text, standard answer, and detailed explanation. The explanation corresponds to the logical relationship of knowledge points in the reference subgraph. For example, the explanation for multiple-choice questions needs to explain the knowledge points corresponding to the correct options and the cognitive traps corresponding to the distractors. Finally, the generated question combinations are stored according to question type.

[0041] Step S350: Combine the question groups of each question type according to the preset test paper layout rules to generate the test paper template. Specifically, a preset test paper layout rule library is constructed, which contains the structural specifications of the test paper, such as the question type order: multiple choice questions → fill-in-the-blank questions → calculation questions; the question numbering rule: consecutive numbering; and the position of the score label: right side of each question. Using the test paper combination algorithm, the question groups of each question type are arranged sequentially according to the specifications in the rule library. For example, first arrange 10 multiple choice questions, then arrange 5 fill-in-the-blank questions, and finally arrange 3 calculation questions. During the arrangement process, header information such as the test paper title, exam time, and total score, as well as footer information such as the answer area and score column, are added. The format of the arranged test paper is validated to ensure that the question numbers are consecutive, the score labels are correct, and the explanation content is complete. Finally, the validated test paper is exported as an editable document format or a printable format to generate the final test paper template.

[0042] In one possible implementation, the reference subgraph and quantity requirement encoding for each question type are input into a conditional text generation model to generate a question combination for each question type. Step S340 further includes step S341, which establishes a cognitive trap map representing cognitive traps for the reference subgraph of each question type. The cognitive trap map includes a first trap map linking any knowledge point to common learner misunderstandings and misinterpretations, a second trap map linking a knowledge point to other easily confused knowledge points, and a third trap map linking a knowledge point to prerequisites or implicit assumptions necessary for correct understanding and application. Specifically, cognitive traps refer to misunderstandings or application errors that learners are prone to when learning knowledge points. A dataset of common cognitive traps in various subjects is collected. The dataset contains information such as knowledge points, misunderstandings, easily confused knowledge points, and prerequisites. For example, a common misunderstanding of a quadratic equation is that the equation has no solution when the discriminant is less than zero; an easily confused knowledge point is a linear equation, and the prerequisite is that the coefficient of the quadratic term is not zero. Based on this dataset, three types of cognitive trap graphs were constructed: the first trap graph uses knowledge points as source nodes and incorrect expressions as target nodes, with edge types of error association; the second trap graph uses knowledge points as source nodes and easily confused knowledge points as target nodes, with edge types of confusion association; the third trap graph uses knowledge points as source nodes and prerequisites as target nodes, with edge types of dependency association. A graph database was used to store the cognitive trap graphs, adding attribute information to each node and edge, such as error frequency, confusion level, and dependency strength. For example, the confusion level attribute value between quadratic equations and linear equations was set to high. Finally, the cognitive trap graphs were associated with a question type reference subgraph to ensure that the knowledge points in the trap graphs were consistent with those in the reference subgraphs.

[0043] Step S342: The cognitive trap map, along with the reference subgraph and quantity requirement for each question type, are encoded and input into the conditional text generation model to generate a question combination for each question type. The conditional text generation model is a large language model, and different question types correspond to different large language models. The corresponding large language model is called according to the question type to analyze the reference subgraph and quantity requirement for each question type, generating intermediate question combinations. The large language model is then used to perform cognitive defect placement on the intermediate question combinations based on the cognitive trap map, generating the final question combination. Specifically, for multiple-choice questions, cognitive defect placement is performed on the options based on the cognitive trap map; for fill-in-the-blank and short-answer questions, cognitive defect placement is performed on the question stem based on the cognitive trap map. Specifically, the cognitive trap map is encoded using a graph neural network to generate trap vectors, with the vector dimension consistent with the reference subgraph vector dimension. The trap vector, reference subgraph vector, and quantity requirement vector are concatenated and used as input to the large language model. Dedicated large language models are deployed for different question types. For example, the multiple-choice question model uses a variant of the bidirectional transformer model, and the fill-in-the-blank question model uses a variant of the unidirectional transformer model. Both models are fine-tuned using subject-specific test questions and cognitive trap datasets. The corresponding large language model is invoked. First, an intermediate question set is generated based on the reference subgraph and quantity requirements. This intermediate question set consists of basic questions without cognitive traps. Then, cognitive deficit placement is performed based on the cognitive trap map. For multiple-choice questions, cognitive traps are placed in the options; for example, incorrect statements from the first trap map are set as distractors, and statements related to easily confused knowledge points from the second trap map are set as distractors. For fill-in-the-blank and short-answer questions, cognitive traps are placed in the question stem; for example, preconditions from the third trap map are intentionally omitted, or incorrect conceptual statements are inserted into the question stem. After placement, the model generates the final question set, including questions with cognitive traps, correct answers, and explanations. The explanations point out the cognitive traps in the questions and the corresponding correct understanding methods. For example, the analysis of multiple-choice questions should explain the error type of the distractor options, and the analysis of fill-in-the-blank questions should explain the prerequisites that are missing in the question stem.

[0044] Step S400: Perform difficulty prediction and logical consistency detection on the test paper template, and optimize it based on the detection results.

[0045] Specifically, difficulty prediction refers to the quantitative assessment of the overall difficulty of the test paper, while logical consistency detection verifies whether the content of the test questions is consistent with the logic of the knowledge points in the knowledge graph. A test paper feature extraction module is constructed to extract independent difficulty features and overall structural features of the test paper. For difficulty prediction, a gradient boosting regression tree model is used as the test paper difficulty prediction model. During model training, the feature data of historical test papers and manually labeled difficulty values ​​are used as the training set, and the mean squared error is used as the loss function. The extracted features are input into the trained model, and the overall predicted difficulty value of the test paper is output. For logical consistency detection, a knowledge reasoning rule base is constructed. For example, rules such as "a quadratic equation has no real roots when the discriminant is less than zero" and "Newton's second law states that the resultant force equals mass multiplied by acceleration" are used. A rule matching algorithm is used to compare the key expressions in the test questions, answers, and explanations with the rule base to verify the correct application of core concepts and theorems. If the predicted difficulty value deviates from the required value by more than a preset threshold, or if the logical consistency detection finds an error, an optimization mechanism is triggered to regenerate abnormal questions until they meet the requirements.

[0046] In one possible implementation, the test paper template undergoes difficulty prediction and logical consistency detection, and optimization is performed based on the detection results. Step S400 further includes step S410, extracting independent difficulty features and overall structural features from the test paper template. Independent difficulty features include the difficulty of the knowledge points associated with the question text and text complexity features. Overall structural features include the total number of questions, question type distribution, and the dispersion of knowledge point distribution. Specifically, a feature extraction module is constructed, divided into two parts: independent difficulty feature extraction and overall structural feature extraction. During independent difficulty feature extraction, for each question, the difficulty value of its associated knowledge points is extracted from the knowledge graph, and the complexity features of the question text are calculated. Text complexity features include average sentence length, number of technical terms, and number of nested sentence layers. For example, the average sentence length is calculated by statistically analyzing the ratio of the number of characters to the number of sentences in the question stem, and the number of technical terms is calculated by matching a subject-specific terminology database. When extracting overall structural features, the following parameters are considered: the total number of questions in the exam paper (i.e., the sum of the number of questions for each question type), the question type distribution (i.e., the proportion of each question type to the total number of questions), and the dispersion of knowledge point distribution. The dispersion of knowledge point distribution is calculated using the entropy method; a higher entropy value indicates a more uniform distribution of knowledge points. All extracted features are converted into numerical feature vectors. The vector dimension is determined based on the number of features. For example, the independent difficulty feature includes two dimensions: knowledge point difficulty and text complexity. The overall structural feature includes three dimensions: total number of questions, question type distribution ratio, and knowledge point dispersion. Therefore, the total dimension of the feature vector is 5.

[0047] Step S420: Input the independent difficulty features and overall structural features into the trained test paper difficulty prediction model, and output the overall predicted difficulty value of the test paper template. Specifically, the test paper difficulty prediction model adopts a gradient boosting regression tree model. The training process of the model is as follows: Collect historical test paper datasets, and label each test paper with a manually evaluated difficulty value, ranging from 0 to 1, where 0 is the easiest and 1 is the most difficult. Extract the independent difficulty features and overall structural features of each test paper to form a feature matrix. Input the feature matrix and the labeled difficulty values ​​into the model, and set the model parameters: 100 decision trees, a decision tree depth of 6, and a learning rate of 0.1. Use five-fold cross-validation to optimize the model parameters to ensure that the model's prediction error is within a preset range. Input the feature vector of the test paper to be tested into the trained model, and the model outputs the overall predicted difficulty value.

[0048] Step S430: For all questions, answers, and explanations in the test paper template, verify whether the core concepts, theorems, and factual relationships conform to the preset knowledge reasoning graph, and obtain the logical consistency detection result. Specifically, a preset knowledge reasoning graph is constructed. This graph is based on the core concepts, theorems, and factual relationships of the subject and includes logical reasoning rules between knowledge points. For example, if 'a' is a quadratic equation, then its quadratic coefficient is not zero; if an object is undergoing uniformly accelerated linear motion, then its acceleration is constant. A detection algorithm combining rule matching and semantic similarity calculation is used to detect the questions, answers, and explanations in the test paper. First, the core concepts and theorem expressions in the text are extracted and matched with the nodes and rules in the knowledge reasoning graph. Then, the semantic similarity between the expressions and the graph rules is calculated, with a similarity threshold set to 0.85. If the match is successful and the similarity is higher than the threshold, logical consistency is determined; if the match fails or the similarity is lower than the threshold, a logical error is determined. Finally, the detection results of all questions are summarized to generate a logical consistency detection report, which includes information such as the error question number, error type, and error content.

[0049] Step S440: If the overall predicted difficulty value does not meet the difficulty corresponding to the question-generating instruction and / or the logical consistency detection result is inconsistent, the abnormal questions are regenerated. Specifically, a difficulty deviation threshold is set. If the deviation between the overall predicted difficulty value and the user's required difficulty value exceeds this threshold, the difficulty is determined to be unacceptable. If there are erroneous questions in the logical consistency detection report, logical inconsistency is determined. For cases where the difficulty does not meet the requirements, the difficulty attribute filtering conditions of the reference subgraph are adjusted. For example, if the predicted difficulty is too low, the proportion of medium and difficult knowledge points is increased, and the questions are regenerated. For cases where logical inconsistency is determined, the knowledge points and cognitive traps corresponding to the erroneous questions are located, the relevant rules in the knowledge reasoning rule base are corrected, or the cognitive trap placement strategy is adjusted, and the questions are regenerated. After regenerating the questions, the detection process of steps S410-S430 is executed again until the predicted difficulty value of the test paper meets the requirements and the logical consistency detection result is consistent. Finally, the optimized test paper template is output as the final version.

[0050] This application's embodiments employ an integrated multidisciplinary textbook and test question database. Through knowledge extraction and fusion, a multidisciplinary knowledge graph containing knowledge point entities, semantic relationships, and difficulty and cognitive level attributes is constructed. It receives question-generating instructions containing subject type, target knowledge point, question type, and difficulty requirements. Relevant subgraphs are extracted from the corresponding subject knowledge graph. After encoding the relevant subgraphs and question type requirements, they are input into a conditional text generation model to automatically generate a test paper template containing test question text, answers, and explanations. The test paper template undergoes difficulty prediction and logical consistency detection, and the test paper is optimized based on the detection results. These technical means solve the technical problem of insufficient accuracy in question matching in existing multidisciplinary automatic question generation, achieving the technical effect of improving the accuracy of question matching.

[0051] In the above text, refer to Figure 1 This paper describes in detail a multi-disciplinary automatic question generation method based on a knowledge graph AI algorithm according to embodiments of the present invention. Next, we will refer to... Figure 2 This invention describes a knowledge graph-based AI algorithm-based multi-disciplinary automatic question generation system according to an embodiment of the present invention.

[0052] The knowledge graph-based AI algorithm-based multi-disciplinary automatic question generation system according to embodiments of the present invention addresses the technical problem of insufficient accuracy in question matching in existing multi-disciplinary automatic question generation systems, thereby improving the accuracy of question matching. The knowledge graph-based AI algorithm-based multi-disciplinary automatic question generation system includes: a subject knowledge graph construction module 10, a question generation instruction receiving module 20, a conditional text generation module 30, and an optimization module 40.

[0053] The subject knowledge graph construction module 10 integrates textbook and test question databases from multiple subjects. Through knowledge extraction and fusion, it constructs multiple subject knowledge graphs containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. The question-generating instruction receiving module 20 receives question-generating instructions containing subject type, target knowledge point, question type, and difficulty requirements, and extracts relevant subgraphs from the subject knowledge graph corresponding to the subject type. The conditional text generation module 30 encodes the relevant subgraphs and the question type requirements in the question-generating instructions and inputs them into the conditional text generation model to automatically generate a test paper template, including test question text, answers, and explanations. The optimization module 40 performs difficulty prediction and logical consistency detection on the test paper template and optimizes it based on the detection results.

[0054] The specific configuration of the subject knowledge graph construction module 10 is described in detail below: As mentioned above, it integrates textbook and test question databases of multiple subjects, and constructs multiple subject knowledge graphs containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes through knowledge extraction and fusion. The subject knowledge graph construction module 10 may further include: a knowledge extraction unit for extracting knowledge based on the first textbook data of the first subject to generate a first set of knowledge point entities; a first subject knowledge graph establishment unit for identifying and extracting a set of questions associated with the first set of knowledge point entities in the first test question database corresponding to the first subject, and establishing the semantic relationships, difficulty, and cognitive level attributes between different knowledge point entities by analyzing single-knowledge-point questions and multi-knowledge-point questions to establish the first subject knowledge graph corresponding to the first subject; and a multiple subject knowledge graph generation unit for adding the first subject knowledge graph into the multiple subject knowledge graphs.

[0055] Specifically, by analyzing single-knowledge-point questions and multi-knowledge-point questions, semantic relationships, difficulty, and cognitive level attributes between entities of different knowledge points are established, thereby constructing a first-subject knowledge graph corresponding to the first subject. The first-subject knowledge graph construction unit may further include: a question set extraction subunit for extracting single-knowledge-point question sets and multi-knowledge-point question sets from the question set; a single-knowledge-point graph construction subunit for identifying the difficulty and cognitive level attributes of the single-knowledge-point question sets based on expert labeling to construct a single-knowledge-point graph; a multi-knowledge-point fusion knowledge graph generation subunit for performing problem-solving semantic relationships between associated entities and identifying the difficulty and cognitive level attributes based on expert labeling to generate a multi-knowledge-point fusion knowledge graph; and a graph splicing subunit for splicing the single-knowledge-point graph and the multi-knowledge-point fusion knowledge graph to generate the first-subject knowledge graph.

[0056] The detailed description of the specific configuration of the conditional text generation module 30 is as follows: As mentioned above, the relevant subgraph and the question type requirements in the question generation instruction are encoded and input into the conditional text generation model to automatically generate a test paper template. The conditional text generation module 30 may further include: a distribution feature generation unit for parsing the question type requirements and generating question type knowledge point distribution features and question type quantity distribution features; a reference subgraph acquisition unit for acquiring a reference subgraph for each question type in the relevant subgraph based on the question type knowledge point distribution features; a quantity requirement acquisition unit for acquiring the quantity requirement for each question type based on the question type quantity distribution features; a question combination generation unit for inputting the encoded reference subgraph and quantity requirement for each question type into the conditional text generation model to generate a question combination for each question type, including the question, answer, and explanation; and a test paper template generation unit for combining the question combinations for each question type according to a preset test paper layout rule to generate the test paper template.

[0057] Specifically, the reference sub-graph and quantity requirements for each question type are encoded and input into the conditional text generation model to generate a question combination for each question type. The question combination generation unit may further include: a cognitive trap map establishment sub-unit for establishing a cognitive trap map representing cognitive traps from the reference sub-graph for each question type; and a question combination generation sub-unit for encoding the cognitive trap map together with the reference sub-graph and quantity requirements for each question type and inputting them into the conditional text generation model to generate a corresponding question combination for each question type.

[0058] The cognitive trap map establishment subunit may further include: the cognitive trap map includes a first trap map for linking any knowledge point with common learner misunderstandings and expressions, a second trap map for linking a knowledge point with other easily confused knowledge points, and a third trap map for linking a knowledge point with the prerequisites or implicit assumptions that must be relied upon for correct understanding and application.

[0059] The question combination generation subunit may further include: the conditional text generation model is a large language model, and different question types correspond to different large language models; according to the question type, the corresponding large language model is called to analyze the reference subgraph and quantity requirements of each question type to generate intermediate question combinations; the large language model is used to deploy cognitive defects in the intermediate question combinations based on the cognitive trap map to generate the final question combination.

[0060] The question combination generation subunit may further include: when setting cognitive deficits, setting cognitive deficits for the options based on the cognitive trap map for multiple-choice questions, and setting cognitive deficits for the question stems based on the cognitive trap map for fill-in-the-blank and short-answer questions.

[0061] The detailed description of the specific configuration of the optimization module 40 is as follows: As mentioned above, the optimization module 40 performs difficulty prediction and logical consistency detection on the test paper template, and optimizes it based on the detection results. The optimization module 40 may further include: a test paper template feature extraction unit for extracting independent difficulty features and overall structural features from the test paper template. The independent difficulty features include the difficulty of the knowledge points associated with the question text and the text complexity features. The overall structural features include the total number of questions, the distribution of question types, and the dispersion of the distribution of knowledge points. The test paper difficulty prediction unit is used to input the independent difficulty features and overall structural features into the trained test paper difficulty prediction model and output the overall predicted difficulty value of the test paper template. The logical consistency detection unit is used to verify whether the core concepts, theorems, and factual relationships of all questions, answers, and explanations in the test paper template conform to the preset knowledge reasoning graph and obtain the logical consistency detection result. The re-question generation unit is used to re-generate questions for abnormal questions if the overall predicted difficulty value does not conform to the difficulty corresponding to the question generation instruction and / or the logical consistency detection result is inconsistent.

[0062] The knowledge graph-based AI algorithm multi-disciplinary automatic question generation system provided in this embodiment of the invention can execute the knowledge graph-based AI algorithm multi-disciplinary automatic question generation method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0063] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A knowledge graph-based AI algorithm for automatic question generation across multiple disciplines, characterized in that: include: By integrating textbook and test question databases from multiple disciplines, and through knowledge extraction and fusion, a multi-disciplinary knowledge graph is constructed, which includes knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. Receive a question-generating instruction containing the subject type, target knowledge points, question type, and difficulty requirements, and extract relevant subgraphs from the subject knowledge graph corresponding to the subject type; The relevant subgraphs and the question type requirements in the question generation instructions are encoded and input into the conditional text generation model to automatically generate a test paper template, including the question text, answers, and explanations; The test paper template is subjected to difficulty prediction and logical consistency detection, and optimized based on the detection results. The step of encoding the relevant subgraphs and the question type requirements in the question-generating instructions and inputting them into the conditional text generation model to automatically generate a test paper template includes: Analyze the question type requirements to generate the distribution characteristics of knowledge points and the distribution characteristics of the number of questions for each question type; Based on the distribution characteristics of the knowledge points of the question type, a reference sub-graph is obtained for each question type in the relevant sub-graph; The quantity requirement for each question type is obtained based on the quantity distribution characteristics of the question types. The reference sub-graph and quantity requirement code for each question type are input into the conditional text generation model to generate a question combination for each question type, including the question, answer, and explanation. The questions of each question type are combined according to the preset test paper layout rules to generate the test paper template; The question type knowledge point distribution characteristics refer to the distribution of knowledge point cognitive levels and difficulty appropriate for different question types, and the question type quantity distribution characteristics refer to the distribution of the number of questions and scores for different question types. These two characteristics are converted into numerical feature vectors for subgraph filtering and question generation. Filtering criteria are set based on the distribution characteristics of knowledge points for each question type; A subgraph filtering algorithm is used to perform conditional filtering in relevant subgraphs to extract knowledge point entities and relationships that meet the filtering conditions for each question type. The selected entities and relationships are combined to generate a reference subgraph for each question type. The reference subgraph contains all the knowledge points applicable to the question type.

2. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 1, characterized in that, By integrating textbook and test question databases from multiple disciplines, and through knowledge extraction and fusion, a multi-disciplinary knowledge graph is constructed, containing knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes, including: Knowledge is extracted from the first textbook data of the first subject to generate a set of entities for the first knowledge point; Identify and extract a set of questions associated with the first knowledge point entity set in the first test question database corresponding to the first subject. By analyzing single knowledge point questions and multi-knowledge point questions, establish semantic relationships, difficulty and cognitive level attributes between different knowledge point entities, and build a first subject knowledge graph corresponding to the first subject. Add the first subject knowledge graph to the multiple subject knowledge graphs.

3. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 2, characterized in that, By analyzing single-knowledge-point questions and multi-knowledge-point questions, semantic relationships, difficulty, and cognitive level attributes between entities of different knowledge points are established, and a first-subject knowledge graph corresponding to the first subject is constructed, including: Extract single-knowledge-point question sets and multi-knowledge-point question sets from the question set; Based on the identification of the execution difficulty and cognitive level attributes of the single knowledge point question set using expert-marked tags, a single knowledge point graph is established; The problem set containing multiple knowledge points is processed to perform semantic relationship analysis between related entities and to identify difficulty and cognitive level attributes based on expert-marked labels, thereby generating a multi-knowledge-point fusion knowledge graph. The single-knowledge-point graph is spliced ​​with the multi-knowledge-point fused knowledge graph to generate the first subject knowledge graph.

4. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 1, characterized in that, The reference sub-graph and quantity requirement code for each question type are input into the conditional text generation model to generate question combinations for each question type, including: For each question type, a cognitive trap map representing cognitive traps is constructed. The cognitive trap map, along with the reference sub-map for each question type and the quantity requirement, are encoded together and then input into the conditional text generation model to generate the corresponding question combination for each question type.

5. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 4, characterized in that, The cognitive trap map includes a first trap map for linking any knowledge point with common learner misunderstandings and misinterpretations, a second trap map for linking a knowledge point with other easily confused knowledge points, and a third trap map for linking a knowledge point with the prerequisites or implicit assumptions that must be relied upon for correct understanding and application.

6. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 5, characterized in that, The conditional text generation model is a large language model, and different question types correspond to different large language models; Based on the question type, the corresponding large language model is invoked to analyze the reference subgraph and quantity requirements for each question type, and intermediate question combinations are generated. Using a large language model based on a cognitive trap graph, cognitive defects are identified in the intermediate question combinations to generate the final question combinations.

7. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 6, characterized in that, When setting up cognitive deficits, cognitive deficits for the options are set up based on the cognitive trap map for multiple-choice questions, and cognitive deficits for the question stems are set up based on the cognitive trap map for fill-in-the-blank and short-answer questions.

8. The method for automatic question generation across multiple disciplines using AI algorithms based on knowledge graphs as described in claim 1, characterized in that, The test paper template is subjected to difficulty prediction and logical consistency checks, and optimization is performed based on the check results, including: Independent difficulty features and overall structural features are extracted from the test paper template. Independent difficulty features include the difficulty of the knowledge points associated with the question text and the text complexity features. Overall structural features include the total number of questions, the distribution of question types, and the dispersion of the distribution of knowledge points. The independent difficulty features and overall structural features are input into the trained test paper difficulty prediction model, and the overall predicted difficulty value of the test paper template is output. For all questions, answers, and explanations in the test paper template, verify whether the core concepts, theorems, and factual relationships conform to the preset knowledge reasoning map, and obtain the logical consistency test results; If the overall predicted difficulty value does not match the difficulty corresponding to the question-generating instruction and / or the logical consistency detection result is inconsistent, a new question will be generated for the abnormal question.

9. A multi-disciplinary automatic question generation system based on knowledge graph AI algorithms, characterized in that: The system is used to implement the knowledge graph-based AI algorithm for multi-disciplinary automatic question generation method as described in any one of claims 1-8, and the system comprises: The subject knowledge graph construction module is used to integrate textbook and test question databases from multiple subjects. Through knowledge extraction and fusion, it constructs multiple subject knowledge graphs that include knowledge point entities, semantic relationships between entities, and difficulty and cognitive level attributes. The question-generating instruction receiving module is used to receive question-generating instructions containing subject type, target knowledge point, question type and difficulty requirements, and extract relevant subgraphs from the subject knowledge graph corresponding to the subject type; The conditional text generation module is used to encode the relevant subgraphs and the question type requirements in the question generation instructions and input them into the conditional text generation model to automatically generate a test paper template, including the question text, answers and explanations; The optimization module is used to predict the difficulty and detect the logical consistency of the test paper template, and optimize it based on the detection results. The conditional text generation module includes: The distribution feature generation unit is used to parse the question type requirements and generate the distribution features of question type knowledge points and the distribution features of question type quantity. The reference subgraph acquisition unit is used to acquire a reference subgraph for each question type from the relevant subgraphs based on the distribution characteristics of the knowledge points of the question type. The quantity requirement acquisition unit is used to acquire the quantity requirement of each question type based on the quantity distribution characteristics of the question types. The question combination generation unit is used to input the reference sub-graph and quantity requirement encoding of each question type into the conditional text generation model to generate a question combination for each question type, including the question, answer and explanation. The test paper template generation unit is used to combine the questions of each question type according to the preset test paper layout rules to generate the test paper template.

Images