Method and system for stipulating answering steps of mathematical subjective questions

By using a pre-trained language model and knowledge graph based on Transformer architecture for deep semantic understanding and matching, combined with graph neural network reasoning, the problems of low efficiency and inaccurate analysis of traditional mathematical subjective problem correction methods are solved, and efficient and intelligent correction and feedback are achieved.

CN120197696AInactive Publication Date: 2025-06-24新疆理工学院
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202510266880.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-03-07
Publication Date
2025-06-24
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The traditional mathematical subjective problem correction method relies on manual paper-based judgment and simple keyword matching technology, making it difficult to accurately analyze complex problem-solving processes, especially the parts that include semantic understanding and logical reasoning, and are inefficient and cannot meet the needs of large-scale exams.

Method used

The pre-trained language model based on the Transformer architecture is used for deep semantic understanding, generate semantic representation information, and divide the problem-solving information into semantic coherent atomic steps. Through matching with the knowledge graph and reasoning of graph neural networks, a solution step diagram is constructed, reference answers are generated and semantic matching is performed, and batch feedback is provided.

Benefits of technology

It realizes accurate analysis and correction of the answer steps of subjective mathematical questions, improves the automation and intelligence level of the problem-solving process, improves the correction efficiency and teaching effect, and helps the answerers better understand the problem-solving ideas and improve their learning methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120197696A_ABST
    Figure CN120197696A_ABST
Patent Text Reader

Abstract

The invention relates to a stipulation processing method and system for mathematical subjective question answering steps. The method comprises the steps of performing deep semantic understanding on problem solving information of mathematical subjective problems based on a pre-training language model of a Transform architecture, generating semantic representation information, and segmenting the problem solving information into semantically coherent atomic steps based on the information; matching the atomic steps with a knowledge graph to obtain a first matching result, and determining a connection relationship between the atomic steps based on the knowledge graph; based on the connection relation of the atomic steps and the atomic steps, constructing an answer step graph, and performing reasoning through a graph neural network to obtain a reasoning result; a reference answer is generated according to the reasoning result, semantic matching is conducted on the atomic step and the reference answer, a second matching result is obtained, and the result is used for indicating generation of correction feedback. According to the method, through semantic understanding, the knowledge graph and the graph neural network, automatic and accurate answer step analysis can be realized, and the question solving efficiency and the feedback quality are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and particularly relates to a method and system for regularizing the solution steps of mathematical subjective questions. Background Art

[0002] With the rapid development of computer science and information technology, in the field of educational evaluation, the technology of automatic marking and grading has become a research hotspot and development trend. However, traditional automatic marking and grading technologies mainly focus on the judgment of objective questions such as multiple-choice questions and fill-in-the-blank questions, and the correctness of students' answers can be automatically judged through clear standard answers. Although this method further reduces the workload of manual marking, there are certain limitations in the analysis of traditional subjective questions.

[0003] Currently, the methods for grading mathematical subjective questions mainly rely on manual marking and simple keyword matching technologies. Although manual marking can relatively accurately understand students' problem-solving ideas, it is inefficient and difficult to meet the needs of large-scale examinations. The simple keyword matching technology can judge the correctness of the answer by extracting keywords in the question and the answer, but for complex problem-solving processes, especially those involving semantic understanding and logical reasoning, it is often unable to accurately analyze. Summary of the Invention

[0004] Based on this, in view of the above technical problems, the purpose of the present invention is to provide a method and system for regularizing the solution steps of mathematical subjective questions to accurately analyze the problem-solving steps and logical relationships and improve the automation and intelligence levels of the problem-solving marking process.

[0005] In a first aspect, the present application provides a method for regularizing the solution steps of mathematical subjective questions, including:

[0006] Using a pre-trained language model based on the Transformer architecture to perform in-depth semantic understanding on the problem-solving information of mathematical subjective questions, generating semantic representation information, and segmenting the problem-solving information into semantically coherent atomic steps based on the semantic representation information; wherein, the problem-solving information includes text information and graphic information;

[0007] Matching the atomic steps with a knowledge graph to obtain a first matching result, and determining the connection relationship between the atomic steps based on the knowledge graph, the first matching result including matched atomic steps and unmatched atomic steps; wherein, the knowledge graph is constructed according to the mathematical entities included in the mathematical discipline system and the relationships between the mathematical entities, and the mathematical entities are used to represent mathematical knowledge point information and / or knowledge carriers carrying mathematical knowledge point information;

[0008] Construct an answer step graph based on the connection relationships and atomic steps of each atomic step, and perform reasoning on the answer step graph through a graph neural network to obtain a reasoning result;

[0009] Generate a reference answer according to the reasoning result, and perform semantic matching between the atomic steps and the reference answer to obtain a second matching result, which is used to indicate the generation of marking feedback, and the marking feedback includes correct or wrong prompts, marking scores, and associated knowledge point recommendations.

[0010] In one embodiment, a pre-trained language model based on the Transformer architecture is used to perform in-depth semantic understanding on the problem-solving information to generate semantic representation information, including:

[0011] Extract features from the problem-solving information to obtain text features and graphic features, and fuse the text features and graphic features to obtain fused features;

[0012] Perform embedding processing on the fused features to obtain embedding processing features, and the embedding processing includes word embedding, position embedding, segment embedding, and semantic enhancement embedding;

[0013] Divide approximate steps according to the text structure of the problem-solving information, and calculate the weight coefficients corresponding to the approximate steps through a preset function;

[0014] Calculate the attention scores of the embedding processing features based on the weight coefficients through a dynamic attention mechanism;

[0015] Process the embedding processing features corresponding to the attention scores through a feed-forward neural network based on residual connection and layer normalization to obtain semantic representation information.

[0016] In one embodiment, generate a reference answer according to the reasoning result, and the semantic matching between the atomic steps and the reference answer to obtain a second matching result includes:

[0017] Based on a generative adversarial network, compare and optimize the reasoning result with the standard answer to generate a reference answer;

[0018] Convert the atomic steps and the reference answer into word vector representations, and calculate the cosine similarity and word moving distance between the atomic steps and the reference answer;

[0019] Input the cosine similarity and the word moving distance into a preset matching function to obtain a second matching result.

[0020] In one embodiment, the preset matching function is:

[0021]

[0022] where S is the result of the preset matching function, that is, the comprehensive matching score, α is the proportionality coefficient, arg(Cij is the cosine similarity C ij and the average value of a i represents the set of word vectors of atomic steps, and b j represents the set of word vectors of the reference answer. WMD(a i , b j ) is the word moving distance from the atomic step to the reference answer.

[0023] In one embodiment, based on the first matching result, processing is performed based on the knowledge graph, and determining the connection relationship of each atomic step includes:

[0024] Normalize the matched atomic steps through the standard expressions of the corresponding matching nodes in the knowledge graph to obtain the normalized steps;

[0025] Infer the unmatched atomic steps based on the inference mechanism of the knowledge graph to obtain the corresponding approximate nodes, and generate the complement steps based on the approximate nodes;

[0026] Calculate the semantic association degree of each atomic step based on the normalized steps and the complement steps, and obtain the path information between each atomic step through the graph search algorithm;

[0027] Calculate according to the semantic association degree and the path information to determine the connection relationship of each atomic step.

[0028] In one embodiment, constructing the answer step graph based on the connection relationship of each atomic step and the atomic step includes:

[0029] Use the atomic steps as the nodes of the answer step graph, and the feature vectors of the nodes include semantic representation information, matching quality features, and position features;

[0030] Use the connection relationship of each atomic step as the edge of the answer step graph, and the edge weight is calculated through the semantic association degree and the path information;

[0031] Based on the nodes and edges, obtain the answer step graph.

[0032] In one embodiment, the method further includes:

[0033] Compare the second matching result with the preset matching threshold to obtain the marking feedback;

[0034] Among them, when the second matching result is higher than the preset matching threshold, generate a correct prompt, and generate a marking score based on the second matching result; obtain the associated knowledge point recommendation according to the atomic steps and the knowledge graph;

[0035] When the second matching result is lower than the preset matching threshold, an error prompt is generated, and a marking score is generated based on the second matching result. By comparing the atomic steps with the reference answer, correction suggestions for the incorrect steps and / or completion prompts for the missing steps are generated; according to the atomic steps and the knowledge graph, recommended related knowledge points are obtained.

[0036] In a second aspect, the present application also provides a system for regularizing the solution steps of mathematical subjective questions, including:

[0037] An atomic step division module, which is used to perform in-depth semantic understanding on the problem-solving information of mathematical subjective questions using a pre-trained language model based on the Transformer architecture, generate semantic representation information, and segment the problem-solving information into semantically coherent atomic steps based on the semantic representation information; where the problem-solving information includes text information and graphical information;

[0038] A step relationship acquisition module, which is used to match the atomic steps with the knowledge graph to obtain a first matching result, and determine the connection relationship between the atomic steps based on the knowledge graph. The first matching result includes the matched atomic steps and the unmatched atomic steps; where the knowledge graph is constructed according to the mathematical entities included in the mathematics discipline system and the relationships between the mathematical entities, and the mathematical entities are used to represent mathematical knowledge point information and / or knowledge carriers carrying mathematical knowledge point information;

[0039] A step reasoning connection module, which is used to construct a solution step graph based on the connection relationship of the atomic steps and the atomic steps, and perform reasoning on the solution step graph through a graph neural network to obtain a reasoning result;

[0040] An answer matching feedback module, which is used to generate a reference answer according to the reasoning result, and perform semantic matching between the atomic steps and the reference answer to obtain a second matching result. The second matching result is used to indicate the generation of marking feedback, and the marking feedback includes correct or error prompts, marking scores, and recommended related knowledge points.

[0041] In a third aspect, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method in any one of the first aspects are implemented.

[0042] In a fourth aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in any one of the first aspects are implemented.

[0043] The above-mentioned method and system for standardizing the steps of solving subjective mathematical problems, computer equipment and storage medium, use a pre-trained language model based on the Transformer architecture to perform deep semantic understanding of the solution information of subjective mathematical problems and generate semantic representation information. Through this process, the text information and graphic information in the question can be fully understood, providing semantic support for the subsequent segmentation of the solution steps. Based on the semantic representation information, the text information and graphic information in the solution information can be effectively parsed and automatically broken down into atomic steps that are easier to handle, ensuring that each step has a clear logical relationship. Compared with traditional single text analysis, it greatly improves the depth of understanding of the solution information and can more accurately grasp the solution ideas.

[0044] Secondly, the atomic steps are matched with the knowledge graph constructed based on the mathematical entities and their relationships in the mathematical discipline system to obtain the first matching result containing matched and unmatched atomic steps, and the connection relationship between each atomic step is determined based on the knowledge graph. Through this process, the problem-solving steps are closely connected with the mathematical knowledge system, and the logical context of the problem-solving steps can be clearly sorted out, which improves the scientificity and rigor of the atomic steps. Based on each atomic step and its connection relationship, a solution step graph can be constructed, and the step graph can be reasoned through the graph neural network, which can analyze and mine the potential logical connection between the steps from the overall structure, further improving the accuracy and comprehensiveness of the reasoning. Finally, a reference answer is generated based on the obtained reasoning result, and the atomic steps are semantically matched with the reference answer to obtain the second matching result, which indicates the generation of correction feedback containing correct or incorrect prompts, correction scores and related knowledge point recommendations. This process not only realizes the automatic correction of problem-solving information, but also provides comprehensive feedback and knowledge expansion for the solver, greatly improving the correction efficiency and teaching effect, and helping to improve mathematics learning ability and learning effect in a targeted manner.

[0045] Compared with traditional methods for grading and analyzing subjective math questions, this method can comprehensively and deeply process text and graphic information in problem-solving information through deep semantic understanding and intelligent matching of knowledge graphs, thereby breaking down the problem-solving process into more detailed and clear steps, and can automatically evaluate and provide feedback, greatly improving the accuracy and efficiency of automatic grading. In addition, by combining the reasoning ability of graph neural networks, the understanding of the complex relationship between problem-solving steps is further enhanced, and the dependencies and reasoning paths between steps can be captured in multiple dimensions, which not only improves the intelligence of the reasoning process, but also can efficiently identify potential logical errors or omissions between steps. By generating corresponding reference answers through reasoning results, it is possible to dynamically adjust the grading strategy and generate targeted grading feedback, further improving learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] To more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0047] Figure 1 A flowchart of a method for regularizing the solution steps of a mathematical subjective question provided for an exemplary embodiment of the present invention;

[0048] Figure 2 A schematic structural diagram of a system for regularizing the solution steps of a mathematical subjective question provided for an exemplary embodiment of the present invention. Detailed implementation manners

[0049] In order to make the objectives, technical solutions, and advantages of the present application more clearly understood, the following further details the present application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] In one embodiment, as Figure 1 shown, a method for regularizing the solution steps of a mathematical subjective question is provided. In this embodiment, this method is exemplified by being applied to a terminal. It can be understood that this method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S101: Use a pre-trained language model based on the Transformer architecture to perform in-depth semantic understanding on the problem-solving information of the mathematical subjective question, generate semantic representation information, and segment the problem-solving information into semantically coherent atomic steps based on the semantic representation information.

[0052] Specifically, the problem-solving information may include text information and graphical information. Transformer is a deep learning architecture based on the self-attention mechanism, which can perform weighted processing on each word or unit of the input through the self-attention mechanism, so as to capture the global dependencies in the sentence or text. By inputting the problem-solving information of the mathematical subjective questions into the pre-trained language model based on the Transformer architecture for semantic understanding, that is, for text information such as the textual description of the problem-solving idea, the explanation of theorem citation, the interpretation of the calculation process, etc., the self-attention mechanism is used to perform weighted processing on each word or symbol, and calculate the degree of association between each position and other positions, so as to not only understand the meaning of individual words, but also capture the long-distance dependencies between words, thus accurately grasping the semantics of the entire sentence or even paragraph. For graphical information, the pre-trained language model can convert it into a numerical representation form suitable for model processing, such as digitizing the features of geometric figures such as side lengths and angles, discretizing the function image into a series of data points, and capturing the relationships between these graphical features based on the self-attention mechanism. Through this process, the pre-trained language model outputs semantic representation information. Through this semantic representation information, the problem-solving process can be divided into several smallest atomic steps with independent meanings according to the coherence and logic of the semantics. Schematically, this semantic representation information is an abstract vector representation, which contains the key semantic content in the problem-solving information and provides a basis for subsequent step segmentation.

[0053] S102: Match the atomic steps with the knowledge graph to obtain the first matching result, and determine the connection relationship between the atomic steps based on the knowledge graph. The first matching result includes the matched atomic steps and the unmatched atomic steps.

[0054] A knowledge graph is a semantic network in a graphical structure, which can be constructed based on mathematical entities contained in the mathematical discipline system and the relationships between mathematical entities. Among them, mathematical entities can be used to represent specific mathematical knowledge point information such as the Pythagorean theorem, the general term formula of an arithmetic sequence, etc., and can also be used to represent knowledge carriers that carry mathematical knowledge point information such as the chapter division of mathematics textbooks, the classification of different mathematical question types, and the presentation forms of mathematical concepts, theorems, and formulas at different learning stages. The transmission relationships between mathematical entities can be reflected as the logical derivation relationships between mathematical knowledge points, such as the process of deriving other theorems from basic axioms, and the application relationships of mathematical knowledge points in different question types, that is, the usage methods and frequencies of a certain knowledge point in different types of questions, as well as the progressive relationship of the difficulty levels of knowledge points. And the divided atomic steps can be semantically matched with the knowledge graph through cosine similarity to obtain the first matching result. This first matching result is divided into matched atomic steps and unmatched atomic steps. Specifically, the matched atomic steps can indicate that the atomic steps can correspond to the existing mathematical knowledge system, while the unmatched atomic steps mean that no directly corresponding content can be found in the knowledge graph. After obtaining the first matching result, the order and logical connection between different atomic steps can be judged through the relationships between mathematical entities in the knowledge graph, and then the connection relationships between each atomic step can be obtained. Through this process, not only can the problem-solving process of the solver be analyzed more deeply, but also an important basis can be provided for subsequent reasoning and feedback, such as judging whether the problem-solving is correct and whether the steps are complete, etc.

[0055] S103: Construct a problem-solving step graph based on the connection relationships and atomic steps of each atomic step, and perform reasoning on the problem-solving step graph through a graph neural network to obtain a reasoning result.

[0056] A problem-solving step graph is a structure that presents the problem-solving process in a graphical form, where atomic steps serve as the nodes of the graph, and the connection relationships between each atomic step form the edges of the graph. Among them, node features can include semantic vectors, knowledge graph matching status, knowledge point type labels, and procedural features such as derivation direction or transmission dependency relationships, etc. And based on this problem-solving step graph, a graph neural network can use the message passing mechanism to receive information from other nodes (atomic steps) and update in combination with its own features, and then can comprehensively analyze and reason about the entire problem-solving step graph and output a reasoning result. Optionally, the reliability of the problem-solving path can be evaluated by calculating the path confidence based on the node matching degree, that is, the matching degree between the atomic step and the entity in the knowledge graph, and the edge weight, that is, the importance degree of the connection relationship represented by the edge, etc., and then the problem-solving reasoning path with high confidence can be screened out.

[0057] S104: Generate a reference answer based on the inference result, perform semantic matching between the atomic steps and the reference answer to obtain a second matching result, which is used to indicate the generation of marking feedback. The marking feedback includes correct or incorrect prompts, marking scores, and recommended related knowledge points.

[0058] Specifically, the inference result obtained by reasoning on the solution step graph through a graph neural network can reflect whether the steps are complete, whether the logic is correct, and whether the knowledge points are used reasonably. Based on this inference result, a corresponding reference answer can be generated. Schematically, matching can be performed through semantic similarity calculation and logical consistency detection, etc., to obtain a second matching result. This result can include the matching status of each atomic step with the reference answer, error types, and the mastery level of related knowledge points. Then, corresponding prompts and marking scores can be generated according to the matching status of each atomic step with the reference answer, and corresponding knowledge points can be recommended according to the error types or the mastery level of related knowledge points.

[0059] In the above method and system for regularizing the solution steps of a mathematical subjective question, by using a pre-trained language model based on the Transformer architecture, it is possible to deeply understand the semantics of the problem-solving information of the mathematical subjective question, and segment the problem-solving process from complex text and graphic information into semantically coherent atomic steps. Furthermore, it can avoid overly simple rules or superficial parsing, deeply understand the logical relationships in mathematical problems, and thus accurately regularize the complex problem-solving process. This not only improves the refinement degree of the problem-solving steps but also helps to automatically identify and process various variables and steps in problem-solving. And through the matching of the knowledge graph and the atomic steps, it is possible to judge whether the atomic steps have been correctly matched to the corresponding mathematical knowledge points based on the mathematical entities constructed according to the mathematical discipline system and the relationships between them, effectively avoiding the problems of incorrect or incomplete answers caused by improper knowledge point matching in traditional methods, and determining the relationships between atomic steps. Secondly, by constructing a solution step graph and using a graph neural network for reasoning, it is possible to conduct in-depth analysis based on the connection relationships between atomic steps, reason from the relevance and global structure between multiple steps, enhance the accuracy and comprehensiveness of the problem-solving process, and reduce errors caused by the blind spots of a single solution path.

[0060] Finally, after generating the reference answers based on the reasoning results, the second matching result is obtained by semantically matching the atomic steps with the reference answers, which provides an accurate basis for generating marking feedback. The marking feedback not only includes correct / incorrect prompts, but also gives marking scores and recommended relevant knowledge points. This marking method based on personalized learning feedback enables the answerer to more clearly understand the specific problems in their problem-solving process and supplement their learning in a timely manner according to the recommended knowledge points, forming a knowledge loop, which helps to make up for the weak links in learning. Compared with the traditional marking method, this method comprehensively applies atomic step semantic parsing, knowledge graph matching, and inference networks, not only greatly improving the marking efficiency, but also quickly and accurately identifying the possible knowledge blind spots of students, thus providing more efficient and accurate learning support for the answerer.

[0061] In an exemplary embodiment, a pre-trained language model based on the Transformer architecture is used to perform deep semantic understanding on the problem-solving information to generate semantic representation information, including:

[0062] Extract features from the problem-solving information to obtain text features and graphic features, and fuse the text features and graphic features to obtain fused features;

[0063] Perform embedding processing on the fused features to obtain embedding processed features, and the embedding processing includes word embedding, position embedding, segment embedding, and semantic enhancement embedding;

[0064] Divide approximate steps according to the text structure of the problem-solving information, and calculate the weight coefficients corresponding to the approximate steps through a preset function;

[0065] Calculate the attention scores of the embedding processed features based on the weight coefficients through a dynamic attention mechanism;

[0066] Process the embedding processed features corresponding to the attention scores through a feed-forward neural network based on residual connection and layer normalization to obtain semantic representation information.

[0067] Schematically, the text part of the problem-solving information contains content such as text descriptions and formula expressions. Natural language processing techniques, such as lexical analysis and syntactic analysis, can be used to extract key information from the text and convert it into numerical features that the model can understand, namely text features. For graphical information such as geometric figures or function images, computer vision techniques can be used to extract features and vectorize them to obtain graphical features. And through operations such as concatenation and weighted summation of the text feature vector and the graphical feature vector, a fused feature can be obtained, enabling the pre-trained language model to comprehensively utilize text and graphical information. Embedding processing is performed on the fused feature to obtain an embedded processing feature, which can convert the fused feature into a form more suitable for model processing and enhance the model's semantic understanding ability. Among them, word embedding can map the vocabulary in the text part of the fused feature to a low-dimensional vector space, so that each vocabulary corresponds to a low-dimensional dense vector, which helps the model understand the semantic relationship of the vocabulary. Position embedding can encode the vocabulary position information in the text information as a vector and add it to the word embedding vector, enabling the model to perceive the position of the vocabulary in the text. Segment embedding can distinguish information in different paragraphs or parts, which helps the model grasp the text structure. And semantic enhancement embedding can incorporate context-related semantic information into the features, or combine with the corresponding question knowledge base to perform semantic enhancement on the involved mathematical concepts, further enriching the semantic representation and improving the model's understanding accuracy of the problem-solving information.

[0068] Optionally, the problem-solving information can be divided into multiple approximate steps with certain semantic coherence by analyzing the grammatical structure, semantic logic of the problem-solving information text, and the conventional process of mathematical problem-solving. For example, in the process of proving a quadratic equation, "transforming the equation into the general form", "calculating the discriminant", "solving the equation according to the discriminant", etc. can be regarded as different approximate steps. For each approximate step, a preset function is used to calculate its weight coefficient. The preset function can be constructed based on the importance of the step in the entire problem-solving process, the degree of association with other steps, the degree of influence on the final result, etc., so that the model can focus on important steps more pertinently when processing information, improving the efficiency and accuracy of information processing. By combining the weight coefficient to calculate the attention score of the embedded processing features, the model can reasonably allocate attention resources according to the importance of the problem-solving steps, thus more effectively extracting key information and improving the ability to understand and analyze problem-solving information. By adding the embedded processing features processed by the attention mechanism to the original embedded processing features, that is, residual connection, it helps to learn the differences and changes between features, avoid the problem of gradient disappearance or gradient explosion in deep networks, and make the training of the model more stable and efficient. And performing layer normalization on the features after residual connection, that is, normalizing the data and standardizing it to a distribution with a mean of 0 and a variance of 1, helps to improve the convergence speed of the model. Inputting the features after residual connection and layer normalization into a feedforward neural network for processing can gradually extract higher-level and abstract feature representations through multi-layer non-linear transformations and combinations, and finally obtain semantic representation information, laying a solid foundation for subsequent in-depth analysis and processing of problem-solving information.

[0069] In an exemplary embodiment, based on the first matching result, processing is performed according to the knowledge graph, and determining the connection relationship of each atomic step includes:

[0070] Normalize the matched atomic steps through the standard expressions of the corresponding matching nodes in the knowledge graph to obtain normalized steps;

[0071] Infer the unmatched atomic steps based on the inference mechanism of the knowledge graph to obtain corresponding approximate nodes, and generate complementary steps based on the approximate nodes;

[0072] Calculate the semantic association degree of each atomic step based on the normalized steps and the complementary steps, and obtain the path information between each atomic step through a graph search algorithm;

[0073] Calculate according to the semantic association degree and the path information to determine the connection relationship of each atomic step.

[0074] Specifically, after matching the atomic steps with the knowledge graph, the matched atomic steps have found corresponding nodes in the knowledge graph. The nodes in the knowledge graph usually have standard and normative expression forms. By adjusting the matched atomic steps according to the standard expressions of the corresponding matched nodes in the knowledge graph, the expressions of the corresponding atomic steps can be made more unified and standardized, avoiding the understanding ambiguity caused by expression differences, and providing a standardized basis for determining the connection relationships between atomic steps. For the unmatched atomic steps, approximate nodes can be obtained through the reasoning mechanism of the knowledge graph, namely the preset entity relationships, rules, etc., and complementary steps can be generated based on the approximate nodes and the corresponding nodes of the associated matched atomic steps. Exemplarily, based on the normalized steps and the complementary steps, the semantic association degree can be obtained by calculating the semantic distance and semantic similarity based on the knowledge graph, and based on the nodes corresponding to the normalized steps and the complementary steps, relevant subgraphs can be extracted from the knowledge graph, and graph search algorithms such as Dijkstra algorithm can be used to search for the shortest paths between nodes in the subgraph, thereby obtaining the path information between each atomic step. Finally, according to the semantic association degree, it can be judged whether there is a direct or indirect logical relationship between the steps, and according to the path information, the sequence and dependency relationships between the steps can be determined, and then the connection relationships of each atomic step can be obtained. Through the semantic reasoning and graph search algorithms of the knowledge graph, this process can not only handle expression differences, but also complement missing steps, and determine the logical sequence between steps through the semantic association degree and path information, providing a basis for the subsequent construction and reasoning of the solution step graph.

[0075] In an exemplary embodiment, constructing a solution step graph based on the connection relationships of each atomic step and the atomic steps includes:

[0076] Taking the atomic steps as the nodes of the solution step graph, and the feature vectors of the nodes include semantic representation information, matching quality features, and position features;

[0077] Taking the connection relationships of each atomic step as the edges of the solution step graph, and the edge weights are calculated through the semantic association degree and path information;

[0078] Based on the nodes and edges, the solution step graph is obtained.

[0079] Among them, the semantic representation information is output based on a pre-trained language model of the Transformer architecture, which can include the encoding of the semantics of atomic steps and can reflect the specific meaning of the corresponding atomic steps in the problem-solving process. The matching quality feature is obtained by quantifying the first matching result and can reflect the degree of matching between the atomic steps and the knowledge graph. The position feature can be obtained from the connection relationships of the atomic steps, thereby clarifying the positions and roles of the individual atomic steps in the overall solution step diagram. The connection relationships between the atomic steps can be used to form the edges of the solution step diagram, and thus the logical connections between the individual atomic steps in the problem-solving process can be reflected, such as causal relationships, sequential relationships, etc. The edge weights are calculated based on the semantic association degrees and path information of the atomic steps corresponding to the edges. Schematically, if the semantic association degree between two atomic steps is high and there is a short and direct connection in the path obtained by graph search, the corresponding edge weight value is high. Taking the feature vectors of all atomic steps as the node set, taking all connection relationships and their weights as the edge set, and a directed weighted graph, that is, the solution step diagram, can be combined based on the node set and the edge set. The solution step diagram not only provides a basis for subsequent graph neural network reasoning but also can intuitively display the semantic associations and logical order between the atomic steps in the problem-solving process, thereby achieving a deep understanding and evaluation of the problem-solving information. Optionally, the solution step diagram can be represented using an adjacency matrix or an adjacency list.

[0080] In an exemplary embodiment, a reference answer is generated according to the inference result, and the atomic steps are semantically matched with the reference answer, and the second matching result includes:

[0081] Based on a generative adversarial network, the inference result is compared and optimized with the standard answer to generate a reference answer;

[0082] The atomic steps and the reference answer are converted into word vector representations, and the cosine similarity and word mover's distance between the atomic steps and the reference answer are calculated;

[0083] The cosine similarity and the word mover's distance are input into a preset matching function to obtain the second matching result.

[0084] Specifically, a generative adversarial network is a deep learning model composed of a generator and a discriminator. The inference result is the analysis and evaluation of the problem-solving process obtained by inferring the solution step diagram. The preliminary reference answer can be generated by inputting the inference result into the generator. And through the discriminator, this preliminary answer can be compared in detail with the standard answer, and the differences between the reference answer and the standard answer in terms of content integrity, logical accuracy, knowledge point application, etc. can be analyzed and the feedback can be output to the generator for iterative adjustment, so as to obtain a more accurate reference answer. By taking advantage of the generative adversarial network, this process can make full use of the information in the inference result and combine the standardization of the standard answer to generate high-quality reference answers, which can not only effectively make up for the possible deficiencies in the inference result, but also avoid simply relying on the standard answer and ignoring the reasoning ideas in the problem-solving process, providing a reliable reference basis for the subsequent semantic matching with atomic steps. A word vector is a representation method that maps the words in a text to a low-dimensional vector space and can capture the semantic information of the words. The atomic steps and the reference answer are converted into word vector representations, and by calculating the cosine similarity of all corresponding word vectors and performing appropriate aggregation such as averaging, the overall cosine similarity between the atomic steps and the reference answer can be obtained. The higher the cosine similarity, the more similar the directions of the atomic steps and the reference answer in semantics, that is, the closer the expressed semantics are.

[0085] The word mover's distance can measure the minimum moving distance required to convert the words in one text to the words in another text. It can be obtained by finding the optimal matching path between the word vectors in the atomic steps and the reference answer and calculating the total distance required to move the word vectors along this optimal matching path. The smaller the word mover's distance, the easier it is to convert the semantics from the atomic steps to the reference answer, and the more similar the two semantics are. The cosine similarity and the word mover's distance can be used as inputs and calculated through specific algorithms or rules to output a result representing the matching degree between the atomic steps and the reference answer, that is, the second matching result, which lays a foundation for generating accurate and targeted marking feedback subsequently. Schematically, this second matching result is a comprehensive evaluation of the semantic matching degree between the atomic steps and the reference answer, and can be a specific numerical value, a classification result, etc.

[0086] In an exemplary embodiment, the preset matching function is:

[0087]

[0088] where S is the result of the preset matching function, that is, the comprehensive matching score, α is the proportionality coefficient, arg(C ij is the average value of the cosine similarity C ij and a i represents the set of word vectors of the atomic steps, b jThe set of word vectors representing the reference answer, WMD(a i ,b j ) is the word moving distance from the atomic step to the reference answer.

[0089] Specifically, for the set of word vectors a i of the atomic step and the set of word vectors b j of the reference answer, the cosine similarity of each group of word vectors is calculated through the formula . This measures the similarity in direction between the atomic step and the reference answer, with a value range between [-1, 1]. The closer the value is to 1, the more similar the directions of the two vectors are, that is, the more similar the semantics of the corresponding words in the atomic step and the reference answer. The word moving distance WMD(a i ,b j ) measures the difficulty of semantic conversion from the semantics of the atomic step to the semantics of the reference answer. The smaller this value, the easier the semantic conversion from the atomic step to the reference answer, and the more similar their semantics. This preset matching function comprehensively evaluates from two dimensions: the similarity in semantic direction and the difficulty of semantic conversion by combining cosine similarity and word moving distance, and can more comprehensively and accurately reflect the semantic matching degree between the atomic step and the reference answer, providing a more reliable basis for subsequent marking feedback.

[0090] In an exemplary embodiment, the method further includes:

[0091] Comparing the second matching result with a preset matching threshold to obtain marking feedback;

[0092] Among them, when the second matching result is higher than the preset matching threshold, a correct prompt is generated, and a marking score is generated based on the second matching result; relevant knowledge point recommendations are obtained according to the atomic step and the knowledge graph;

[0093] When the second matching result is lower than the preset matching threshold, an error prompt is generated, and a marking score is generated based on the second matching result. By comparing the atomic step and the reference answer, a correction suggestion for the wrong step and / or a completion prompt for the missing step are generated; relevant knowledge point recommendations are obtained according to the atomic step and the knowledge graph.

[0094] Specifically, the second matching result quantifies the matching degree between the atomic steps and the reference answer. The marking score can be generated according to a certain scoring rule based on the corresponding matching degree. For example, a full score is given when the matching result reaches a certain high range, and the score is deducted accordingly as the matching degree decreases. When the second matching result is higher than the preset matching threshold, after generating the correct prompt, relevant knowledge points can be recommended by analyzing the knowledge points involved in the atomic steps and combining the association relationships of these knowledge points in the knowledge graph to further consolidate the knowledge points. When the second matching result is lower than the preset matching threshold, after generating the error prompt, the student's atomic steps and the reference answer can be carefully compared to find the differences between the two, such as through semantic similarity. If the difference shows that there is an error in the atomic steps, such as using the wrong formula or having incorrect reasoning logic, then a correction suggestion for the wrong steps is generated, clearly pointing out the error and providing the correct ideas and methods through the knowledge graph; if there are missing key steps in the atomic steps, a prompt for completing the missing steps is generated, and it can be informed through the knowledge graph which steps need to be supplemented to make the problem-solving process complete and correct. And relevant knowledge points can be recommended for the student based on the knowledge points involved in the atomic steps and the association relationships of the knowledge graph to improve the learning effect.

[0095] Based on the same inventive concept, as Figure 2 shown, an embodiment of the present application further provides a regularization processing system 200 for the solution steps of a mathematical subjective question. The system includes:

[0096] An atomic step division module 201, configured to perform in-depth semantic understanding on the problem-solving information of a mathematical subjective question by using a pre-trained language model based on the Transformer architecture, generate semantic representation information, and segment the problem-solving information into semantically coherent atomic steps based on the semantic representation information; wherein, the problem-solving information includes text information and graphic information;

[0097] A step relationship acquisition module 202, configured to match the atomic steps with the knowledge graph to obtain a first matching result, and determine the connection relationship between each atomic step based on the knowledge graph. The first matching result includes the matched atomic steps and the unmatched atomic steps; wherein, the knowledge graph is constructed according to the mathematical entities included in the mathematical discipline system and the relationships between the mathematical entities, and the mathematical entities are used to represent mathematical knowledge point information and / or knowledge carriers carrying mathematical knowledge point information;

[0098] A step reasoning connection module 203, configured to construct an answer step graph based on the connection relationship of each atomic step and the atomic steps, and perform reasoning on the answer step graph through a graph neural network to obtain a reasoning result;

[0099] The answer matching and feedback module 204 is used to generate a reference answer based on the reasoning result, semantically match the atomic steps with the reference answer to obtain a second matching result, and the second matching result is used to indicate the generation of marking feedback, where the marking feedback includes correct or incorrect prompts, marking scores, and recommended related knowledge points.

[0100] The system mainly includes four modules. Among them, the atomic step division module 201 uses a pre-trained language model based on the Transformer architecture to deeply semantically understand the problem-solving information, can parse from text information and graphic information, accurately capture the key steps and information in the problem-solving process, and thus can better divide the problem-solving steps, avoiding omissions and misjudgments during the manual decomposition of the problem-solving process. The step relationship acquisition module 202 can use the mathematical entities and their mutual relationships to determine the logical order and dependencies between atomic steps through matching with the mathematical knowledge graph, avoiding the ambiguity in the understanding of the relationships between steps in traditional methods, and then ensuring the scientificity and correctness of the problem-solving steps. The step reasoning connection module 203 can efficiently simulate and deduce the problem-solving process by constructing an answer step graph and using a graph neural network for reasoning. It not only improves the system's ability to handle complex problems but also generates reasoning results related to the problem-solving steps, making the answer process clearer and more accurate.

[0101] The answer matching and feedback module 204 generates a reference answer through the reasoning result and performs semantic matching, so that it can generate an automatic marking based on the problem-solving steps submitted by the solver, give prompts on whether it is correct or not, and even recommend relevant knowledge points according to the feedback. By combining technologies such as deep semantic understanding, knowledge graph matching, and graph neural network reasoning, the system not only effectively solves the problems of inaccurate understanding, fuzzy step relationships, and low marking efficiency in the traditional automatic marking method for solving mathematical subjective questions but also can provide customized feedback and improvement suggestions according to the problem-solving process of each solver, thereby helping the solver more efficiently master the problem-solving skills and improve the learning effect.

[0102] Furthermore, the atomic step division module 201 includes a semantic analysis sub-module, which is used for:

[0103] Extract features from the problem-solving information to obtain text features and graphic features, and fuse the text features and graphic features to obtain fused features;

[0104] Perform embedding processing on the fused features to obtain embedded processing features, and the embedding processing includes word embedding, position embedding, segment embedding, and semantic enhancement embedding;

[0105] Divide approximate steps according to the text structure of the problem-solving information, and calculate the weight coefficients corresponding to the approximate steps through a preset function;

[0106] Calculate the attention score of the embedded processing features through the weight coefficient based on the dynamic attention mechanism;

[0107] Process the embedded processing features corresponding to the attention scores through a feed-forward neural network based on residual connection and layer normalization to obtain semantic representation information.

[0108] Furthermore, the answer matching feedback module 204 includes:

[0109] A reference answer generation sub-module for comparing and optimizing the inference result with the standard answer based on the generative adversarial network to generate a reference answer;

[0110] A reference answer matching sub-module for:

[0111] Convert the atomic steps and the reference answer into word vector representations, and calculate the cosine similarity and word mover's distance between the atomic steps and the reference answer;

[0112] Input the cosine similarity and word mover's distance into a preset matching function to obtain a second matching result.

[0113] Furthermore, the preset matching function in the reference answer matching sub-module is:

[0114]

[0115] Where S is the result of the preset matching function, that is, the comprehensive matching score, α is the proportionality coefficient, arg(C ij is the average value of the cosine similarity C ij , and a i represents the set of word vectors of the atomic steps, b j represents the set of word vectors of the reference answer, and WMD(a i ,b j ) is the word mover's distance from the atomic steps to the reference answer.

[0116] Furthermore, the step relationship acquisition module 202 includes a connection relationship determination sub-module for:

[0117] Normalize the already matched atomic steps through the standard expressions of the corresponding matching nodes in the knowledge graph to obtain normalized steps;

[0118] Infer the unmatched atomic steps based on the inference mechanism of the knowledge graph to obtain the corresponding approximate nodes, and generate complementary steps based on the approximate nodes;

[0119] Calculate the semantic association degree of each atomic step based on the normalized steps and the complementary steps, and obtain the path information between each atomic step through the graph search algorithm;

[0120] Calculate according to the semantic relevance and path information to determine the connection relationships of the atomic steps.

[0121] Further, the step inference connection module 203 includes an answer step graph construction sub-module, which is used for:

[0122] Taking the atomic steps as the nodes of the answer step graph, and the feature vectors of the nodes include semantic representation information, matching quality features, and position features;

[0123] Taking the connection relationships of the atomic steps as the edges of the answer step graph, and the edge weights are calculated through the semantic relevance and path information;

[0124] Based on the nodes and edges, obtain the answer step graph.

[0125] Further, the system further includes a marking feedback generation module, which is used for:

[0126] Comparing the second matching result with a preset matching threshold to obtain marking feedback;

[0127] Wherein, when the second matching result is higher than the preset matching threshold, generate a correct prompt, and generate a marking score based on the second matching result; obtain associated knowledge point recommendations according to the atomic steps and the knowledge graph;

[0128] When the second matching result is lower than the preset matching threshold, generate an error prompt, and generate a marking score based on the second matching result. By comparing the atomic steps with the reference answers, generate correction suggestions for the error steps and / or completion prompts for the missing steps; obtain associated knowledge point recommendations according to the atomic steps and the knowledge graph.

[0129] In an exemplary embodiment, the present invention further provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of a method for regularizing the solution steps of a mathematical subjective question in the present application. A multi-core processor is preferably used to improve the parallel processing ability of the system. Memory: Provide sufficient temporary storage space to support the operation of the program and the processing of data. The memory capacity should be large enough to accommodate a large amount of supply information and computing tasks.

[0130] In an exemplary embodiment, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for regularizing the solution steps of a mathematical subjective question in the present application are implemented. The computer-readable storage medium may include: a read-only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a solid-state drive (SSD, Solid State Drives), or an optical disc, etc. Among them, the random access memory may include a resistive random access memory (ReRAM, Resistance Random Access Memory) and a dynamic random access memory (DRAM, Dynamic Random Access Memory).

[0131] The above-described embodiments merely represent several implementation manners of the embodiments of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the embodiments of the application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A standardized processing method for solving subjective mathematical problems, characterized in that: The method comprises: Using a pre-trained language model based on the Transformer architecture to perform deep semantic understanding on the solution information of a subjective math problem, generating semantic representation information, and dividing the solution information into semantically coherent atomic steps based on the semantic representation information; wherein the solution information includes text information and graphic information; Matching the atomic steps with the knowledge graph to obtain a first matching result, and determining the connection relationship between the atomic steps based on the knowledge graph, wherein the first matching result includes matched atomic steps and unmatched atomic steps; wherein the knowledge graph is constructed based on the mathematical entities included in the mathematical subject system and the relationship between the mathematical entities, and the mathematical entities are used to represent mathematical knowledge point information and / or knowledge carriers that carry the mathematical knowledge point information; Constructing a solution step graph based on the connection relationship between each of the atomic steps and the atomic steps, and reasoning the solution step graph through a graph neural network to obtain a reasoning result; A reference answer is generated according to the reasoning result, and the atomic step is semantically matched with the reference answer to obtain a second matching result, wherein the second matching result is used to indicate generation of correction feedback, wherein the correction feedback includes a correct or incorrect prompt, a correction score, and a recommendation of related knowledge points.

2. The method according to claim 1, characterized in that The pre-trained language model based on the Transformer architecture is used to perform deep semantic understanding of the problem-solving information and generate semantic representation information, including: Extracting features from the problem-solving information to obtain text features and graphic features, and fusing the text features and the graphic features to obtain fused features; Performing embedding processing on the fused features to obtain embedded processing features, wherein the embedding processing includes word embedding, position embedding, segment embedding and semantic enhancement embedding; Divide the approximate steps according to the text structure of the problem-solving information, and calculate the weight coefficients corresponding to the approximate steps through a preset function; Calculating the attention score of the embedded processing feature through the weight coefficient based on the dynamic attention mechanism; The embedded processing features corresponding to the attention score are processed through a feedforward neural network based on residual connection and layer normalization to obtain the semantic representation information.

3. The method according to claim 1, characterized in that Generating a reference answer according to the reasoning result, and semantically matching the atomic step with the reference answer to obtain a second matching result includes: Comparing and optimizing the inference result with the standard answer based on a generative adversarial network to generate the reference answer; Convert the atomic steps and the reference answer into word vector representations, and calculate the cosine similarity and word shift distance between the atomic steps and the reference answer; The cosine similarity and the word shift distance are input into a preset matching function to obtain the second matching result.

4. The method according to claim 3, characterized in that The preset matching function is: Where S is the result of the preset matching function, that is, the comprehensive matching score, α is the proportionality coefficient, arg(C ij ) is the cosine similarity C ij The average value of a i The word vector set representing the atomic step, b j Represents the word vector set of the reference answer, WMD(a i ,b j ) is the word shift distance from the atomic step to the reference answer.

5. The method according to claim 1, characterized in that The processing based on the knowledge graph according to the first matching result to determine the connection relationship of each of the atomic steps includes: Normalizing the matched atomic steps by using the standard representation of the corresponding matching nodes in the knowledge graph to obtain normalized steps; Reasoning the unmatched atomic steps based on the reasoning mechanism of the knowledge graph to obtain corresponding approximate nodes, and generating completion steps based on the approximate nodes; Calculating the semantic association of each of the atomic steps based on the normalization step and the completion step, and obtaining path information between each of the atomic steps through a graph search algorithm; The connection relationship between the atomic steps is determined by performing calculations based on the semantic association and the path information.

6. The method according to claim 5, characterized in that The constructing of a solution step diagram based on the connection relationship of each of the atomic steps and the atomic steps includes: The atomic step is used as a node of the solution step graph, wherein the feature vector of the node includes the semantic representation information, the matching quality feature and the position feature; The connection relationship of each of the atomic steps is used as an edge of the solution step graph, and the edge weight is calculated by the semantic association and the path information; Based on the nodes and the edges, the solution step graph is obtained.

7. According to claim 1 Any one of 6 The method is characterized in that The method further comprises: Comparing the second matching result with a preset matching threshold to obtain the correction feedback; Wherein, when the second matching result is higher than the preset matching threshold, a correct prompt is generated, and the correction score is generated based on the second matching result; and the associated knowledge point recommendation is obtained according to the atomic steps and the knowledge graph; When the second matching result is lower than the preset matching threshold, an error prompt is generated, and the correction score is generated based on the second matching result. By comparing the atomic steps and the reference answers, correction suggestions for the wrong steps and / or completion prompts for the missing steps are generated; based on the atomic steps and the knowledge graph, the related knowledge point recommendations are obtained.

8. A standardized processing system for solving subjective mathematical problems, characterized in that: The system comprises: An atomic step division module is used to use a pre-trained language model based on the Transformer architecture to perform deep semantic understanding on the solution information of a subjective math problem, generate semantic representation information, and divide the solution information into semantically coherent atomic steps based on the semantic representation information; wherein the solution information includes text information and graphic information; A step relationship acquisition module, used to match the atomic steps with the knowledge graph to obtain a first matching result, and determine the connection relationship between the atomic steps based on the knowledge graph, wherein the first matching result includes matched atomic steps and unmatched atomic steps; wherein the knowledge graph is constructed based on the mathematical entities contained in the mathematical subject system and the relationship between the mathematical entities, and the mathematical entities are used to represent mathematical knowledge point information and / or knowledge carriers that carry the mathematical knowledge point information; A step reasoning connection module is used to construct a solution step graph based on the connection relationship of each atomic step and the atomic steps, and to reason about the solution step graph through a graph neural network to obtain a reasoning result; The answer matching feedback module is used to generate a reference answer based on the reasoning result, and semantically match the atomic step with the reference answer to obtain a second matching result, wherein the second matching result is used to indicate the generation of correction feedback, and the correction feedback includes correct or incorrect prompts, correction scores and related knowledge point recommendations.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Citation Information

Cited By

  • Teaching optimization method, system and equipment based on knowledge space and medium

    CN120782083A

  • A knowledge space-based teaching optimization method, system, device and medium

    CN120782083B

  • Mathematical solution process step-level correction and feedback generation method based on deep learning

    CN121836992A

  • Mathematical solution process step grading and feedback generation method based on deep learning

    CN121836992B