A student problem-solving state evaluation method based on a knowledge graph

Abstract

The application discloses a student problem solving state evaluation method based on a knowledge graph, relates to the technical field of intelligent education, and comprises the following steps: performing entity recognition and semantic analysis on test question text data to obtain knowledge point text data; taking the knowledge point text data as knowledge point nodes to construct a problem solving knowledge graph; acquiring a standard problem solving directed path through the problem solving knowledge graph; acquiring historical answering data of a student and mapping the historical answering data to the problem solving knowledge graph; calculating a mastery score of the student on the knowledge point nodes and constructing a student personal ability portrait; acquiring actual answering data of the student; generating a corresponding student cognitive directed acyclic path in the problem solving knowledge graph; constructing an edit distance matrix according to the student cognitive directed acyclic path and the standard problem solving directed path; obtaining an optimal edit path through the edit distance matrix and calculating a logical deviation coefficient; determining a problem solving logical state grade of the student through the logical deviation coefficient; and evaluating the problem solving logical state in combination with the student personal ability portrait.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to a method for evaluating students' problem-solving status based on knowledge graphs. Background Technology

[0002] In recent years, with the vigorous development of smart education and educational informatization, the use of computer technology to assist teaching has become an important driving force for educational reform. In adaptive learning systems and intelligent tutoring platforms, the core prerequisite for achieving personalized instruction is the accurate capture and quantification of learners' cognitive states. How to extract characteristics reflecting students' true thinking levels from massive amounts of learning behavior data, and then provide targeted teaching interventions, is a key issue that urgently needs to be addressed in the field of smart education.

[0003] Currently, Chinese patent application number 202510913974.7 discloses a personalized tutoring method that integrates collaborative cognitive diagnosis and adaptive teaching. By constructing a knowledge hierarchy tree, it uses a large language model to conduct collaborative cognitive diagnosis on students. Based on students' answers to questions, it marks the nodes on the knowledge hierarchy tree as mastered and not mastered, and generates tutoring strategies for fast thinking and slow thinking. However, while the aforementioned technologies have achieved a certain level of tracking students' knowledge mastery, they still have significant limitations in assessing the logical thinking behind complex problem-solving processes. Specifically, these limitations are as follows: Existing technologies mainly rely on tree structures to represent the pre- and post-knowledge relationships. This single hierarchical structure is insufficient to depict the thought processes of students when solving complex problems. It only focuses on whether the final answer is correct or whether a single knowledge node is covered, failing to identify the logical coherence between students' problem-solving steps. Existing technologies simply attribute students' errors to a lack of mastery, lacking in-depth analysis of the mechanisms behind these errors. Existing technologies typically update cognitive states in real time based on current interactive data, lacking in-depth mining and time-sensitive weighting of historical data. This makes it difficult to effectively distinguish between fluctuations in students' states and the solidification of their ability deficiencies, resulting in assessment results that are easily affected by the state of a single problem-solving session, leading to fluctuations in accuracy and thus limiting the guiding value for adjusting teaching strategies. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies only focus on the final answer and the coverage of a single knowledge point, making it difficult to quantitatively represent and accurately diagnose the logical topology of students' thinking during the problem-solving process. They cannot identify specific structural defects such as logical jumps, redundant thinking, and reversed causality from the perspective of the graph path, and lack a stability analysis mechanism that combines historical ability profiles. This makes it difficult to effectively distinguish between atypical mistakes and typical knowledge blind spots, thus limiting the accuracy of the assessment results and the teaching guidance value.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for evaluating student problem-solving status based on knowledge graphs, comprising the following steps:

[0006] Step S1: Perform entity recognition and semantic analysis on the test question text data to obtain knowledge point text data. Use the knowledge point text data as knowledge point nodes to construct a problem-solving knowledge graph. Obtain the standard problem-solving directed path through the problem-solving knowledge graph.

[0007] Step S2: Obtain students' historical answer data and map it into the problem-solving knowledge graph. Calculate the students' mastery scores of knowledge points and construct individual ability profiles of students based on the mastery scores.

[0008] Step S3: Obtain the actual answer data of students to the test questions, and generate the corresponding directed acyclic path of student cognition in the problem-solving knowledge graph based on the actual answer data;

[0009] Step S4: Construct an edit distance matrix based on the student's cognitive directed acyclic path and the standard problem-solving directed path; obtain the optimal edit path through the edit distance matrix and calculate the logical deviation coefficient.

[0010] Step S5: Determine the student's logical state level in solving the test questions by using the logical deviation coefficient, and evaluate the logical state in solving the questions in combination with the student's personal ability profile.

[0011] Preferably, step S1 includes the following sub-steps:

[0012] Step S101: Sequence labeling of test question text data in the question bank is performed through bidirectional long short-term memory network and conditional random field, knowledge point text data is extracted, and a unique identifier code is assigned to each knowledge point text data. The knowledge point text data is input into the pre-trained language model, and the knowledge point text vector corresponding to the knowledge point text data is output.

[0013] Step S102: Perform semantic dependency parsing on the test question text data, identify the relationship between the dominant and subordinate words in the question stem logic of the knowledge point text data, take the knowledge point text data as knowledge point nodes, and take the relationship between the dominant and subordinate words as the edges between the knowledge point text data to construct a problem-solving knowledge graph.

[0014] Preferably, step S1 further includes the following sub-steps:

[0015] Step S103: The standard solution text of the test question is segmented into ordered atomic solution text segments according to punctuation marks and logical connectors, and the ordered atomic solution text segments are used to generate text feature vectors through a pre-trained language model;

[0016] Step S104: Calculate the first cosine similarity between the text feature vector and the knowledge point text vector, select knowledge point nodes with a first cosine similarity greater than a first preset similarity threshold as mapping nodes, and connect the mapping nodes according to the original order of the atomic problem-solving text segments to generate a standard problem-solving directed path.

[0017] Preferably, step S2 includes the following sub-steps:

[0018] Step S201: Obtain the student's historical answer data within a preset time window, and parse the unique identifier code of the test text data, the cumulative number of answers, the answer score rate, and the answer timestamp of the historical answer data;

[0019] Based on the unique identifier code of the test question text data, obtain the knowledge point text data and the assigned unique identifier code corresponding to the test question text data, and use the unique identifier code of the corresponding knowledge point text data to query the problem-solving knowledge graph to obtain the knowledge point node corresponding to the historical answer data.

[0020] Step S202: Calculate the time difference between the answer timestamp and the real-time timestamp. Using the time difference and a preset forgetting rate constant, calculate the memory retention coefficient corresponding to the historical answer data. The mathematical expression for the memory retention coefficient is:

[0021] ;

[0022] in, For memory retention coefficient, This is the time difference. This is a preset forgetting rate constant.

[0023] Preferably, step S2 further includes the following sub-steps:

[0024] Step S203: Locate the student's historical answer data related to the knowledge points, obtain the difficulty coefficient of the corresponding questions in the historical answer data, and calculate the student's mastery score for the knowledge points based on the answer score rate, memory retention coefficient, difficulty coefficient, and cumulative number of answers in the historical answer data. The mathematical expression for the mastery score is:

[0025] ;

[0026] in, For the first The score for mastery of each knowledge point. To accumulate the number of answers, For the first The answer score rate of the historical answer data. For the first The memory retention coefficient of historical answer data. For the first The difficulty level of the historical answer data;

[0027] Step S204: According to the numerical order of the unique identifier encoding, the mastery scores of all knowledge point nodes are encapsulated into one-dimensional feature vectors, and all one-dimensional feature vectors are combined into a student's personal ability profile.

[0028] Preferably, step S3 includes the following sub-steps:

[0029] Step S301: Obtain students' actual answer data for the test questions, use regular expressions to identify explicit segmentation symbols in the answer data, and use a natural language processing model to identify logical connectors that represent deductive relationships in the actual answer data. Use the explicit segmentation symbols and logical connectors as segmentation points to divide the actual answer data into student problem-solving sub-steps arranged in time sequence. Use a pre-trained language model to extract semantic features of student problem-solving sub-steps, and generate corresponding sub-step feature vectors through the semantic features.

[0030] Step S302: Traverse all knowledge point nodes in the problem-solving knowledge graph, calculate the second cosine similarity between the feature vector of the sub-step and the text vector of the knowledge point, and for each student's problem-solving sub-step, select the knowledge point node whose second cosine similarity is greater than the second preset similarity threshold and output it as the cognitive mapping node corresponding to the student's problem-solving sub-step.

[0031] Preferably, step S3 further includes the following sub-steps:

[0032] Step S303: Map the student's problem-solving sub-steps to nodes in the problem-solving knowledge graph. For different student problem-solving sub-steps mapped to the same knowledge point node, normalize them to the same cognitive mapping node in the problem-solving knowledge graph.

[0033] Based on the original time sequence of students' problem-solving sub-steps in the actual answer data, identify the cognitive mapping nodes corresponding to adjacent students' problem-solving sub-steps, and establish directed connection edges from the cognitive mapping nodes corresponding to the previous students' problem-solving sub-steps to the cognitive mapping nodes corresponding to the subsequent students' problem-solving sub-steps.

[0034] When the cognitive mapping node corresponding to the subsequent student's problem-solving sub-step is the same as the cognitive mapping node corresponding to the previous student's problem-solving sub-step, the directed connection edge forms a backward edge pointing to the cognitive mapping node corresponding to the previous student's problem-solving sub-step, generating a student cognitive path graph.

[0035] Step S304: The student cognitive path graph is subjected to logical loop detection using a depth-first search algorithm. During the logical loop detection process, a recursive path stack that records the current traversal path is maintained in real time.

[0036] When a directed connection edge pointing to a node that already exists in the recursive path stack is encountered, the directed connection edge is determined to be a backward edge and confirmed as a logical loop.

[0037] Extract all backward edges that form a logical closed loop and mark them as cyclic logical feature edges. Count the number of cyclic logical feature edges and mask the cyclic logical feature edges in the topology to generate a directed acyclic path for student cognition.

[0038] Preferably, step S4 includes the following sub-steps:

[0039] Step S401: Extract the nodes and corresponding logical pointing features in the student's cognitive directed acyclic path, and serialize them into a student logical unit sequence according to the generation timestamp; extract the nodes and corresponding logical pointing features in the standard problem-solving directed path, and serialize them into a standard logical unit sequence; construct an edit distance matrix based on the student logical unit sequence and the standard logical unit sequence.

[0040] Step S402: Starting from the beginning position of the edit distance matrix, the cumulative edit cost of each position in the edit distance matrix is ​​calculated using a dynamic programming algorithm with preset node insertion cost, node deletion cost, and edge direction reversal cost, until the end position of the edit distance matrix is ​​reached.

[0041] Preferably, step S4 further includes the following sub-steps:

[0042] Step S403: Take the connected path with the minimum cumulative editing cost as the optimal editing path, and count the number of redundant nodes, missing nodes and reverse edges in the optimal editing path.

[0043] The number of redundant nodes is the number of times the node deletion operation is performed in the optimal editing path;

[0044] The number of missing nodes is the number of times the node insertion operation is performed in the optimal editing path;

[0045] The number of reverse edges is the number of operations in the optimal editing path that involve nodes with the same but opposite logical pointing characteristics.

[0046] Step S404: Obtain the total number of standard nodes in the directed path of the standard solution. Calculate the logical deviation coefficient using the total number of standard nodes, the number of redundant nodes, the number of missing nodes, the number of cyclic logical feature edges, and the number of reverse edges. The mathematical expression for the logical deviation coefficient is:

[0047] ;

[0048] in, The logical deviation coefficient, The number of redundant nodes. This represents the number of missing nodes. This represents the number of reverse edges. The number of cyclic logical feature edges. The preset redundancy penalty weight, The default penalty weight is the missing value. The preset reverse order penalty weight, The preset cyclic penalty weights, The total number of standard nodes, As a redundancy penalty item, For missing penalty items, This is a reverse order penalty term. This is a cyclical penalty term.

[0049] Preferably, step S5 includes the following sub-steps:

[0050] Step S501: Compare the logic deviation coefficient with the logic deviation threshold. The logic deviation threshold includes a first logic deviation threshold and a second logic deviation threshold, wherein the first logic deviation threshold is less than the second logic deviation threshold.

[0051] If the logical deviation coefficient is less than the first logical deviation threshold, it indicates that the student's logical state level in solving the test question is logically rigorous.

[0052] If the logical deviation coefficient is greater than the first logical deviation threshold and less than the second logical deviation threshold, it indicates that the student's logical state level in solving the test question is general.

[0053] If the logical deviation coefficient is greater than the second logical deviation threshold, it indicates that the student's logical state of solving the test questions is logically confused.

[0054] Step S502: When the problem-solving logic state level is determined to be logically general or logically chaotic, the values ​​of the redundancy penalty term, the missing penalty term, and the reverse order penalty term are calculated respectively.

[0055] Compare the values ​​of the redundancy penalty term, the missing penalty term, and the reverse order penalty term, and take the one with the largest value as the dominant difference term;

[0056] If the dominant difference term is a redundant penalty term, it indicates that the student's problem-solving logic is redundant, suggesting that the student has invalid derivation steps.

[0057] If the dominant difference item is a missing penalty item, it indicates that the student's problem-solving logic is in a logical leap, suggesting that the student is missing a key deduction step;

[0058] If the dominant difference term is a reverse order penalty term, it indicates that the student's problem-solving logic is reversed, suggesting that the student's reasoning direction is contrary to standard logic;

[0059] If the dominant difference term is a circular penalty term, it indicates that the student's problem-solving logic is in a circular argument, suggesting that the student has a logical dead loop where each element is mutually causal.

[0060] Step S503: Extract the mastery score corresponding to the knowledge point node from the student's personal ability profile, calculate the average value of the mastery score and use it as the historical ability benchmark value, calculate the weighted sum of the historical ability benchmark value and the logical deviation coefficient to obtain the state stability index.

[0061] If the problem-solving logic state level is logical confusion and the historical ability benchmark value is greater than the preset high score threshold, it indicates that the student's state is fluctuating abnormally, and the problem-solving logic state is determined to be an atypical error.

[0062] If the problem-solving logic state level is logically confused and the historical ability benchmark value is less than the preset low score threshold, it indicates that the student's ability deficiency is solidified, and the problem-solving logic state is determined to be a typical knowledge blind spot.

[0063] The beneficial effects of this invention are as follows: By calculating the topological differences between the student's cognitive directed acyclic path and the standard problem-solving directed path, a quantitative logical deviation coefficient is generated. This allows the assessment method to break through the limitation of simply relying on the answer score and achieve a precise measurement of the student's logical rigor. By identifying missing nodes, redundant nodes, and reverse edges in the optimal editing path, and associating them with specific problem-solving logical states such as logical jumps, cognitive redundancy, and causal inversion, it can delve into the problem-solving process and provide detailed data support for targeted error correction tutoring. Using the student's personal ability profile as a benchmark, historical ability benchmark values ​​are calculated, and combined with current performance to obtain a state stability index. This effectively solves the one-sidedness of a single assessment and can accurately distinguish between atypical errors caused by abnormal state fluctuations and typical knowledge blind spots caused by solidified ability deficiencies, thereby significantly improving the accuracy of problem-solving logical state assessment. Attached Figure Description

[0064] Figure 1 The flowchart illustrates a method for evaluating student problem-solving status based on knowledge graphs, as provided in one embodiment of the present invention. Detailed Implementation

[0065] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] Example, refer to Figure 1 This paper presents a knowledge graph-based method for evaluating students' problem-solving status, which includes the following steps:

[0067] Step S1: Perform entity recognition and semantic analysis on the test question text data to obtain knowledge point text data. Use the knowledge point text data as knowledge point nodes to construct a problem-solving knowledge graph. Obtain the standard problem-solving directed path through the problem-solving knowledge graph.

[0068] Step S2: Obtain students' historical answer data, map the historical answer data to the problem-solving knowledge graph, calculate the students' mastery scores of knowledge point nodes, and construct a personal ability profile of students based on the mastery scores.

[0069] Step S3: Obtain the actual answer data of students to the test questions, and generate the corresponding directed acyclic path of student cognition in the problem-solving knowledge graph based on the actual answer data.

[0070] Step S4: Construct an edit distance matrix based on the student's cognitive directed acyclic path and the standard problem-solving directed path, obtain the optimal edit path through the edit distance matrix, and calculate the logical deviation coefficient.

[0071] Step S5: Determine the level of the student's logical reasoning state for solving the test questions by using the logical deviation coefficient, and evaluate the logical reasoning state in conjunction with the student's personal ability profile.

[0072] This invention generates a quantified logical deviation coefficient by calculating the topological differences between the student's cognitive directed acyclic path and the standard problem-solving directed path. This allows the assessment method to break through the limitation of simply relying on the answer score and achieve a precise measurement of the student's logical rigor. By identifying missing nodes, redundant nodes, and reverse edges in the optimal editing path, it is associated with specific problem-solving logical states such as logical jumps, cognitive redundancy, and causal inversion. This provides detailed data support for targeted error correction tutoring by delving into the problem-solving process. Using the student's individual ability profile as a benchmark, historical ability benchmark values ​​are calculated and combined with current performance to obtain a state stability index. This effectively solves the one-sidedness of a single assessment and can accurately distinguish between atypical errors caused by abnormal state fluctuations and typical knowledge blind spots caused by solidified ability deficiencies, thereby significantly improving the accuracy of problem-solving logical state assessment.

[0073] In a specific embodiment, step S1 includes the following sub-steps:

[0074] Step S101: Sequence labeling of test question text data in the question bank is performed using a bidirectional long short-term memory network and a conditional random field to extract knowledge point text data. A unique identifier code is assigned to each knowledge point text data. The knowledge point text data is then input into a pre-trained language model, and a knowledge point text vector corresponding to each knowledge point text data is output.

[0075] In this embodiment, the pre-trained language model is a BERT pre-trained model. It uses a bidirectional long short-term memory network (BSSN) and a conditional random field (CRF) to input word vector sequences from the question bank's text data and output entity label sequences. The BSSN and CRF identify text fragments labeled as knowledge points (e.g., the Pythagorean theorem, Newton's second law), extract these fragments into independent knowledge point text data, input this knowledge point text data into the BERT pre-trained model, and output knowledge point text vectors that correspond one-to-one with the knowledge point text data.

[0076] Step S102: Perform semantic dependency parsing on the test question text data, identify the relationship between the dominant and subordinate words in the question stem logic of the knowledge point text data, use the knowledge point text data as knowledge point nodes, and use the relationship between the dominant and subordinate words as edges between the knowledge point text data to construct a problem-solving knowledge graph.

[0077] The governing words include the core predicate and the central noun in the question stem.

[0078] Subordinate terms include nouns and attribute descriptors of knowledge points in text data controlled by the governed term.

[0079] The relationship between a governing word and a subordinate word includes verb-object relationship, attributive-head relationship, and prepositional phrase relationship.

[0080] This invention extracts entities using a bidirectional long short-term memory network and a conditional random field, and combines semantic dependency parsing to identify the logical direction of the dominant and subordinate words in the question stem. This effectively solves the problem that traditional knowledge graph construction relies solely on keyword matching and lacks logical depth. It can automatically parse the causal, deductive, and inclusion relationships implicit in the test question text data, thereby constructing a problem-solving knowledge graph with a clear logical flow. This provides a high-precision underlying data foundation for accurately mapping unstructured student problem-solving steps into structured topological paths.

[0081] In a specific embodiment, step S1 further includes the following sub-steps:

[0082] Step S103: The standard solution text of the test question is segmented into ordered atomic solution text segments according to punctuation marks and logical connectors. The ordered atomic solution text segments are then used to generate text feature vectors through a pre-trained language model.

[0083] The pre-trained language model is a BERT pre-trained model.

[0084] Step S104: Calculate the first cosine similarity between the text feature vector and the knowledge point text vector, select knowledge point nodes with a first cosine similarity greater than the first preset similarity threshold as mapping nodes, and connect the mapping nodes according to the original order of the atomic problem-solving text segments to generate a standard problem-solving directed path.

[0085] This invention segments standard problem-solving text into ordered atomic problem-solving text segments and uses the cosine similarity between text feature vectors and knowledge point vectors for mapping. This solves the technical problem that standard answers are usually continuous long texts, making it difficult for computers to perform fine-grained logical comparisons directly. By discretizing continuous text into ordered logical node sequences, a logical standard for evaluating students' problem-solving process is established, enabling the algorithm to accurately identify minute deviations in the problem-solving process.

[0086] In a specific embodiment, step S2 includes the following sub-steps:

[0087] Step S201: Obtain the student's historical answer data within a preset time window, and parse the unique identifier code of the test question text data, the cumulative number of answers, the answer score rate, and the answer timestamp of the historical answer data.

[0088] Based on the unique identifier code of the test question text data, obtain the knowledge point text data and the assigned unique identifier code corresponding to the test question text data. Use the unique identifier code of the corresponding knowledge point text data to query the problem-solving knowledge graph and obtain the knowledge point node corresponding to the historical answer data.

[0089] Step S202: Calculate the time difference between the answer timestamp and the real-time timestamp. Using the time difference and a preset forgetting rate constant, calculate the memory retention coefficient corresponding to the historical answer data. The mathematical expression for the memory retention coefficient is:

[0090] ;

[0091] in, For memory retention coefficient, This is the time difference. This is a preset forgetting rate constant, which is used to control the rate at which the memory retention coefficient decays as the time difference increases.

[0092] The larger the time difference, the smaller the memory retention coefficient, thus giving higher weight to recent responses.

[0093] It should be noted that the preset forgetting rate constant is set based on the mathematical model of the Ebbinghaus forgetting curve and is adjusted according to the knowledge review cycle of different subjects. For science questions with strong logic, the preset forgetting rate constant can be set to a smaller value. For humanities questions with strong memorization, the preset forgetting rate constant can be set to a larger value. The preset forgetting rate constant is used to control the decay rate of the memory retention coefficient as the time difference increases, reflecting the retention pattern of students' memory of textual data of knowledge points. When the time difference is in days, the preset forgetting rate constant ranges from 0.1 to 0.3, which ensures that within the preset time window (such as one semester), recent historical answer data has a higher weight, while the weight of older data decreases smoothly.

[0094] This invention introduces a memory retention coefficient based on the Ebbinghaus forgetting curve when processing students' historical answer data, and establishes a two-level indexing mechanism from the unique identifier code of the test question to the unique identifier code of the knowledge point text data. This effectively overcomes the shortcomings of traditional assessments that rely solely on average scores and ignore the forgetting patterns of knowledge. By giving higher weight to recent answer data, it can more realistically and dynamically restore the student's current knowledge reserve level, providing a precise time dimension benchmark for distinguishing between state fluctuations and ability solidification.

[0095] In a specific embodiment, step S2 further includes the following sub-steps:

[0096] Step S203: Locate the student's historical answer data related to the knowledge points, and obtain the difficulty coefficient of the corresponding questions in the historical answer data. The mathematical expression for the difficulty coefficient is:

[0097] ;

[0098] in, For the first The difficulty level of each historical answer data point corresponds to a specific question. This represents the total number of students who have answered the corresponding questions in the question bank. For the first Each student's score on the corresponding test question. This is the full score for this question.

[0099] It should be noted that the difficulty coefficient of a question is calculated statistically based on the total historical response data for that question in the question bank. The difficulty coefficient characterizes the ease or difficulty of solving a question, and its value ranges from [value missing]. This indicates that the lower the overall average score rate of a test question, the closer the calculated difficulty coefficient is to 1, indicating that the question is more difficult. Conversely, the higher the overall average score rate, the closer the difficulty coefficient is to 0, indicating that the question is easier. By introducing a difficulty coefficient, students' scores when answering highly difficult questions can be converted into a higher weighting for their mastery of the knowledge points.

[0100] Based on the historical answer data, including the score rate, memory retention coefficient, difficulty coefficient, and cumulative number of answers, the student's mastery score for each knowledge point is calculated. The mathematical expression for the mastery score is:

[0101] ;

[0102] in, For the first The score for mastery of each knowledge point. To accumulate the number of answers, For the first The answer score rate of the historical answer data. For the first The memory retention coefficient of historical answer data. For the first The difficulty level of each historical answer data point.

[0103] Step S204: According to the numerical order of the unique identifier code, the mastery scores of all knowledge point nodes are encapsulated into one-dimensional feature vectors, and all one-dimensional feature vectors are combined into a student's personal ability profile.

[0104] This invention constructs a comprehensive mastery calculation method that includes answer score rate, memory retention coefficient, test question difficulty coefficient, and number of answers. It eliminates the assessment bias caused by fluctuations in the difficulty of test questions. By introducing a difficulty coefficient for weighting, it ensures that students' scores when answering high-difficulty test questions can be converted into higher ability weights, thereby generating a denoised individual student ability profile and providing an objective reference coordinate system for personalized diagnosis.

[0105] In a specific embodiment, step S3 includes the following sub-steps:

[0106] Step S301: Obtain students' actual answers to the test questions, use regular expressions to identify explicit segmentation symbols in the answer data, and use a natural language processing model to identify logical connectors that represent deductive relationships in the actual answer data. Use explicit segmentation symbols and logical connectors as dividing points to segment the actual answer data into student problem-solving sub-steps arranged in time sequence, use a pre-trained language model to extract semantic features of student problem-solving sub-steps, and generate corresponding sub-step feature vectors through semantic features.

[0107] In this embodiment, the natural language processing model employs a sequence labeling model based on a BERT pre-trained model, a bidirectional long short-term memory network, and a conditional random field architecture. The natural language processing model, from bottom to top, includes a BERT pre-trained model embedding layer, a bidirectional long short-term memory network feature extraction layer, and a conditional random field decoding layer.

[0108] The BERT pre-trained model embedding layer serves as the bottom layer of the natural language processing model, receiving text sequences of actual response data. Leveraging the powerful semantic representation capabilities of the BERT pre-trained model, discrete characters are converted into dynamic word vectors containing contextual semantic information.

[0109] The bidirectional long short-term memory network feature extraction layer is located in the middle layer. It uses a bidirectional long short-term memory network to receive dynamic character vectors, capture the long-distance dependencies of the text sequence from the forward and backward directions respectively, and output the context feature vector corresponding to each character.

[0110] The Conditional Random Field (CRF) decoding layer, as the top layer of the model, uses CRF to constrain the decoding of the context feature vector. By combining the transition probabilities between adjacent labels, it predicts the BIO label sequence with the highest global probability, thereby accurately identifying the position of logical connectors.

[0111] Character-level word embedding vectors are extracted through the embedding layer of the BERT pre-trained model. These vectors are then input into the feature extraction layer of the bidirectional long short-term memory network to capture long-distance dependency features of the context. Finally, the label sequence is constrained and decoded through the conditional random field decoding layer to predict the label category corresponding to each character. The output is a BIO label sequence with the same length as the input text. The character segments covered by B-Logic and I-Logic labels (such as "because" and "therefore") are identified as logical connectors. The start and end indexes of these logical connectors in the text are recorded, and the logical connectors are used as segmentation anchors.

[0112] Step S302: Traverse all knowledge point nodes in the problem-solving knowledge graph, calculate the second cosine similarity between the feature vector of the sub-step and the text vector of the knowledge point, for each student's problem-solving sub-step, select the knowledge point node whose second cosine similarity is greater than the second preset similarity threshold and output it as the cognitive mapping node corresponding to the student's problem-solving sub-step, combine all cognitive mapping nodes into a cognitive mapping node set, if the maximum cosine similarity of the problem-solving sub-step is less than the preset similarity threshold, then mark the corresponding problem-solving sub-step as an invalid noise step and remove it.

[0113] It should be noted that a preset similarity threshold is obtained through statistical analysis of sample questions in the question bank. This threshold is used to determine whether a valid semantic mapping relationship exists between the text feature vector and the knowledge point text vector. The preset similarity threshold is set to 0.8. If it is lower than 0.8, irrelevant text segments will be incorrectly mapped to cognitive mapping nodes, generating noise.

[0114] By extracting a set of candidate knowledge point nodes with a cosine similarity greater than a preset similarity threshold, rather than a single knowledge point node with the highest similarity, we can effectively capture the complex knowledge point information involved in a single problem-solving step. This ensures the consistency of the student's cognitive directed acyclic graph with the standard path in terms of logical granularity and avoids logical jump misjudgments caused by overly strict semantic mapping.

[0115] This invention utilizes a natural language processing model to identify implicit logical connectors, combines explicit segmentation symbols to perform fine-grained segmentation of the answer data, and uses a similarity threshold to filter out invalid noise. It effectively solves the parsing difficulties caused by students' non-standard answer styles (such as lack of punctuation or redundant writing), and can accurately restore continuous answer text into a sequence of sub-steps arranged according to logical thinking. It avoids misjudging irrelevant nonsense as valid deduction steps and ensures the accuracy of cognitive path generation.

[0116] In a specific embodiment, step S3 further includes the following sub-steps:

[0117] Step S303: Map the student's problem-solving sub-steps to nodes in the problem-solving knowledge graph. For different student problem-solving sub-steps mapped to the same knowledge point node, normalize them to the same cognitive mapping node in the problem-solving knowledge graph.

[0118] Based on the original time sequence of students' problem-solving sub-steps in the actual answer data, identify the cognitive mapping nodes corresponding to adjacent students' problem-solving sub-steps, and establish directed connections from the cognitive mapping nodes corresponding to the previous students' problem-solving sub-steps to the cognitive mapping nodes corresponding to the subsequent students' problem-solving sub-steps.

[0119] When the cognitive mapping node corresponding to the subsequent student's problem-solving sub-step is the same as the cognitive mapping node corresponding to the previous student's problem-solving sub-step, the directed connection edge forms a backward edge pointing to the cognitive mapping node corresponding to the previous student's problem-solving sub-step, generating a student cognitive path graph.

[0120] Step S304: Logical loop detection is performed on the student's cognitive path graph using a depth-first search algorithm. During the logical loop detection process, a recursive path stack that records the current traversal path is maintained in real time.

[0121] When a directed edge pointing to a node that already exists in the recursive path stack is encountered, the directed edge is determined to be a backward edge and confirmed as a logical loop.

[0122] Extract all backward edges that form a logical closed loop and mark them as cyclic logical feature edges. Count the number of cyclic logical feature edges and mask the cyclic logical feature edges in the topology to generate directed acyclic paths for student cognition.

[0123] This invention addresses the specific logical fallacy of circular reasoning by proposing an innovative processing strategy of feature labeling and topological masking. Specifically, when a logical loop is detected, the number of circular feature edges is counted, but when generating the path graph, the circular logical feature edges are only masked in the topological structure. This resolves the mathematical conflict between logical loops and the requirement of a directed acyclic graph in the edit distance algorithm. It ensures the executability of the subsequent dynamic programming algorithm, fully preserves the statistical evidence of serious logical errors by students, and avoids the loss of logical depth caused by forced disconnection.

[0124] In a specific embodiment, step S4 includes the following sub-steps:

[0125] Step S401: Extract the knowledge point nodes and corresponding logical pointing features from the student's directed acyclic path of cognition, and serialize them into a sequence of student logical units according to the generation timestamp. Specifically, this includes:

[0126] Based on the original generation timestamps of the students' problem-solving sub-steps, the knowledge point nodes in the directed acyclic graph of students' cognition are sorted in ascending order to obtain an ordered list of student nodes. .

[0127] For each current node in the ordered list of student nodes Searching for the next time-series node in the directed acyclic graph of student cognition. The directed connecting edges.

[0128] Extract the direction vector of the directed connection edge as the logical pointing feature of the current knowledge point node.

[0129] If the current knowledge point node is the last node, then the logical pointer feature is marked as a terminator. The current knowledge point node and its corresponding logical pointer feature are encapsulated into a tuple, which serves as the logical unit in the student's logical unit sequence.

[0130] Extract the nodes and corresponding logical pointer features from the directed path of the standard solution, and serialize them into a standard logical unit sequence, specifically including:

[0131] Based on the standard directed path for problem-solving, knowledge point nodes are extracted according to the logical order of the atomic problem-solving text segments, resulting in an ordered list of standard nodes. .

[0132] For each standard node in the ordered list of standard nodes Identify which standard nodes point to successor nodes in the standard path. The directed logical relationship defines the standard logical flow as the logical pointing characteristic of the standard node.

[0133] The standard node and its corresponding logical pointer feature are encapsulated into a tuple, which serves as the logical unit in the standard logical unit sequence.

[0134] An edit distance matrix is ​​constructed based on the student logic unit sequence and the standard logic unit sequence, where the row dimension of the edit distance matrix corresponds to the student logic unit sequence and the column dimension of the edit distance matrix corresponds to the standard logic unit sequence.

[0135] Step S402: Starting from the beginning position of the edit distance matrix, the cumulative edit cost of each position in the edit distance matrix is ​​calculated using a dynamic programming search algorithm with preset node insertion cost, node deletion cost, and edge direction reversal cost, until the end position of the edit distance matrix is ​​reached.

[0136] The preset node insertion cost is the preset computational cost assigned in dynamic programming search to complete the student logic unit sequence for missing units relative to the standard logic unit sequence.

[0137] The preset node deletion cost is the preset computational cost assigned in dynamic programming search to remove redundant units in the student logic unit sequence relative to the standard logic unit sequence.

[0138] The preset edge direction reversal cost is the preset computational cost assigned to correct the difference in logical pointing characteristics when the matched student logical unit and the standard logical unit have the same node but opposite logical pointing characteristics in dynamic programming search.

[0139] It should be noted that when constructing the edit distance matrix and executing the dynamic programming search algorithm, the computational cost of each atomic operation is defined. The degree of damage to the correctness of the solution is weighted according to different types of logical errors, with logical direction errors (reversal) being more severe than missing steps (insertion) and redundant steps (deletion).

[0140] The cost of node deletion is set between 0.5 and 0.8. Since redundant thinking usually only represents invalid steps written by the student, a minor flaw, it is assigned a low cost. This makes the dynamic programming algorithm tend to identify non-critical differences as redundant nodes rather than forcibly matching them as erroneous logic. The cost of node insertion is set between 1.0 and 1.2. Logical jumps mean a break in the derivation chain, a moderately serious error, and its cost is used as the baseline. The cost of edge direction reversal is set between 1.5 and 2.0. Causal reversal is a fundamental logical construction error, which has the greatest impact on the correctness of the solution, and therefore is assigned the highest cost.

[0141] This invention encapsulates node entities and their corresponding logical pointing features into a unified serialization unit to construct an edit distance matrix. This solves the problem that traditional edit distance algorithms can only compare node differences but cannot detect dimensional errors in direction. It enables the algorithm to keenly capture the reverse logic of correct nodes but opposite directions during the comparison process, providing a rigorous mathematical structure to accurately calculate the cost of edge direction reversal.

[0142] In a specific embodiment, step S4 further includes the following sub-steps:

[0143] Step S403: Select the connected path with the minimum cumulative editing cost as the optimal editing path, and count the number of redundant nodes, missing nodes, and reverse edges in the optimal editing path.

[0144] The number of redundant nodes is the number of times a node deletion operation is performed in the optimal editing path.

[0145] The number of missing nodes is the number of times node insertion operations are performed in the optimal editing path.

[0146] The number of reverse edges is the number of operations in the optimal editing path that involve nodes with the same name but opposite logical pointing characteristics.

[0147] Step S404: Obtain the total number of standard nodes in the directed path of the standard solution. Calculate the logical deviation coefficient using the total number of standard nodes, the number of redundant nodes, the number of missing nodes, the number of cyclic logical feature edges, and the number of reverse edges. The mathematical expression for the logical deviation coefficient is:

[0148] ;

[0149] in, The logical deviation coefficient, The number of redundant nodes. This represents the number of missing nodes. This represents the number of reverse edges. The number of cyclic logical feature edges. The preset redundancy penalty weight, The default penalty weight is the missing value. The preset reverse order penalty weight, The preset cyclic penalty weights, The total number of standard nodes, As a redundancy penalty item, For missing penalty items, This is a reverse order penalty term. This is a cyclical penalty term.

[0150] The preset redundancy penalty weight, missing penalty weight, reverse order penalty weight, and loop penalty weight are used to distinguish the degree of impact of different logical error types on the problem-solving state. The total number of standard nodes is used to normalize the deviation coefficient and eliminate the influence of the length of the test steps on the evaluation results.

[0151] It should be noted that the preset weights for redundancy penalties, missing penalty penalties, and reverse order penalties are determined according to the teaching syllabus's requirements for problem-solving standards. If the teaching objective emphasizes the completeness of steps, the missing penalty weight is increased; if the emphasis is on the rigor of logical deduction, the reverse order penalty weight is increased. The redundancy penalty weight ranges from 0.5 to 0.8, as redundant thinking is generally considered a minor flaw. The missing penalty weight ranges from 1.0 to 1.2, as logical jumps affect the completeness of the deduction chain. The reverse order penalty weight ranges from 1.5 to 2.0, as causal reversal is a fundamental logical error. The circular reasoning penalty weight ranges from 2.0 to 2.5, as circular reasoning is a serious logical fallacy, meaning the problem-solving path is invalid and self-contained. Compared to causal reversal, circular reasoning is more destructive to the problem-solving logic, therefore it is given the highest penalty weight.

[0152] This invention achieves a quantitative representation of the problem-solving logic state by using dynamic programming to backtrack to the optimal path and calculate a normalized logical deviation coefficient. By assigning higher penalty weights to circular reasoning and causal inversion, it objectively reflects the degree of damage to the correctness of the problem-solving by different error types. At the same time, by using the total number of standard nodes for normalization, it eliminates the impact of differences in the length of the test steps on the evaluation results and achieves horizontal comparability across questions.

[0153] In a specific embodiment, step S5 includes the following sub-steps:

[0154] Step S501: Compare the logic deviation coefficient with the logic deviation threshold. The logic deviation threshold includes a first logic deviation threshold and a second logic deviation threshold, wherein the first logic deviation threshold is less than the second logic deviation threshold.

[0155] If the logical deviation coefficient is less than the first logical deviation threshold, it indicates that the student's logical state level in solving the test question is logically rigorous.

[0156] If the logical deviation coefficient is greater than the first logical deviation threshold and less than the second logical deviation threshold, it indicates that the student's logical state level in solving the test question is "general logic".

[0157] If the logical deviation coefficient is greater than the second logical deviation threshold, it indicates that the student's logical state of solving the test questions is logically confused.

[0158] Step S502: When the problem-solving logic state level is determined to be either logically normal or logically chaotic, calculate the values ​​of the redundancy penalty term, the missing penalty term, and the reverse order penalty term, respectively.

[0159] Compare the values ​​of the redundancy penalty, the missing penalty, and the reverse order penalty, and select the one with the largest value as the dominant difference term.

[0160] If the dominant difference term is a redundant penalty term, it indicates that the student's problem-solving logic is redundant, suggesting that the student has invalid derivation steps.

[0161] If the dominant difference item is a missing penalty item, it indicates that the student's problem-solving logic is in a logical leap, suggesting that the student is missing a key derivation step.

[0162] If the dominant difference term is a reverse order penalty term, it indicates that the student's problem-solving logic is reversed, suggesting that the student's reasoning direction is contrary to standard logic.

[0163] If the dominant difference term is a cyclical penalty term, it indicates that the student's problem-solving logic is in a cyclical argument, suggesting that the student has a logical dead loop where each element is mutually causal.

[0164] Step S503: Extract the mastery scores corresponding to knowledge point nodes from the student's individual ability profile, calculate the average of the mastery scores and use it as the historical ability benchmark value, calculate the weighted sum of the historical ability benchmark value and the logical deviation coefficient to obtain the state stability index. The mathematical expression of the state stability index is:

[0165] ;

[0166] in, It is a state stability index. The weighting coefficients are the historical capability benchmark values. This is the weighting factor for the logic deviation coefficient. Based on historical capability benchmarks, This is the logical deviation coefficient.

[0167] If the student's problem-solving logic state is classified as "logical confusion" and the historical ability benchmark value is greater than the preset high score threshold, it indicates that the student's state is fluctuating abnormally, and the problem-solving logic state is judged to be an atypical error.

[0168] If the problem-solving logic state level is logically confused and the historical ability benchmark value is less than the preset low score threshold, it indicates that the student's ability deficiency has become entrenched, and the problem-solving logic state is judged to be a typical knowledge blind spot.

[0169] It should be noted that, based on the statistical analysis of the distribution of logical deviation coefficients from a large number of historical test questions, a first logical deviation threshold and a second logical deviation threshold were obtained. These two thresholds are used to classify the calculated continuous logical deviation coefficients into discrete levels of problem-solving logic states. The first logical deviation threshold ranges from 0.2 to 0.3, and the second logical deviation threshold ranges from 0.6 to 0.7.

[0170] It should be noted that, in order to ensure that the range of values ​​for the state stability index is normalized, Within the interval, set . The default value range is 0.6 to 0.7. Because... It is generated based on the cumulative data of students' multiple responses within a preset time window, and has a high statistical confidence level. Therefore, it is given a high weight and represents the basic level of students' abilities. The default value range is 0.3 to 0.4. This reflects the logical quality of the solution in this single attempt; a lower weight can prevent errors caused by a single accidental mistake (such as carelessness). An increase in the stability index leads to a sharp drop, while maintaining sensitivity to serious logical errors. This is achieved through calculation... This transforms logical deviation (a negative indicator) into logical correctness (a positive indicator), thereby enabling [the system to] [follow the principle of logical correctness]. (Positive indicators) are weighted and summed. If The value is significantly lower than This indicates that the current performance is significantly hindering stability, and the system will trigger an abnormal state warning accordingly.

[0171] It should be noted that the preset high-score threshold and preset low-score threshold are determined based on the statistical distribution of the average mastery score in history for the subject. These preset thresholds serve as baselines for determining whether a student's historical ability benchmark in their individual ability profile falls within the high- or low-level range. The preset high-score threshold is set at 0.8 (corresponding to 80% mastery). A score above this threshold indicates a solid foundation in history; if logical inconsistencies occur, it is more likely to be an atypical error or a fluctuation in performance. The preset low-score threshold is set at 0.5 (corresponding to 50% mastery). A score below this threshold indicates a weak foundation in history; if logical inconsistencies occur, it is more likely to be a typical knowledge gap or a entrenched ability deficiency.

[0172] This invention, based on qualitative diagnosis of dominant difference items and longitudinal analysis mechanism based on state stability index, can not only output specific diagnostic conclusions on logical jumps, redundant thinking, causal inversion and circular reasoning according to the magnitude of various penalty items, but also accurately distinguish atypical mistakes (strong ability but poor state) and typical knowledge blind spots (weak ability and disordered logic) by combining historical ability benchmark values ​​and weighted index of current performance. This provides a deep decision-making basis for achieving precise individualized teaching and tutoring.

[0173] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for evaluating student problem-solving status based on knowledge graphs, characterized in that, Includes the following steps: Step S1: Perform entity recognition and semantic analysis on the test question text data to obtain knowledge point text data. Use the knowledge point text data as knowledge point nodes to construct a problem-solving knowledge graph. Obtain the standard problem-solving directed path through the problem-solving knowledge graph. Step S2: Obtain students' historical answer data and map it into the problem-solving knowledge graph. Calculate the students' mastery scores of knowledge points and construct individual ability profiles of students based on the mastery scores. Step S3: Obtain the actual answer data of students to the test questions, and generate the corresponding directed acyclic path of student cognition in the problem-solving knowledge graph based on the actual answer data; Step S4: Construct an edit distance matrix based on the student's cognitive directed acyclic path and the standard problem-solving directed path; obtain the optimal edit path through the edit distance matrix and calculate the logical deviation coefficient. Step S5: Determine the student's logical state level in solving the test questions by using the logical deviation coefficient, and evaluate the logical state in solving the questions in combination with the student's personal ability profile.

2. The student problem-solving status assessment method based on knowledge graph as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Sequence labeling of test question text data in the question bank is performed through bidirectional long short-term memory network and conditional random field, knowledge point text data is extracted, and a unique identifier code is assigned to each knowledge point text data. The knowledge point text data is input into the pre-trained language model, and the knowledge point text vector corresponding to the knowledge point text data is output. Step S102: Perform semantic dependency parsing on the test question text data, identify the relationship between the dominant and subordinate words in the question stem logic of the knowledge point text data, take the knowledge point text data as knowledge point nodes, and take the relationship between the dominant and subordinate words as the edges between the knowledge point text data to construct a problem-solving knowledge graph.

3. The student problem-solving status assessment method based on knowledge graph as described in claim 2, characterized in that, Step S1 further includes the following sub-steps: Step S103: The standard solution text of the test question is segmented into ordered atomic solution text segments according to punctuation marks and logical connectors, and the ordered atomic solution text segments are used to generate text feature vectors through a pre-trained language model; Step S104: Calculate the first cosine similarity between the text feature vector and the knowledge point text vector, select knowledge point nodes with a first cosine similarity greater than a first preset similarity threshold as mapping nodes, and connect the mapping nodes according to the original order of the atomic problem-solving text segments to generate a standard problem-solving directed path.

4. The student problem-solving status assessment method based on knowledge graph as described in claim 3, characterized in that, Step S2 includes the following sub-steps: Step S201: Obtain the student's historical answer data within a preset time window, and parse the unique identifier code of the test text data, the cumulative number of answers, the answer score rate, and the answer timestamp of the historical answer data; Based on the unique identifier code of the test question text data, obtain the knowledge point text data and the assigned unique identifier code corresponding to the test question text data, and use the unique identifier code of the corresponding knowledge point text data to query the problem-solving knowledge graph to obtain the knowledge point node corresponding to the historical answer data. Step S202: Calculate the time difference between the answer timestamp and the real-time timestamp. Using the time difference and a preset forgetting rate constant, calculate the memory retention coefficient corresponding to the historical answer data. The mathematical expression for the memory retention coefficient is: ； in, For memory retention coefficient, This is the time difference. This is a preset forgetting rate constant.

5. The student problem-solving status assessment method based on knowledge graph as described in claim 4, characterized in that, Step S2 further includes the following sub-steps: Step S203: Locate the student's historical answer data related to the knowledge points, obtain the difficulty coefficient of the corresponding questions in the historical answer data, and calculate the student's mastery score for the knowledge points based on the answer score rate, memory retention coefficient, difficulty coefficient, and cumulative number of answers in the historical answer data. The mathematical expression for the mastery score is: ； in, For the first The score for mastery of each knowledge point. To accumulate the number of answers, For the first The answer score rate of the historical answer data. For the first The memory retention coefficient of historical answer data. For the first The difficulty level of the historical answer data; Step S204: According to the numerical order of the unique identifier encoding, the mastery scores of all knowledge point nodes are encapsulated into one-dimensional feature vectors, and all one-dimensional feature vectors are combined into a student's personal ability profile.

6. The student problem-solving status assessment method based on knowledge graph as described in claim 5, characterized in that, Step S3 includes the following sub-steps: Step S301: Obtain students' actual answer data for the test questions, use regular expressions to identify explicit segmentation symbols in the answer data, and use a natural language processing model to identify logical connectors that represent deductive relationships in the actual answer data. Use the explicit segmentation symbols and logical connectors as segmentation points to divide the actual answer data into student problem-solving sub-steps arranged in time sequence. Use a pre-trained language model to extract semantic features of student problem-solving sub-steps, and generate corresponding sub-step feature vectors through the semantic features. Step S302: Traverse all knowledge point nodes in the problem-solving knowledge graph, calculate the second cosine similarity between the feature vector of the sub-step and the text vector of the knowledge point, and for each student's problem-solving sub-step, select the knowledge point node whose second cosine similarity is greater than the second preset similarity threshold and output it as the cognitive mapping node corresponding to the student's problem-solving sub-step.

7. The student problem-solving status assessment method based on knowledge graph as described in claim 6, characterized in that, Step S3 further includes the following sub-steps: Step S303: Map the student's problem-solving sub-steps to nodes in the problem-solving knowledge graph. For different student problem-solving sub-steps mapped to the same knowledge point node, normalize them to the same cognitive mapping node in the problem-solving knowledge graph. Based on the original time sequence of students' problem-solving sub-steps in the actual answer data, identify the cognitive mapping nodes corresponding to adjacent students' problem-solving sub-steps, and establish directed connection edges from the cognitive mapping nodes corresponding to the previous students' problem-solving sub-steps to the cognitive mapping nodes corresponding to the subsequent students' problem-solving sub-steps. When the cognitive mapping node corresponding to the subsequent student's problem-solving sub-step is the same as the cognitive mapping node corresponding to the previous student's problem-solving sub-step, the directed connection edge forms a backward edge pointing to the cognitive mapping node corresponding to the previous student's problem-solving sub-step, generating a student cognitive path graph. Step S304: The student cognitive path graph is subjected to logical loop detection using a depth-first search algorithm. During the logical loop detection process, a recursive path stack that records the current traversal path is maintained in real time. When a directed connection edge pointing to a node that already exists in the recursive path stack is encountered, the directed connection edge is determined to be a backward edge and confirmed as a logical loop. Extract all backward edges that form a logical closed loop and mark them as cyclic logical feature edges. Count the number of cyclic logical feature edges and mask the cyclic logical feature edges in the topology to generate a directed acyclic path for student cognition.

8. The student problem-solving status assessment method based on knowledge graph as described in claim 7, characterized in that, Step S4 includes the following sub-steps: Step S401: Extract the nodes and corresponding logical pointing features in the student's cognitive directed acyclic path, and serialize them into a student logical unit sequence according to the generation timestamp; extract the nodes and corresponding logical pointing features in the standard problem-solving directed path, and serialize them into a standard logical unit sequence; construct an edit distance matrix based on the student logical unit sequence and the standard logical unit sequence. Step S402: Starting from the beginning position of the edit distance matrix, the cumulative edit cost of each position in the edit distance matrix is ​​calculated using a dynamic programming algorithm with preset node insertion cost, node deletion cost, and edge direction reversal cost, until the end position of the edit distance matrix is ​​reached.

9. The student problem-solving status assessment method based on knowledge graph as described in claim 8, characterized in that, Step S4 further includes the following sub-steps: Step S403: Take the connected path with the minimum cumulative editing cost as the optimal editing path, and count the number of redundant nodes, missing nodes and reverse edges in the optimal editing path. The number of redundant nodes is the number of times the node deletion operation is performed in the optimal editing path; The number of missing nodes is the number of times the node insertion operation is performed in the optimal editing path; The number of reverse edges is the number of operations in the optimal editing path that involve nodes with the same but opposite logical pointing characteristics. Step S404: Obtain the total number of standard nodes in the directed path of the standard solution. Calculate the logical deviation coefficient using the total number of standard nodes, the number of redundant nodes, the number of missing nodes, the number of cyclic logical feature edges, and the number of reverse edges. The mathematical expression for the logical deviation coefficient is: ； in, The logical deviation coefficient, The number of redundant nodes. This represents the number of missing nodes. This represents the number of reverse edges. The number of cyclic logical feature edges. The preset redundancy penalty weight, The default penalty weight is the missing value. The preset reverse order penalty weight, The preset cyclic penalty weights, The total number of standard nodes, As a redundancy penalty item, For missing penalty items, This is a reverse order penalty term. This is a cyclical penalty term.

10. The student problem-solving status assessment method based on knowledge graph as described in claim 9, characterized in that, Step S5 includes the following sub-steps: Step S501: Compare the logic deviation coefficient with the logic deviation threshold. The logic deviation threshold includes a first logic deviation threshold and a second logic deviation threshold, wherein the first logic deviation threshold is less than the second logic deviation threshold. If the logical deviation coefficient is less than the first logical deviation threshold, it indicates that the student's logical state level in solving the test question is logically rigorous. If the logical deviation coefficient is greater than the first logical deviation threshold and less than the second logical deviation threshold, it indicates that the student's logical state level in solving the test question is general. If the logical deviation coefficient is greater than the second logical deviation threshold, it indicates that the student's logical state of solving the test questions is logically confused. Step S502: When the problem-solving logic state level is determined to be logically general or logically chaotic, the values ​​of the redundancy penalty term, the missing penalty term, and the reverse order penalty term are calculated respectively. Compare the values ​​of the redundancy penalty term, the missing penalty term, and the reverse order penalty term, and take the one with the largest value as the dominant difference term; If the dominant difference term is a redundant penalty term, it indicates that the student's problem-solving logic is redundant, suggesting that the student has invalid derivation steps. If the dominant difference item is a missing penalty item, it indicates that the student's problem-solving logic is in a logical leap, suggesting that the student is missing a key deduction step; If the dominant difference term is a reverse order penalty term, it indicates that the student's problem-solving logic is reversed, suggesting that the student's reasoning direction is contrary to standard logic; If the dominant difference term is a circular penalty term, it indicates that the student's problem-solving logic is in a circular argument, suggesting that the student has a logical dead loop where each element is mutually causal. Step S503: Extract the mastery score corresponding to the knowledge point node from the student's personal ability profile, calculate the average value of the mastery score and use it as the historical ability benchmark value, calculate the weighted sum of the historical ability benchmark value and the logical deviation coefficient to obtain the state stability index. If the problem-solving logic state level is logical confusion and the historical ability benchmark value is greater than the preset high score threshold, it indicates that the student's state is fluctuating abnormally, and the problem-solving logic state is determined to be an atypical error. If the problem-solving logic state level is logically confused and the historical ability benchmark value is less than the preset low score threshold, it indicates that the student's ability deficiency is solidified, and the problem-solving logic state is determined to be a typical knowledge blind spot.

Images