Mathematical test question correcting method and computing device

Abstract

The invention provides a mathematical test question correction method and computing equipment, relates to the technical field of computers, and can improve the accuracy of test question correction. The specific scheme is applied to a computing device and comprises the steps that test question information is obtained, and the test question information comprises questions and answer content of the questions; searching knowledge point information related to the question from a mathematical knowledge graph which is a knowledge graph constructed by using mathematical knowledge points in advance; based on the questions, the knowledge point information, answer information of the questions and a multi-modal mathematical test question correction model, the correction result of the answer content is determined, and the answer information comprises standard answers and step scoring standards; the question, the knowledge point information and the answer information comprise at least one of a mathematical picture and a mathematical text, and the mathematical picture is used for indicating at least one of a mathematical formula, a mathematical table and a mathematical graph.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a mathematical test question correcting method and a computing device. BACKGROUND

[0002] In the field of education, test question automatic correction will be the inevitable development direction, which can not only release a large number of repetitive work of teachers, but also is an important data analysis basis for intelligent learning and precise teaching in future education.

[0003] Generally, a deep learning model can be used for test question correction, and a large amount of mathematical knowledge can be learned in advance by the mathematical text data taking the mathematical subject as an example. SUMMARY

[0004] The present application provides a mathematical test question correcting method and a computing device, which can improve the accuracy of test question correction.

[0005] To achieve the above technical purposes, the present application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a mathematical test question correcting method, which is applied to a computing device, and the method comprises: obtaining test question information, wherein the test question information comprises a question and an answer content of the question; retrieving knowledge point information related to the question from a mathematical knowledge graph, wherein the mathematical knowledge graph is a knowledge graph constructed in advance by using mathematical knowledge points; determining a correction result of the answer content based on the question, the knowledge point information, answer information of the question and a multi-modal mathematical test question correcting model, wherein the answer information comprises a standard answer and a step scoring standard; the question, the knowledge point information and the answer information comprise at least one of mathematical pictures and mathematical texts, and the mathematical pictures are used to indicate at least one of mathematical formulas, mathematical tables and mathematical graphs.

[0007] It can be understood that the multi-modal mathematical test question correcting model is a model trained based on multi-modal mathematical data (such as mathematical texts, mathematical graphs, mathematical formulas, mathematical tables, etc.), which can understand the formulas and graphs in the mathematical test question and has certain logical reasoning ability. The computing device determines the knowledge point information related to the question by using the mathematical knowledge graph constructed in advance and containing rich semantic information of mathematical knowledge points; the knowledge point information and the answer information (the standard answer and the step scoring standard) are used as prompt information of the multi-modal mathematical test question correcting model, so as to enhance the understanding of the question and the step analysis ability of the answer content of the question by the multi-modal mathematical test question correcting model, and further improve the correction accuracy of the answer content of the question by the multi-modal mathematical test question correcting model.

[0008] In a possible implementation, the above-mentioned mathematical knowledge points comprise mathematical concepts, mathematical theorems, mathematical formulas and mathematical methods.

[0009] It can be understood that the mathematical knowledge graph comprehensively covers mathematical knowledge, so that the mathematical knowledge graph can accurately retrieve the knowledge point information involved in the question, facilitate the focusing on knowledge points in the correction process of the multi-modal mathematical question correction model, and improve the correction accuracy.

[0010] In a possible implementation, the method further includes: determining a first entity in the target mathematical textbook, wherein the first entity is a mathematical knowledge point; determining a first association relationship between the first entities, wherein the first association relationship includes a dependency relationship and a progression relationship, wherein the dependency relationship is used to indicate a logical association between the first entities; and the progression relationship is used to indicate a learning order between the first entities; and constructing a mathematical knowledge graph based on the first entities and the first association relationship.

[0011] It can be understood that the target mathematical textbook can be multiple mathematical textbooks covering different grades and different versions, the computing device can comprehensively extract the first entities (mathematical concepts, mathematical theorems, mathematical formulas, and mathematical methods, etc.) from the target mathematical textbook, and establish the first association relationship (dependency relationship, progression relationship, etc.) between the first entities, and based on the first entities and the first association relationship between the first entities, a mathematical knowledge graph with rich semantics and high structure is established, thereby improving the retrieval efficiency of the knowledge point information involved in the question.

[0012] In another possible implementation, the determination of the first entity in the target mathematical textbook includes: extracting mathematical knowledge points from the target mathematical textbook; and performing knowledge fusion on a pair of mathematical knowledge points that satisfy a first preset condition in the mathematical knowledge points to obtain the first entity; wherein the first preset condition is that the mathematical knowledge point names are the same or the mathematical knowledge point contents have a similarity greater than a preset similarity threshold.

[0013] It can be understood that different versions of mathematical textbooks have different explanations for the same mathematical knowledge point, so the computing device performs knowledge fusion on the mathematical knowledge points after extracting the mathematical knowledge points from the target mathematical textbook to ensure the consistency of the knowledge. Further, the data format in the mathematical knowledge point includes two parts: the mathematical knowledge point name and the mathematical knowledge point content, for example, the mathematical knowledge point name is Pythagorean theorem, and the mathematical knowledge point content is that the square of the two right-angle sides of a right-angle triangle is equal to the square of the hypotenuse. The computing device quickly identifies a pair of mathematical knowledge points representing the same knowledge based on whether the names of the mathematical knowledge points are the same and whether the similarity of the mathematical knowledge point contents is greater than a preset similarity threshold, thereby quickly implementing knowledge fusion.

[0014] In another possible implementation, the target mathematical teaching material includes a plurality of mathematical teaching materials; and the constructing the mathematical knowledge graph based on the first entity and the first association relationship includes: determining a second entity in the target mathematical teaching material, where the second entity includes a mathematical subject, a teaching material version, an education stage, a teaching material name, and a chapter name; determining a second association relationship between the first entity and the second entity and between the second entities, where the second association relationship includes a containing relationship between the entities; and constructing the mathematical knowledge graph based on the first entity, the second entity, the dependency relationship, the progressive relationship, and the containing relationship.

[0015] It can be understood that the computing device also comprehensively extracts the second entity (a mathematical subject, a teaching material version, an education stage, a teaching material name, and a chapter name, etc.) from the target mathematical teaching material, and establishes the second association relationship (a containing relationship, etc.) between the first entity and the second entity and between the second entities, and establishes a hierarchical structure that is gradually refined from a mathematical subject, a teaching material version, an education stage, a teaching material name, a chapter name, and a mathematical knowledge point based on the first entity, the second entity, the first association relationship, and the second association relationship, so as to facilitate the retrieval of the knowledge point involved in the question and the backtracking of the knowledge point position.

[0016] In another possible implementation, the retrieving the knowledge point information involved in the question from the mathematical knowledge graph includes: determining entity information of the first entity involved in the question, where the entity information includes a mathematical knowledge point name, or the mathematical knowledge point name and at least one second entity having a containing relationship with the first entity involved in the question; retrieving the first entity involved in the question from the mathematical knowledge graph based on the entity information; and determining the mathematical knowledge point content corresponding to the first entity as the knowledge point information involved in the question.

[0017] It can be understood that the computing device can determine the entity information (i.e., the name of the mathematical knowledge point, or the name and at least one of the mathematical subject, the teaching material version, the education stage, the teaching material name, and the chapter name) involved in the question, so as to quickly retrieve the knowledge point information involved in the question from the mathematical knowledge graph based on the entity information.

[0018] In another possible implementation, the scoring standard in the scoring step includes a deduction rule corresponding to the step error type; and the determining the correction result of the answer content based on the question, the knowledge point information, the answer information of the question, and the multi-modal mathematical test question correction model includes: generating a prompt word by using the knowledge point information, the standard answer, and the deduction rule corresponding to the step error type; and inputting the question, the answer content, and the prompt information into the multi-modal mathematical test question correction model to obtain the correction result output by the multi-modal mathematical test question correction model.

[0019] It can be understood that the step scoring standard gives the step error type and the corresponding deduction rule, and the multi-modal mathematical test correction model is based on the correct answer, the step error type and the corresponding deduction rule to deeply analyze the rationality and correctness of the problem solving steps in the answer content, thereby improving the accuracy of the test correction.

[0020] In another possible implementation manner, the correction result includes: step score information, the step score information including a score value, a step function description, a step score reason and a step deduction reason.

[0021] It can be understood that the correction result output by the multi-modal mathematical test correction model clearly presents the score of each step, including the score value, the step function description, the step score reason and the step deduction reason, and the like, thereby facilitating targeted knowledge consolidation and learning improvement.

[0022] In another possible implementation manner, the multi-modal mathematical test correction model is obtained by performing mathematical test correction transfer learning on a multi-modal large model; the multi-modal large model is trained by using a mathematical picture-text pair; and the mathematical picture-text pair includes a mathematical picture and corresponding mathematical text.

[0023] It can be understood that the mathematical test question includes at least one of a mathematical picture and mathematical text, and the mathematical picture is generally presented in the form of a mathematical formula, a mathematical table, a mathematical graph, and the like. If the information represented by the mathematical picture cannot be accurately understood, the accuracy of the mathematical test correction will be seriously affected. Therefore, the electronic device uses the multi-modal mathematical test correction model to correct the mathematical test, and the multi-modal mathematical test correction model is obtained by performing mathematical test correction transfer learning on a multi-modal large model, and the multi-modal large model is trained by using a large amount of mathematical pictures and corresponding mathematical texts. Therefore, the multi-modal mathematical test correction model has good mathematical picture understanding ability, and can ensure the accuracy of the mathematical test correction.

[0024] In a second aspect, the present application provides a computing device. The computing device includes various modules applied to the method of the first aspect or any possible design of the first aspect.

[0025] In a third aspect, the present application provides a computing device including a memory and a processor. The memory and the processor are coupled; the memory is used to store computer program code, and the computer program code includes computer instructions. When the processor executes the computer instructions, the computing device executes the mathematical test correction method of the first aspect and any possible implementation manner thereof.

[0026] In a fourth aspect, the present application provides a computing device including a processor, wherein the processor executes the mathematical test correction method of the first aspect and any possible implementation manner thereof.

[0027] Exemplarily, the computing device can be a server, a tablet, a desktop, a laptop, a notebook, a netbook, and the like.

[0028] In a fifth aspect, a computer readable storage medium is provided, which includes computer instructions. When the computer instructions are run on a computing device, the computing device is caused to perform the mathematical test grading method according to the first aspect and any possible implementation manner thereof.

[0029] In a sixth aspect, a computer program product is provided, which includes computer instructions. When the computer instructions are run on a computing device, the computing device is caused to perform the mathematical test grading method according to the first aspect and any possible implementation manner thereof.

[0030] The detailed description of the second aspect to the sixth aspect and various implementation manners thereof in the present application can refer to the detailed description in the first aspect and various implementation manners thereof, and the beneficial effects of the second aspect to the sixth aspect and various implementation manners thereof can refer to the beneficial effect analysis in the first aspect and various implementation manners thereof, which will not be repeated here.

[0031] These aspects or other aspects of the present application will be more apparent in the following description. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 An implementation environment schematic diagram related to a mathematical test grading method provided by an embodiment of the present application;

[0033] Figure 2 A hardware structure schematic diagram of a computing device provided by an embodiment of the present application;

[0034] Figure 3 A flow schematic diagram of a mathematical test grading method provided by an embodiment of the present application;

[0035] Figure 4 A flow schematic diagram of constructing a mathematical knowledge graph provided by an embodiment of the present application;

[0036] Figure 5 A flow schematic diagram of another mathematical test grading method provided by an embodiment of the present application;

[0037] Figure 6 A structure schematic diagram of a mathematical test grading device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] In the following, in order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, the terms "first", "second", etc. are used to distinguish the same or similar items with basically the same function and role. Those skilled in the art can understand that the terms "first", "second", etc. do not limit the quantity and execution order, and the terms "first", "second", etc. also do not necessarily mean different.

[0039] Meanwhile, in the description of the present application, unless otherwise specified, "multiple" means two or more than two. The words "exemplary" or "for example" are used to mean an example, illustration, or description, and any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the words "exemplary" or "for example" are intended to present the relevant concept in a specific manner for ease of understanding.

[0040] For the convenience of understanding, the following will first introduce the related terms involved in the embodiments of the present application:

[0041] (1) Large Language Model (LLM): It is an artificial intelligence model obtained by training a large amount of text data using a deep learning algorithm, which can learn the patterns and structures of language and thus can understand and generate natural language text.

[0042] (2) Retrieval-augmented Generation (RAG): It combines language models and information retrieval techniques and is one of the current popular large model frontiers. Specifically, when a language model needs to generate text or answer a question, it will first retrieve relevant information from a large set of knowledge base documents, and then use the retrieved information to guide the language model to generate text or answer the question, thereby improving the output accuracy of the language model.

[0043] (3) Optical Character Recognition (OCR): It refers to the process of scanning text on paper by electronic devices (such as scanners or digital cameras), then determining its shape through detection algorithms, and finally translating the shape into computer text using character recognition methods. In the embodiments of the present application, optical character recognition is used to extract the test questions and the corresponding answer results on the mathematical test paper. Paddle OCR is an optical character recognition (OCR) tool library that supports multi-language text detection, recognition, and end-to-end OCR tasks.

[0044] (4)Contrastive Language–Image Pre-training (CLIP): is a model trained on a large number of image-text pairs, which can link language and image information together, and is good at image-text matching tasks, such as determining the matching degree of handwritten pictures and text.

[0045] (5)Multimodal Large Language Model (MLLM): is an artificial intelligence model that integrates the processing capabilities of multiple modal data. It can not only process text data like large language models, but also understand and analyze non-text information such as images, audio, and video. Multimodal large models usually refer to the Transformer model architecture, which captures complex relationships between different modal data through self-attention mechanisms. In the modal pre-training phase, a large amount of multimodal data (such as image-text pairs, videos and subtitles, etc.) is used for training to learn how to extract, fuse and correlate information from different modalities. With strong cross-modal representation learning capabilities, multimodal large models can achieve joint understanding and generation of multimodal content, playing an important role in image description generation, video question answering and other tasks, greatly expanding the application boundaries of artificial intelligence in complex scenarios. In the embodiments of the present application, the multimodal large model is a model trained using mathematical multimodal data (such as mathematical text, mathematical graphics, mathematical formulas, mathematical tables, etc.) in the mathematical field, which can understand formulas and graphics in mathematical test questions and has certain logical reasoning ability. Then, the multimodal large model is subjected to mathematical test question correction transfer learning to obtain a multimodal mathematical test question correction model, making it more suitable for downstream tasks of mathematical test question correction.

[0046] (6)You Only Look Once version 8 (YOLOv8) is a core technology in the field of computer vision, mainly used for identifying object categories and their positions in images or videos. It is widely used in industrial detection, autonomous driving, security monitoring and other fields.

[0047] (7)Named Entity Recognition (NER) is one of the core tasks of Natural Language Processing (NLP), aiming to identify and classify predefined entity categories (such as names, places, times, etc.) from unstructured text.

[0048] The embodiment of the application provides a mathematical test grading method, first, obtaining a question, a solution content of the question and answer information, wherein the answer information comprises a standard answer and a step scoring standard; knowledge point information related to the question is retrieved from a mathematical knowledge graph; wherein the question, the knowledge point information and the answer information comprise at least one of a mathematical picture and a mathematical text, the mathematical picture is used for indicating at least one of a mathematical formula, a mathematical table and a mathematical graph; the knowledge point information and the answer information are used as prompt words of a multi-modal mathematical test grading model, the solution content of the question is graded by the multi-modal mathematical test grading model, and the accuracy of grading the solution content by the multi-modal mathematical test grading model is improved.

[0049] The implementation manners of the embodiment of the application will be described in detail below with reference to the drawings.

[0050] Figure 1 An implementation environment involved in a mathematical test grading method provided by the embodiment of the application is shown in FIG. 1, which comprises a computing device 101. Figure 1

[0051] The computing device 101 can be a smart terminal and can be a server.

[0052] In the embodiment of the application, the computing device 101 is used to obtain test information, wherein the test information comprises a question and a solution content of the question; knowledge point information related to the question is retrieved from a mathematical knowledge graph; based on the question, the knowledge point information, answer information of the question and a multi-modal mathematical test grading model, a grading result of the solution content is determined, wherein the answer information comprises a standard answer and a step scoring standard.

[0053] The mathematical knowledge graph is a knowledge graph constructed in advance by using mathematical knowledge points. Specifically, the mathematical knowledge graph represents mathematical concepts, mathematical theorems, mathematical methods and the like in a graphical or networked form.

[0054] In a possible implementation manner, the computing device 101 comprises a component for information interaction with a user, such as a touch screen, a keyboard, a mouse and the like. The user can send test information to the computing device 101 through the component. The test information can be a mathematical test paper to be graded (such as a photo or a scanned file of the test paper), or a single mathematical test question to be graded. When the test information is a mathematical test paper to be graded, the computing device 101 is further used to split the mathematical test paper into a plurality of mathematical test questions by using an optical character recognition technology, each mathematical test question comprising a question and a solution content of the question.

[0055] ​In a possible implementation, the computing device 101 pre-constructs and stores the mathematical knowledge graph. After obtaining the test question information, the computing device 101 retrieves the knowledge point information involved in the question in the test question information by using the mathematical knowledge graph.

[0056] In a possible implementation, the implementation environment further includes a first server 102, which pre-constructs and stores the mathematical knowledge graph and provides a knowledge retrieval service of the mathematical knowledge graph.

[0057] After obtaining the test question information, the computing device 101 sends the question in the test question information to the first server 102 to obtain the knowledge point information involved in the question returned by the first server 102.

[0058] In a possible implementation, the computing device 101 stores the multi-modal mathematical test question correction model. The computing device 101 inputs the knowledge point information and the answer information as prompt information into the multi-modal mathematical test question correction model, enhances the understanding of the question and the step analysis capability of the answer content of the question by the multi-modal mathematical test question correction model, and then corrects the answer content of the question by using the multi-modal mathematical test question correction model to obtain the correction result.

[0059] In a possible implementation, the implementation environment further includes a second server 103, which stores the multi-modal mathematical test question correction model. The computing device 101 sends the question, the knowledge point information involved in the question, the answer information of the question, and the answer content of the question to the second server 103 to make the second server 103 return the correction result of the answer content, wherein the second server 103 inputs the knowledge point information and the answer information as prompt information into the multi-modal mathematical test question correction model, enhances the understanding of the question and the step analysis capability of the answer content of the question by the multi-modal mathematical test question correction model, and then corrects the answer content of the question by using the multi-modal mathematical test question correction model to obtain the correction result.

[0060] The intelligent terminal can be at least one of a smart phone, a smart watch, a desktop computer, a laptop computer, a virtual reality terminal, an augmented reality terminal, a wireless terminal, and a laptop computer.

[0061] The server, the first server and the second server can be at least one of an independent physical server such as a general server, a graphics processing unit (GPU) server, a data processing unit (DPU) server and the like, or a server cluster or a distributed file system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content distribution network and big data. The embodiments of the present application are not limited thereto.

[0062] Figure 2 A hardware structure schematic diagram of a computing device provided by the embodiments of the present application is shown in FIG. 2. The computing device 200 shown in FIG. 2 can include a processor 201, a memory 202, a communication interface 203 and a bus 204. The processor 201, the memory 202 and the communication interface 203 can be connected through the bus 204. Figure 2 Figure 2 The processor 201 is the control center of the computing device, which can be a general central processing unit such as a CPU, or other general-purpose processors and the like. The general-purpose processor can be a microprocessor or any conventional processor and the like.

[0063] As an example, the processor 201 can include one or more CPUs, such as the CPU 0 and the CPU 1 shown in FIG. 2.

[0064] As an example, the processor 201 can include one or more CPUs, such as the CPU 0 and the CPU 1 shown in FIG. 2. Figure 2

[0065] The memory 202 can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a magnetic disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited thereto.

[0066] ​​In a possible implementation, the memory 202 can exist independently of the processor 201. The memory 202 can be connected to the processor 201 through the bus 204, for storing data, instructions or program codes. When the processor 201 invokes and executes the instructions or program codes stored in the memory 202, the document display method provided by the embodiments of the present application can be implemented.

[0067] In another possible implementation, the memory 202 can also be integrated with the processor 201.

[0068] The communication interface 203 is configured to connect the computing device to other devices through a communication network, which can be an Ethernet, a radio access network (RAN), a wireless local area network (WLAN) or the like. The communication interface 203 can include a receiving unit configured to receive data, and a sending unit configured to send data.

[0069] The bus 204 can be an industry standard architecture (ISA) bus, a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus or the like. The bus can be divided into an address bus, a data bus, a control bus and the like. For the convenience of representation, Figure 2 In the drawings, only one thick line is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0070] It should be noted that, Figure 2 The structure shown in the drawings does not constitute a limitation on the computing device, except Figure 2 The computing device can include more or fewer components than shown in the drawings, or combine certain components, or different component arrangements.

[0071] Figure 3 is a flow diagram of a mathematical test grading method provided by the embodiments of the present application. The mathematical test grading method is executed by the computing device 101 in the implementation environment shown in Figure 1 As shown in the drawings, the method includes the following steps. Figure 3 S301, obtaining test information, wherein the test information includes a question and an answer content of the question.

[0072] In the embodiments of the present application, the question is a mathematical question, which is used to test the mastery of the mathematical concepts, mathematical formulas, mathematical methods and the like of the answerer.

[0073] Specifically, the above-mentioned question mainly refers to a subjective question (i.e., a mathematical question with a solution step), including but not limited to a proof question and a solution question, and the above-mentioned solution content is the answer of the questioner to the question.

[0074] The above-mentioned question contains specific problem statements designed around a specific theme or task. For example, the question of a proof question can include known information or assumptions, conclusions or propositions that need to be proved, etc. The question of a solution question can include problem scenarios, known conditions, and solution goals, etc.

[0075] In one possible implementation, the computing device can obtain the above-mentioned test question information from an electronically connected device (e.g., a teacher's end of remote teaching); or obtain the test question information uploaded by the user locally.

[0076] In one possible implementation, the above-mentioned execution subject can obtain the test question information by the following steps:

[0077] First, obtain the image taken for the test question as the test question image.

[0078] The above-mentioned test question image can be a test paper including at least one subjective question and other types of mathematical questions, or a test paper including only at least one subjective question. The above-mentioned other types of mathematical questions can be multiple-choice questions and fill-in-the-blank questions, etc.

[0079] Second, determine the test question area based on the test question image. The above-mentioned test question area is the area where the mathematical questions and the solution contents of the mathematical questions are concentrated.

[0080] The above-mentioned determination of the test question area based on the test question image can be the determination of the test question area by using a model for detecting a rotating target, such as a refined single stage detector (RSDet model). The RSDet model is suitable for the situation of test paper shooting, which helps to accurately determine the test question area.

[0081] Third, divide the test question area by using a pre-trained target detection model to generate a sub-test question area.

[0082] The above-mentioned target detection model can be a YOLOv8 model focusing on image segmentation. By using the model, the area of each mathematical question and its solution content (hereinafter referred to as a sub-test question area) and the corresponding question type can be determined. The above-mentioned type can be divided into subjective questions and non-subjective questions. The subjective questions can be further divided into solution questions and proof questions, etc. The non-subjective questions can be further divided into multiple-choice questions and fill-in-the-blank questions, etc.

[0083] In the fourth step, the sub-question area corresponding to the subjective question is subjected to character recognition to obtain the question and the corresponding answer content, and further obtain the question information, thereby providing a basis for subsequent correction.

[0084] The character recognition described above can be implemented in various ways. For example, it can be implemented by using a FOTS (fast oriented text spotting) model, or by using a Paddle OCR model.

[0085] In the above manner, the photographed question image can be recognized to accurately obtain the question information.

[0086] In S302, knowledge point information involved by the question is retrieved from a mathematical knowledge graph, wherein the mathematical knowledge graph is a knowledge graph constructed by using mathematical knowledge points.

[0087] As an example, the mathematical knowledge graph can be constructed by using mathematical knowledge points in the contents of mathematical textbooks of different grades and versions. It can realize the structured representation of mathematical knowledge points and the relationship therebetween, and support fast retrieval of mathematical knowledge points.

[0088] The mathematical knowledge points generally include a name and content, the knowledge point information is the content of the mathematical knowledge point involved by the question, and the mathematical knowledge points include mathematical concepts, mathematical theorems, mathematical formulas, and mathematical methods.

[0089] In a possible implementation manner, the mathematical knowledge graph can be a mesh knowledge base constituted by first entities and first association relationships between the first entities. The first entity is a mathematical knowledge point, and the relationship between the first entities includes a dependency relationship and a progressive relationship.

[0090] The dependency relationship is used to indicate the logical association between the first entities. The trigonometric function is a derived output concept based on the right triangle, and the two have logical concepts. The progressive relationship is used to indicate the learning order between the first entities, for example, learning integer four arithmetic operations first, and then learning decimal four arithmetic operations.

[0091] In this case, the retrieval of the knowledge point information involved by the question from the mathematical knowledge graph can include: determining the name of the mathematical knowledge point involved by the question based on the first detection model; and retrieving the knowledge point information involved by the question from the mathematical knowledge graph based on the name of the mathematical knowledge point involved by the question.

[0092] The name of the mathematical knowledge point involved in the question is determined based on the first detection model, including: inputting the question into the first detection model to obtain the name of the mathematical knowledge point output by the first detection model; wherein the first detection model can be trained in advance using a large-scale first training sample, and the first training sample includes mathematical questions and the names of mathematical knowledge points involved.

[0093] In a possible implementation, the above-mentioned mathematical knowledge graph can be a network knowledge base composed of first entities, second entities, first association relationships between the first entities, second association relationships between the first entities and the second entities, and second association relationships between the second entities. The second entities include: mathematical disciplines, textbook versions, education stages, textbook names, chapter names, etc. The second association relationship includes a containing relationship. The containing relationship, for example, a chapter contains mathematical knowledge points, a textbook contains chapter names, etc.

[0094] In this case, the above-mentioned retrieval of the knowledge point information involved in the question from the mathematical knowledge graph can include: determining the name of the mathematical knowledge point involved in the question and the association information based on the second detection model, the association information being used to indicate at least one of the mathematical disciplines, the textbook versions, the education stages, the textbook names, and the chapter names corresponding to the mathematical knowledge point involved in the question, and retrieving the knowledge point information involved in the question from the mathematical knowledge graph based on the name of the mathematical knowledge point and the association information.

[0095] It can be understood that the more types of association information, the higher the speed and accuracy of subsequent retrieval.

[0096] The above-mentioned determination of the name of the mathematical knowledge point involved in the question and the association information based on the second detection model includes: inputting the question into the second detection model to obtain the name of the mathematical knowledge point and the association information output by the second detection model. The above-mentioned second detection model is trained in advance using a large-scale second training sample, and the second training sample includes questions, the names of mathematical knowledge points involved in the questions, the mathematical disciplines corresponding to the mathematical knowledge points, the textbook versions, the education stages, the textbook names, and the chapter names.

[0097] The first detection model and the second detection model can be a large language model (LLM), or other types of models, which are not limited here.

[0098] The above-mentioned retrieval of the knowledge point information involved in the question from the mathematical knowledge graph can be implemented using the RAG technology.

[0099] In the embodiments of the present application, the name of the mathematical knowledge point involved in the question is determined by using the first detection model, or the name of the mathematical knowledge point involved in the question and the associated information are determined by using the second detection model, so as to narrow the search range, and then the mathematical knowledge graph is searched to quickly obtain the knowledge point information.

[0100] S303, based on the question, the knowledge point information, the answer information of the question and the multi-modal mathematical test correction model, determining the correction result of the answer content, wherein the answer information includes the standard answer and the step scoring standard; the question, the knowledge point information and the answer information include at least one of mathematical picture and mathematical text, and the mathematical picture is used to indicate at least one of mathematical formula, mathematical table and mathematical graph.

[0101] In a possible implementation, the answer information of the question can include the standard problem solving process of the subjective question, the key step scoring point, the step scoring standard and the like, so that the multi-modal mathematical test correction model scores the problem solving step process of the answer content.

[0102] The progress of optical character recognition (OCR) technology enables electronic devices to efficiently convert paper text into computer-readable text, promoting the process of test correction automation. On this basis, a variety of test automatic correction methods have been derived, for example:

[0103] Relying on a large exercise library to realize test automatic correction: this method constructs a large-scale exercise library, which includes multiple test questions and correct answers corresponding to the test questions. When correcting the test question to be corrected, search for similar test questions in the exercise library. If a test question is found, correct the test question to be corrected according to the correct answer of the searched test question in the exercise library. If not, manual correction is required.

[0104] However, although this method realizes test automatic correction, it relies too much on the exercise library, and the correction can only give the result, but cannot provide detailed step analysis. Subjective questions focus on logical reasoning and step analysis, so this method is not suitable for subjective question correction.

[0105] Using a large language model to realize test automatic correction: with the development of artificial intelligence technology, especially the application of deep learning in natural language processing, large language models (LLM) have shown their powerful capabilities in many tasks. Through pre-training of large-scale domain text data, it not only masters rich domain knowledge, but also has certain reasoning, which provides the possibility of automatically identifying, understanding and correcting test content in the education field.

[0106] However, large language models (LLMs) are trained using pure text and lack the ability to understand and analyze related reasoning for formulas and graphics. Mathematical test questions, especially subjective questions, contain a large number of formulas and graphics, and the correction focuses on reasoning ability and step analysis process. Therefore, large language models are not suitable for correcting subjective questions.

[0107] To address the above situation, the embodiments of the present application introduce a multi-modal mathematical test correction model. The multi-modal mathematical test correction model can effectively compensate for this deficiency and meet the demand for mathematical subjective question correction, thanks to its advanced architecture and training mechanism.

[0108] Specifically, the multi-modal mathematical test correction model is obtained by transferring learning on a multi-modal large model for mathematical test correction. The multi-modal large model is trained using large-scale mathematical pictures (including mathematical graphics, mathematical formulas, mathematical tables, etc.) and their corresponding mathematical texts.

[0109] During training, the multi-modal large model can use a Transformer architecture as the core, use a convolutional neural network (CNN) to extract image features such as formula structures and graphic outlines from mathematical pictures, and use word embedding technology to encode mathematical texts to obtain text features. Then, the self-attention mechanism can capture the complex relationship between image features and text features, thereby realizing cross-modal semantic learning of mathematical pictures and mathematical texts and improving cross-modal semantic understanding ability. Since the multi-modal large model has good cross-modal semantic understanding ability, such as understanding the internal relationship between function images and corresponding formulas and geometric figures and geometric theorems, the multi-modal mathematical test correction model obtained by transferring learning on the multi-modal large model can be applied to the correction of subjective questions.

[0110] In one possible implementation, the above-mentioned multi-modal mathematical test correction model based on the question, knowledge point information, answer information of the question, and the multi-modal mathematical test correction model determines the correction result of the answer content, including:

[0111] The question, knowledge point information, answer information of the question, and the answer content are input into the multi-modal mathematical test correction model to obtain the correction result of the answer content output by the multi-modal mathematical test correction model.

[0112] During specific correction, the question, answer content, knowledge point information related to the question, and answer information of the question are input into the multi-modal mathematical test correction model to enhance the understanding of the multi-modal mathematical test correction model for the question and the step analysis ability of the answer content for the question, thereby improving the correction accuracy of the answer content.

[0113] The step analysis capability mainly reflects whether the analysis step follows mathematical logic, etc. For example, in processing a geometry proof question, the multi-modal mathematical test correction model can analyze the matching degree of the line segment and angle relationship in the figure and the theorem used in the proof step, so as to judge the correctness of the final answer, and the logical rationality and logical coherence of the proof step.

[0114] In addition, the multi-modal mathematical test correction model has excellent generalization ability and can break through the limitations of textbook knowledge to accurately identify and recognize correct problem solving methods other than standard answers.

[0115] In actual application, when using innovative ideas outside the textbook to solve problems, the multi-modal mathematical test correction model can analyze the rationality and correctness of the problem solving steps in depth by virtue of cross-modal learning and logical reasoning ability. Once it is determined that the solution is effective, the answer will be added to the innovative solution label. This label can help quickly locate the thinking highlights of the answerer, and through certain encouragement measures, it can fully stimulate the enthusiasm of the answerer to explore multiple problem solving paths, and help cultivate innovative thinking and mathematical literacy.

[0116] In one possible implementation, the multi-modal mathematical test correction model uses the step score standard in the answer information as the core to evaluate the step score of the subjective question.

[0117] For example, the step score standard will indicate the key steps in the problem solving process. If the answer content completely contains all the key steps, and each step is logically correct and reasonably inferred, full marks for the corresponding step will be obtained.

[0118] Further, the step score standard includes the deduction rules corresponding to the step error types.

[0119] For example, the logical relationship between steps is an important scoring point. The multi-modal mathematical test correction model will analyze whether the answer steps in the answer content are developed in a reasonable mathematical logical order, whether the derivation from the known conditions to the conclusion is smooth. If there is a logical jump or step reversal, the corresponding score will be deducted according to the degree of influence. For example, in a proof question, using the conclusion to derive the conditions, this logical error will result in serious deduction.

[0120] For steps involving formulas and theorems, the multi-modal mathematical test correction model will check whether the formulas and theorems used by the answerer match the question context, whether the formula writing is correct, and whether the parameter substitution is accurate. If the formula is used incorrectly, the step will not score, regardless of the subsequent steps. If there is only a small error in writing, such as missing parentheses, appropriate deductions will be made.

[0121] For the case that the answer step part is partially correct, the multi-modal mathematics test grading model scores according to the matching degree of the correct part in the answer content with the key steps and the contribution degree to the final answer. For example, in a complex application question, the quantity relationship in the question is correctly analyzed, but a calculation error occurs when the equation is solved, and the step of analyzing the quantity relationship will obtain a certain score according to its importance.

[0122] Finally, the grading result output by the multi-modal mathematics test grading model clearly presents the scoring situation of each step, including step description, score, and deduction reason information. At the same time, the step scoring situation is also associated with the book source of the mathematics knowledge point, which is convenient for understanding the position and learning requirements of the mathematics knowledge point in the textbook, and is also helpful for targeted knowledge consolidation and learning improvement.

[0123] It should be noted that the artificial auditing link is introduced in the embodiments of the present application, which can be selectively sampled for auditing to ensure the accuracy of the grading result. If errors or inaccuracies are found, the artificial modification and feedback are provided, and then the multi-modal mathematics test grading model is re-generated to obtain more accurate grading results. In this way, the multi-modal mathematics test grading model can be continuously optimized to improve the grading accuracy.

[0124] The technical scheme provided in the embodiments of the present application accurately finds the knowledge point information involved in the question from the mathematics knowledge graph; and the knowledge point information and the answer of the question are used as the prompt information of the multi-modal mathematics test grading model to enhance the understanding of the question and the step analysis ability of the answer content of the multi-modal mathematics test grading model. On this basis, the multi-modal mathematics test grading model combines the mathematics knowledge learned before to analyze whether the answer steps are reasonable and whether they follow the mathematical logic, to realize the step scoring of the answer content of the question and improve the accuracy and reliability of the grading.

[0125] It should be noted that the embodiments of the present application focus on the grading of subjective questions, but the present application can also realize the grading of non-subjective questions such as multiple-choice questions and fill-in-the-blank questions, and in this case, the multi-modal mathematics test grading model is not needed.

[0126] Specifically, in response to the question being a multiple-choice question, the answer content in the question information is matched with the standard answer, and if the matching is successful, the answer is correct, and if the matching is unsuccessful, the answer is incorrect.

[0127] In response to the question being a fill-in-the-blank question, the answer content in the question information is matched with the standard answer, and if the content is consistent, the answer is correct, and if the content is inconsistent, the Clip model is used for secondary verification to avoid misjudgment caused by low image resolution and non-standard handwriting font.

[0128] The process of using the Clip model for secondary verification is: comparing the handwritten picture corresponding to the answer content with the answer content (i.e., the recognition result of the text in the handwritten picture) and the standard answer respectively, determining the one closest in semantics to the handwritten picture, responding to the handwritten picture that is closer in semantics to the answer content, and the answer is wrong; responding to the handwritten picture that is closer in semantics to the standard answer, and the answer is correct. Before returning the grading result indicating that the answer is correct, a prompt tag can be added to prompt the standard writing.

[0129] When the test questions are multiple-choice questions and fill-in-the-blank questions, the acquisition process of the corresponding questions and answer contents is consistent with the above, and will not be repeated here. The most suitable grading method is selected according to different types of test questions, such as using different methods to process multiple-choice questions, fill-in-the-blank questions and subjective questions, thereby improving the accuracy of grading.

[0130] The technical scheme provided by the embodiments of the present application uses the multi-modal mathematical model and the answer information (including the standard answer and the step scoring standard) of the question as the prompt words of the multi-modal mathematical question grading model, enhances the understanding of the question by the multi-modal mathematical question grading model and the step analysis ability of the answer content of the question, and further improves the grading accuracy of the multi-modal mathematical question grading model on the answer content of the question.

[0131] Figure 4 is a flowchart of constructing a mathematical knowledge graph provided by an embodiment of the present application, as shown in Figure 4 , the flowchart includes:

[0132] S401, determining a first entity and a second entity in a target mathematical textbook, wherein the first entity is a mathematical knowledge point, and the mathematical knowledge point includes a mathematical concept, a mathematical theorem, a mathematical formula and a mathematical method; the second entity includes a mathematical subject, a textbook version, an education stage, a textbook name and a chapter name.

[0133] The above target mathematical textbook includes a plurality of mathematical textbooks, and the learning textbook includes but is not limited to a mathematical textbook, a tutorial book, a popular science article, an academic paper and the like.

[0134] In the embodiments of the present application, the target mathematical textbook is obtained after preprocessing the original mathematical textbook.

[0135] As an example, the process of obtaining the target mathematical textbook includes:

[0136] First, according to the current K12 (primary school to high school) mathematics curriculum standard, the mainstream mathematical textbooks covering the primary, junior high school and high school stages are comprehensively collected, including different versions of mathematical textbooks. At the same time, the authoritative mathematical education research data is collected.

[0137] Second, the collected mathematical textbooks are digitally processed.

[0138] Specifically, the paper content is converted into electronic text format by scanning, text entry, etc. Among them, for the formulas, graphics and other non-text content in the mathematics teaching materials, special mathematical formula recognition tools (such as Mathpix) and graphic recognition technology are used for extraction and conversion into text content to ensure the integrity of the data.

[0139] Third, the collected electronic text is processed by word segmentation, and a special word segmentation tool for the mathematical field is used to accurately divide mathematical terms, symbols and ordinary characters. Remove noise information in the text, such as extra spaces, line breaks, annotation content, etc., and unify the text format.

[0140] It can be understood that a mathematical term or symbol (such as a formula, a picture) is a word segmentation, so the word segmentation processing will not damage the integrity of the information.

[0141] Fourth, the mathematical formulas and graphics are processed for format specification and standardization to ensure the consistency of the expression form. For example, the formulas in the teaching materials are uniformly converted into structured expressions (such as AST abstract syntax tree), and the graphics in the geometric knowledge points are stored in SVG structure, and the key geometric properties (such as triangle edge length, angle, coordinate system parameter) are labeled. Through a combination of manual checking and automated checking, check for errors and contradictory information in the data, such as incorrect formulas, mismatched graphics and text descriptions, and make corrections.

[0142] Fifth, the mathematical teaching materials processed by the above are determined as the target mathematical teaching materials.

[0143] In one possible implementation, the first entity in the above determination of the target mathematical teaching material includes: extracting mathematical knowledge points from the target mathematical teaching material; performing knowledge fusion on the mathematical knowledge points that meet the first preset condition in the mathematical knowledge points to obtain the first entity; wherein the first preset condition is that the mathematical knowledge point names are the same, or the mathematical knowledge point contents have a similarity greater than a preset similarity threshold.

[0144] The above extracting mathematical knowledge points from the target mathematical teaching material includes:

[0145] Using named entity recognition (NER) technology, the mathematical knowledge points, including mathematical concepts (such as functions, triangles), mathematical theorems (such as Pythagorean theorem, cosine theorem), mathematical formulas (such as quadratic function root formula), mathematical methods (such as matching method, substitution method), etc. are identified from the target mathematical teaching material.

[0146] In a possible implementation, the knowledge fusion of the pair of mathematical knowledge points in the mathematical knowledge points satisfying the first preset condition is performed to obtain the first entity, including: performing knowledge fusion on the pair of mathematical knowledge points in the extracted mathematical knowledge points satisfying the first preset condition to obtain the first entity; and the first preset condition is that the names are the same or the content similarity is greater than a preset similarity threshold.

[0147] It can be understood that the first entity represents a mathematical knowledge point. Since different versions of mathematical textbooks have different explanations for the same mathematical knowledge point, after extracting the mathematical knowledge point from the target mathematical textbook, the mathematical knowledge point needs to be knowledge fused (i.e., merged and disambiguated) to ensure the consistency of the knowledge. Considering that a mathematical knowledge point generally includes a name and content, for example, the name of the mathematical knowledge point is Pythagorean theorem, and the content of the mathematical knowledge point is that the square of the two right-angle sides of a right-angle triangle is equal to the square of the hypotenuse. Therefore, by judging whether the names of the mathematical knowledge points are the same and whether the content similarity (for example, the cosine similarity based on text semantics) is greater than a preset similarity threshold, a pair of mathematical knowledge points representing the same knowledge is quickly identified, thereby quickly realizing knowledge fusion.

[0148] In a possible implementation, the determination of the second entity in the target mathematical textbook includes:

[0149] The second entity is extracted from the target mathematical textbook by using a named entity recognition (NER) technology; and the extracted second entity includes an education stage, a textbook name, a chapter name, and the like.

[0150] It can be understood that the second entity can reflect the location of the mathematical knowledge point, and the introduction of the second entity when constructing the mathematical knowledge graph facilitates the use of the location to speed up the retrieval of the mathematical knowledge point involved in the question.

[0151] In addition, after the knowledge point information involved in the question is retrieved, the knowledge point location information can be integrated into the knowledge point information, so as to facilitate subsequent knowledge review.

[0152] S402, determine a first association relationship between the first entities, a second association relationship between the first entities and the second entities, and a second association relationship between the second entities, wherein the first association relationship includes a dependency relationship and a progression relationship; and the second association relationship includes a containing relationship.

[0153] It can be understood that in the relationship rules of the mathematical field, the above dependency relationship is used to indicate the logical association between the first entities, the trigonometric function is an output concept derived on the basis of a right triangle, and the two have logical concepts. The above progressive relationship is used to indicate the learning order between the first entities, such as learning integer four arithmetic operations first and then learning decimal four arithmetic operations. The containment relationship is used to indicate that one entity belongs to the range of another entity, such as a mathematical knowledge point belonging to a certain education stage, a mathematical knowledge point belonging to a chapter of a textbook, a textbook belonging to a certain education stage, and the like.

[0154] In a possible implementation, the above determining the first association relationship between the first entities, the second association relationship between the first entities and the second entities, and the second association relationship between the second entities includes:

[0155] The relationship extraction model is used to determine the first association relationship between the first entities, the second association relationship between the first entities and the second entities, and the second association relationship between the second entities.

[0156] The relationship extraction model can be roughly divided into a rule-based relationship extraction model, a machine learning-based relationship extraction model, and a deep learning-based relationship extraction model (such as a Transformer-based relationship extraction model), and these models can automatically learn and extract complex relationships from text.

[0157] It should be noted that the type and network architecture of the relationship extraction model are not limited here.

[0158] S403, generating a hierarchical structure about the first entities and the second entities according to the containment relationship.

[0159] Based on the first entities, the second entities, and the above containment relationship, a hierarchical structure can be established, which is gradually refined from a mathematical subject, a textbook version, an education stage, a textbook name, a chapter name, and a mathematical knowledge point. The above education stage refers to a grade.

[0160] It can be understood that the hierarchical structure only needs to comply with the containment relationship and have obvious levels.

[0161] S404, constructing a mathematical knowledge graph based on the hierarchical structure, the dependency relationship, and the progressive relationship.

[0162] In a possible implementation, the above constructing a mathematical knowledge graph based on the hierarchical structure, the dependency relationship, and the progressive relationship includes:

[0163] The above constructing an initial mathematical knowledge graph based on the hierarchical structure, the dependency relationship, and the progressive relationship;

[0164] The initial mathematical knowledge graph is converted into a standard format of a knowledge graph, and a mathematical knowledge graph is obtained.

[0165] The standard format can be RDF (Resource Description Framework) or OWL (Web Ontology Language). The format conversion ensures the consistency and interoperability of knowledge, and a suitable graph database (such as Neo4j) is selected to store the knowledge graph. The efficient graph query and analysis function of the graph database is used to facilitate the subsequent retrieval and operation of the mathematical knowledge graph.

[0166] It should be noted that, considering the development and changes of mathematical education, the mathematical knowledge graph needs to be regularly updated with the latest teaching resources and research results to supplement new knowledge points or correct existing content in a timely manner. This includes the release of new textbooks, adjustment of curriculum standards, etc.

[0167] The embodiments of the present application create a highly structured and semantically rich mathematical knowledge graph of mathematical teaching materials through the above steps. The mathematical knowledge graph can be used to quickly and accurately extract the knowledge point information involved in the question.

[0168] Figure 5 is a flowchart of a mathematical test question correction method provided by the embodiments of the present application, as shown in Figure 5 , the method comprises:

[0169] S501, obtaining test question information, wherein the test question information comprises a question and an answer to the question.

[0170] S502, retrieving knowledge point information involved in the question from the mathematical knowledge graph.

[0171] S503, determining a correction result of the answer based on the question, the knowledge point information, the answer information, and a multi-modal mathematical test question correction model, wherein the answer information includes a standard answer and a step scoring standard; the question, the knowledge point information, and the answer information include at least one of mathematical pictures and mathematical texts, and the mathematical pictures are used to indicate at least one of mathematical formulas, mathematical tables, and mathematical graphs.

[0172] The steps S501 to S503 have been described in detail in the steps S301 to S303 of the embodiments as shown in Figure 3 , and will not be repeated here.

[0173] S504, generating a learning situation report based on the correction result.

[0174] The test correction results of the answerer in the current and past period of time are summarized to generate a detailed learning situation report, which reflects the score of a single question, the mastery of knowledge points (determined by the score of related test questions), and common error types, and then provides personalized learning suggestions for the answerer, including targeted review plans, recommended practice questions, and learning resource links, to help the answerer find and fill gaps and improve learning efficiency.

[0175] Meanwhile, the test correction results of multiple answerers in the current and past period of time can also be summarized to generate a learning situation report for multiple answerers as a whole, to analyze common problems and learning trends, and help adjust teaching strategies and optimize teaching content.

[0176] In a possible implementation, after generating the learning situation report, the method further includes collecting the wrong questions of the answerer, establishing a wrong question library, and generating similar practice questions, and outputting the practice questions to help the answerer deepen the understanding of the wrong questions and avoid the repeated occurrence of similar errors, thereby consolidating the review effect. This mechanism not only reduces the teaching burden, but also promotes the self-directed learning of the answerer and improves the scientificity of teaching decisions.

[0177] In a possible implementation, the above generating a learning situation report based on the correction result includes determining weak links and generating a learning situation report for the answerer based on the weak links.

[0178] The weak links can be determined according to error types, mastery of mathematical knowledge points, and completeness of answer steps.

[0179] For example, the weak links determined according to error types can be, for example, conceptual errors (deviation in understanding of theorems and formulas), calculation errors, logical errors (unreasonable proof steps), and graphic understanding errors. If the answerer frequently makes the same type of error, such as repeatedly losing points due to conceptual understanding errors, it can be determined that this type is a weak link.

[0180] The weak links determined according to the mastery of mathematical knowledge points can be, for example, associating the questions with mathematical knowledge points and counting the scores of the answerer on each mathematical knowledge point. For example, if the answerer repeatedly makes mistakes on a specific mathematical knowledge point, it indicates that the answerer has deficiencies in that mathematical knowledge point, which is a weak link.

[0181] The weak links determined according to the completeness of answer steps can be, for example, analyzing the completeness and standardization of the answer steps of the answerer. For example, if the answerer often omits key steps in a geometric proof question, resulting in the loss of step points, it indicates that the answerer has deficiencies in logical expression and problem-solving standards.

[0182] The above learning situation report for the answerer generated based on the weak links can include data visualization display and personalized improvement measures.

[0183] Among them, the data visualization display can be the application of data visualization technology in the form of combination of charts and texts to list the weak mathematical knowledge points and error types of the answerers in detail. Not only the error times are presented, but also the error proportion is calculated to enable the answerers to more intuitively understand the severity of their weak links. For example: in the knowledge point of trigonometric function calculation, the conceptual error occurs 3 times, accounting for 60% of the total errors of this knowledge point; the calculation error occurs 2 times, accounting for 40%; in the answer to the solid geometry proof, 8 points are lost due to incomplete logical steps, accounting for 40% of the total score of the question.

[0184] The personalized improvement measures can be: for conceptual errors, in addition to recommending the key contents of the corresponding chapters in the teaching materials for relearning, online course resources and case analysis can also be provided to help the answerers understand the concepts in depth.

[0185] For calculation errors, according to the calculation error types of the answerers (such as decimal arithmetic error, root simplification error, equation solving error, etc.), layered calculation exercises are customized from basic exercises to expansion and improvement to gradually improve the calculation ability.

[0186] For the answerers who are insufficient in logical expression and problem solving, complete problem solving step templates of typical examples are provided, and detailed logical derivation explanation videos are recorded to enable the answerers to perform imitation training, and one-on-one review and guidance are arranged.

[0187] Considering the weak links, learning time, learning ability and learning habits of the answerers, a scientific and reasonable learning plan is made to target the leakage and make up for the lack. The learning plan not only specifies the learning content and the number of exercises, but also sets learning goals and evaluation standards. For example: “from Monday to Wednesday, review the concept of trigonometric functions, complete 10 basic exercises, and the goal is to achieve a correct rate of more than 80%; from Thursday to Friday, perform logical step imitation training for solid geometry proof questions, complete 3 questions, and require complete steps and clear logic; comprehensive test on the weekend to test the learning effect.

[0188] In one possible implementation, the above-mentioned generation of a learning situation report based on weak links for a plurality of answerers as a whole can include data visualization display, personalized improvement measures, etc.

[0189] The above-mentioned generation of a learning situation report based on correction results includes: determining weak links; generating a learning situation report for a plurality of answerers as a whole based on weak links.

[0190] The above-mentioned generated learning situation report for a plurality of answerers as a whole can include data overall visualization display, error question bank and expansion exercises, teaching strategy adjustment suggestions, etc.

[0191] Overall data visualization can involve statistically analyzing the percentage of various types of errors made by multiple respondents and presenting it in intuitive charts such as pie charts and bar charts. For example, conceptual errors might account for 30%, calculation errors 40%, logical errors 20%, and other errors 10%. Calculating the average score rate for each knowledge point across multiple respondents and creating a line graph of the score rate clearly demonstrates the mastery of each mathematical knowledge point. Special attention should be paid to knowledge points with a score rate below 60%, such as the average score rate for solving the general term formula of a sequence being only 55%, which requires focused analysis and teaching intervention. A score distribution chart can also be created to show the overall performance, including statistics such as the highest score, lowest score, average score, and median, to understand the overall dispersion of scores.

[0192] The error database and extended exercises can aggregate incorrect answers from multiple respondents, categorizing them multidimensionally according to mathematical knowledge points, error types, and difficulty levels to form a personalized error database. Utilizing the natural language processing and knowledge reasoning capabilities of a large-scale model, it generates similar variant questions based on the questions in the error database. These variant questions not only change the question format (e.g., changing coefficients, adding parameters, adjusting the problem angle), but also innovate in areas such as the comprehensive application of knowledge points and the expansion of problem-solving approaches, meeting the learning needs of different respondents. Simultaneously, detailed answers and explanations are provided for each variant question to facilitate self-study.

[0193] Suggestions for adjusting teaching strategies can be based on the results of student learning analysis, combined with educational and teaching theories and practical experience, to provide specific and actionable suggestions. For example, for situations with many calculation errors, it is recommended to add a special calculation training session in class, using timed exercises, group competitions, etc., to improve students' interest and ability in calculation. For knowledge points with low scoring rates, a four-step teaching method of "explain-practice-evaluate-extend" can be used, namely, key explanation, specific practice, timely evaluation, and extension. Group cooperative inquiry learning can also be used to promote communication and understanding among students.

[0194] S505, Output the grading results and learning progress report.

[0195] In the daily grading of math exams, the entire process from uploading the exam paper to generating the final report has been automated, reducing the workload of teachers and providing personalized learning guidance for test takers.

[0196] The technical solution provided in this application can generate personalized reports with real-time feedback for both the test taker and the teacher based on the grading results, helping the test taker to identify weaknesses in their learning and assisting the teacher in adjusting their teaching strategies.

[0197] Figure 6 A schematic diagram of a mathematical test grading device provided in an embodiment of this application. See also... Figure 6The device is applied to a computing device and includes an acquisition module 601, a retrieval module 602, and a determination module 603.

[0198] The acquisition module 601 is configured to acquire test question information, where the test question information includes a question and an answer to the question.

[0199] The retrieval module 602 is configured to retrieve knowledge point information related to the question from a mathematical knowledge graph, where the mathematical knowledge graph is a knowledge graph constructed in advance using mathematical knowledge points.

[0200] The determination module 603 is configured to determine a correction result of the answer based on the question, the knowledge point information, answer information of the question, and a multi-modal mathematical test question correction model, where the answer information includes a standard answer and a step scoring standard; and the question, the knowledge point information, and the answer information include at least one of a mathematical picture and a mathematical text, and the mathematical picture is used to indicate at least one of a mathematical formula, a mathematical table, and a mathematical graph.

[0201] The technical scheme provided by the embodiments of the present application uses the multi-modal mathematical model and the answer information of the question (including the standard answer and the step scoring standard) as the prompt words of the multi-modal mathematical test question correction model, enhances the understanding of the question by the multi-modal mathematical test question correction model and the step analysis capability of the answer to the question, and thus improves the correction accuracy of the multi-modal mathematical test question correction model for the answer to the question.

[0202] In a possible implementation, the mathematical knowledge points include mathematical concepts, mathematical theorems, mathematical formulas, and mathematical methods.

[0203] In a possible implementation, the device further includes a construction module, which includes a first determination unit, a second determination unit, and a construction unit, where

[0204] The first determination unit is configured to determine first entities in a target mathematical textbook, where the first entities are mathematical knowledge points.

[0205] The second determination unit is configured to determine first association relationships between the first entities, where the first association relationships include dependency relationships and progressive relationships, where the dependency relationships are used to indicate logical associations between the first entities, and the progressive relationships are used to indicate learning orders between the first entities.

[0206] The construction unit is configured to construct the mathematical knowledge graph based on the first entities and the first association relationships.

[0207] In a possible implementation, the first determination unit includes an extraction subunit and a knowledge fusion subunit, where

[0208] The extraction subunit is configured to extract mathematical knowledge points from the target mathematical textbook.

[0209] The knowledge fusion subunit is configured to fuse the pairs of mathematical knowledge points that meet a first preset condition in the mathematical knowledge points to obtain a first entity. The first preset condition is that the mathematical knowledge points have the same name or the content similarity is greater than a preset similarity threshold.

[0210] In a possible implementation, the target mathematical teaching material includes a plurality of mathematical teaching materials; and the construction unit includes a first determination subunit, a second determination subunit, and a construction subunit, wherein

[0211] The first determination subunit is configured to determine a second entity in the target mathematical teaching material. The second entity includes a mathematical subject, a teaching material version, an education stage, a teaching material name, and a chapter name.

[0212] The second determination subunit is configured to determine a second association relationship between the first entity and the second entity and between the second entities. The second association relationship includes a containing relationship between the entities.

[0213] The construction subunit is configured to construct a mathematical knowledge graph based on the first entity, the second entity, the dependency relationship, the progressive relationship, and the containing relationship.

[0214] In a possible implementation, the retrieval module 602 includes a third determination unit, a retrieval unit, and a fourth determination unit, wherein

[0215] The third determination unit is configured to determine entity information of the first entity involved in the question. The entity information includes a mathematical knowledge point name or a mathematical knowledge point name and at least one second entity having a containing relationship with the first entity involved in the question.

[0216] The retrieval unit is configured to retrieve the first entity involved in the question from the mathematical knowledge graph based on the entity information.

[0217] The fourth determination unit is configured to determine the mathematical knowledge point content corresponding to the first entity as the knowledge point information involved in the question.

[0218] In a possible implementation, the scoring standard includes a step error type and a corresponding deduction rule; and the determination module includes a prompt unit and a fifth determination unit, wherein

[0219] The prompt unit is configured to generate prompt information by using the knowledge point information, the standard answer, and the deduction rule corresponding to the step error type.

[0220] The fifth determination unit is configured to input the question, the answer content, and the prompt information into a multi-modal mathematical test item correction model to obtain a correction result output by the multi-modal mathematical test item correction model.

[0221] In a possible implementation, the correction result includes step score information, and the step score information includes a score value, a step function description, a step score reason, and a step deduction reason.

[0222] In a possible implementation, the multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model; the multi-modal large model is trained using mathematical image-text pairs; and the mathematical image-text pairs include mathematical pictures and corresponding mathematical texts.

[0223] The embodiments of the present application further provide a computing device, including a processor and a memory, which are coupled. The memory is configured to store computer program instructions, and the processor is configured to invoke the computer program instructions in the memory to execute the mathematical test paper correction method shown in the above embodiments.

[0224] The embodiments of the present application further provide a computer readable storage medium, which stores computer program instructions, and the computer program instructions are configured to enable a computing device to execute the mathematical test paper correction method shown in the above embodiments.

[0225] The embodiments of the present application further provide a computer program product, including computer program instructions, which, when executed on a computing device, enable the computing device to execute the mathematical test paper correction method shown in the above embodiments.

[0226] The computing device, the computer readable storage medium, or the computer program product provided by the embodiments of the present application are all configured to execute the corresponding method provided above. Therefore, the beneficial effects they can achieve can refer to the beneficial effects in the corresponding method provided above, which will not be described here.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is taken as an example for illustration. In actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device (such as a computing device) is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device (such as a computing device) and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0228] In several embodiments provided in the present application, it should be understood that the disclosed system, apparatus (such as a computing device) and method can be implemented in other manners. For example, the apparatus (such as a computing device) embodiments described above are merely illustrative. For example, the division of the modules or units can be different, and each can contain a plurality of sub-units. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0229] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place or distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0230] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0231] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method of each embodiment of the present application. The aforementioned storage medium includes: a flash memory, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0232] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of grading a mathematical test question, characterized by, The method comprises: obtaining test question information, wherein the test question information comprises a question and an answer to the question; retrieving knowledge point information related to the question from a mathematical knowledge graph, wherein the mathematical knowledge graph is a knowledge graph constructed in advance using mathematical knowledge points; determining a correction result of the answer based on the question, the knowledge point information, and answer information of the question using a multi-modal mathematical test question correction model, wherein the answer information comprises a standard answer and a step scoring standard; and the question, the knowledge point information, and the answer information comprise at least one of a mathematical picture and a mathematical text, and the mathematical picture is used to indicate at least one of a mathematical formula, a mathematical table, and a mathematical graph.

2. The method of claim 1, wherein, The mathematical knowledge points comprise mathematical concepts, mathematical theorems, mathematical formulas, and mathematical methods.

3. The method of claim 2, wherein, The method further comprises: determining first entities in a target mathematical textbook, wherein the first entities are mathematical knowledge points; determining first association relationships between the first entities, wherein the first association relationships comprise dependency relationships and progressive relationships, wherein the dependency relationships are used to indicate logical associations between the first entities, and the progressive relationships are used to indicate learning orders of the first entities; constructing the mathematical knowledge graph based on the first entities and the first association relationships.

4. The method of claim 3, wherein, The determination of the first entities in the target mathematical textbook comprises: extracting the mathematical knowledge points from the target mathematical textbook; performing knowledge fusion on pairs of mathematical knowledge points that satisfy a first preset condition in the mathematical knowledge points to obtain the first entities, wherein the first preset condition is that the mathematical knowledge points have the same name or have a content similarity greater than a preset similarity threshold.

5. The method of claim 3, wherein, The target mathematical textbook comprises a plurality of mathematical textbooks; and the construction of the mathematical knowledge graph based on the first entities and the first association relationships comprises: determining second entities in the target mathematical textbook, wherein the second entities comprise mathematical disciplines, textbook versions, education stages, textbook names, and chapter names; determining second association relationships between the first entities and the second entities and between the second entities, wherein the second association relationships comprise inclusion relationships between entities; constructing the mathematical knowledge graph based on the first entities, the second entities, the dependency relationships, the progressive relationships, and the inclusion relationships.

6. The method of claim 5, wherein, The retrieval of the knowledge point information related to the question from the mathematical knowledge graph comprises: determining entity information of the first entities related to the question, wherein the entity information comprises a mathematical knowledge point name or a mathematical knowledge point name and at least one second entity related to the first entities related to the question and having an inclusion relationship with the first entities; retrieving the first entities related to the question from the mathematical knowledge graph based on the entity information; determining mathematical knowledge point content corresponding to the first entities as the knowledge point information related to the question.

7. The method of claim 1, wherein, The step scoring standard comprises a deduction rule corresponding to a step error type. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model.

8. The method according to claim 1 or 7, characterized in that, The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model.

9. The method according to claim 1 or 7, characterized in that, The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model.

10. A computing device, comprising: The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer learning on a multi-modal large model. The multi-modal mathematical test correction model is obtained by mathematical test correction transfer

Images