An english knowledge confusion entity pair identification and bundling recommendation method based on answer modification behavior sequence

Abstract

This invention discloses a method for identifying and recommending English knowledge confusion entities based on question-answering modification behavior sequences, belonging to the field of intelligent education and learning behavior analysis technology. The method includes: collecting learners' modification behavior sequences during question-answering and constructing time-series data; extracting modification behavior features using a bidirectional long short-term memory network to identify hesitation, repeated modifications, and self-correction patterns, and establishing associations with English knowledge points; constructing a confusion entity association strength matrix based on the features, introducing a time decay mechanism to calculate cognitive association degree, and forming a confusion entity knowledge graph; identifying confusion entity groups and generating bundle sets through a community detection algorithm; and dynamically optimizing the recommendation order and difficulty gradient using reinforcement learning, combined with learners' historical data. This invention can accurately identify knowledge confusion relationships and improve the targeting and adaptability of learning path recommendations.

Description

Technical Field

[0001] This invention relates to the field of intelligent education and learning behavior analysis technology, and in particular to a method for identifying and recommending English knowledge-confusing entity pairs based on question-answering modification behavior sequences. Background Technology

[0002] With the rapid development of intelligent educational technology and learning analytics, data-driven personalized learning methods have gradually become an important research direction in the field of foreign language teaching. Especially in English learning, learners' modification behaviors during the answering process (such as repeatedly changing answers, delaying submission, and hesitating) are considered to reflect their cognitive state and knowledge mastery. In recent years, the integration of technologies such as knowledge tracing, learning path recommendation, and knowledge graph construction has led to a shift from static answer result analysis to dynamic behavioral process analysis. However, existing technologies mostly focus on statistical modeling of answer accuracy or final answer results, lacking sufficient mining of fine-grained behavioral data (especially modification behavior sequences) during the answering process, making it difficult to deeply characterize learners' confusion relationships at the knowledge point level. Furthermore, in English learning, many knowledge points share semantic similarities or usage similarities (such as vocabulary and grammatical structures), making learners prone to "knowledge confusion." Existing methods lack a modeling mechanism for the relationships between confused knowledge entities, resulting in recommended content often being discretely distributed, lacking relevance and coherence. Meanwhile, existing recommendation mechanisms mostly employ static or weakly dynamic strategies, failing to adaptively optimize based on learners' real-time behavioral changes, thus making it difficult to meet the needs of personalized and precise learning.

[0003] CN119988730B discloses a personalized learning path recommendation system for English majors. This system collects learner behavioral characteristic data and image data, and combines emotion recognition and age analysis to construct a learning path recommendation model, improving the personalized recommendation effect from a multimodal data perspective. While this method expands the learner's state perception dimension to some extent, its core still relies on macro-level behavioral characteristics and emotion recognition results. It lacks in-depth exploration of fine-grained behaviors during the answering process (such as modification trajectory, modification frequency, and dwell time), and fails to construct a knowledge cognition model based on behavioral evolution. Furthermore, this method does not involve modeling the confusion relationships between knowledge points, and cannot identify the sources of learners' cognitive biases between specific knowledge entities. Therefore, it is difficult to achieve aggregated processing and collaborative optimization of easily confused knowledge points during the recommendation path generation process.

[0004] CN115329959B discloses a learning objective recommendation method based on a two-stream knowledge embedding network. By constructing a Q-matrix and combining a knowledge tracking model and a knowledge demand model, it achieves probabilistic prediction and path recommendation of learning objectives. This method has certain advantages in knowledge representation and recommendation accuracy, but it primarily models based on the sequence of answer results, lacking the utilization of dynamic behavioral information during the answering process, especially failing to model and analyze learners' cognitive fluctuations during answer modification. Furthermore, this method focuses on the probability ranking of single knowledge concepts, neglecting the confusion relationships between knowledge points and their structured representation, and failing to construct a relational graph between confused entities, resulting in fragmented recommendation results when facing complex knowledge networks. In addition, the recommendation mechanism of this method is mainly based on predicted probability ranking, lacking dynamic optimization capabilities based on adaptive mechanisms such as reinforcement learning, making it difficult to achieve continuous iterative adjustments to the recommendation strategy as learners' behavior changes. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] In view of the problems of existing personalized English learning recommendation technologies, such as insufficient utilization of answer modification behavior, weak ability to identify knowledge confusion relationships, and weak dynamic adaptability of recommendation strategies, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to accurately identify confusing entity pairs between English knowledge points based on the sequence of modification behaviors during the learner's answering process, and construct knowledge binding relationships with inherent correlation to optimize the learning path.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for identifying and bundling English knowledge-confusing entity pairs based on question-answering modification behavior sequences, comprising, Collect learners' modification behavior sequences during the English test-taking process to form time-series test-taking behavior data; A bidirectional long short-term memory network is used to extract features from the modification behavior sequence, identify the hesitation pattern, repeated modification pattern and self-correction pattern in the modification behavior, and establish an association mapping with the preset English knowledge points; Based on the extracted modification behavior features, a confusion entity association strength matrix is ​​constructed. By setting a time decay weight coefficient, the cognitive association degree between confusion entity pairs is calculated, forming a confusion entity knowledge graph. A community detection algorithm is used to perform cluster analysis on the confusing entity knowledge graph to identify groups of confusing entities with inherent cognitive relationships and generate a set of confusing entity bundles. Based on learners' confused entity bundles and historical learning data, reinforcement learning algorithms are used to dynamically adjust the presentation order and difficulty gradient of recommended content to form an optimized learning path recommendation scheme.

[0009] Compared with existing technologies, the beneficial effects of this invention are as follows: By collecting the sequence of learners' modification behaviors during the question-answering process, it transforms them into time-series question-answering behavior data, preserving key time and sequence information for subsequent cognitive diagnosis; a bidirectional long short-term memory network is used to automatically identify cognitive patterns such as hesitation, repeated modifications, and self-correction from this sequence, and map them to specific knowledge points, achieving accurate tracing of the causes of confusion from behavior to the knowledge layer; on this basis, a confusion entity association strength matrix with time decay weights is constructed to form a confusion entity knowledge graph that dynamically reflects the current cognitive state, making the estimation of confusion associations both cumulative and timely; the community detection algorithm is used to perform cluster analysis on this graph, automatically generating a set of confusion entity bundles with inherent cognitive associations, reducing the dimensionality of the complex network to an interventionizable cognitive group; reinforcement learning algorithms are used to dynamically adjust the recommendation order and difficulty gradient based on the bundles and historical learning data, generating an optimized learning path recommendation scheme, significantly improving the accuracy of English confusion knowledge point identification and the adaptability and efficiency of learning path recommendation. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a method for identifying and bundling English knowledge-confusing entity pairs based on question-answering behavior sequences. Detailed Implementation

[0011] 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.

[0012] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0013] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0014] As mentioned in the background section, existing technologies are unable to fully extract fine-grained behavioral information during the question-answering process, nor can they effectively model the cognitive confusion relationships between knowledge points, resulting in insufficient targeting and coherence of recommendation results. This invention provides a method for identifying and bundling English knowledge confusion entity pairs based on question-answering modification behavior sequences.

[0015] Reference Figure 1 , Figure 1 This is a flowchart illustrating a method for identifying and bundling English knowledge-confusing entity pairs based on question-answering behavior sequences, according to an embodiment of the present invention. Figure 1 As shown, a method for identifying and bundling English knowledge-confusing entity pairs based on question-answering modification behavior sequences includes: S1: Collect the sequence of learners’ modification behaviors during the English test-taking process to form time-series test-taking behavior data. The sequence of modification behaviors includes modification timestamp, content before modification, content after modification, and duration of modification. S2: A bidirectional long short-term memory network is used to extract features from the modification behavior sequence, identify the hesitation pattern, repeated modification pattern and self-correction pattern in the modification behavior, and establish an association mapping with the preset English knowledge points; S3: Based on the extracted modification behavior features, construct the confusion entity association strength matrix, calculate the cognitive association degree between confusion entity pairs by setting the time decay weight coefficient, and form a confusion entity knowledge graph; S4: The community detection algorithm is used to perform cluster analysis on the confusing entity knowledge graph to identify groups of confusing entities with inherent cognitive relationships and generate a set of confusing entity bundles; S5: Based on the learner's confused entity bundle set and historical learning data, a reinforcement learning algorithm is used to dynamically adjust the presentation order and difficulty gradient of recommended content to form an optimal learning path recommendation scheme.

[0016] In this embodiment of the application, step S1 includes: S1.1: Obtain the original operation log data of the target learner during the English quiz from the behavior log database of the learning terminal. The original operation log data shall include at least the operation type field, operation content field, operation timestamp field, question type field, question unique identifier field, and question correct answer text field. S1.2: Based on the original operation log data, filter out the original operation log data with the operation type of modifying the answer, and take each modified answer operation log data obtained after filtering as a candidate modification behavior record; S1.3: Extract the operation timestamp field of the candidate modification behavior record as the modification timestamp data corresponding to this modification behavior, the answer text before modification in the operation content field as the content data before modification, and the answer text after modification in the operation content field as the content data after modification. S1.4: Calculate the difference between the modification timestamp data corresponding to the candidate modification behavior record and the timestamp data of the previous operation record in the same question-answering process, and use the difference as the modification dwell time data corresponding to this modification behavior; S1.4.1: Obtain the modification timestamp data corresponding to each candidate modification behavior record, as well as the complete timestamp sequence of all operation records in the answer session to which this candidate modification behavior record belongs; S1.4.2: For each candidate modification behavior record, locate the operation order corresponding to the candidate modification behavior record in the complete timestamp sequence, and extract the timestamp data of the previous operation record corresponding to the previous operation order. It should be noted that the previous operation record includes, but is not limited to, the question loading operation, option selection operation, answer submission operation, or previous round of modification operation in the same answering process; when the candidate modification behavior record is the first operation record in the current answering session, and there is no previous operation record, the modification dwell time data of this candidate modification behavior record is marked as invalid and discarded. S1.4.3: Subtract the timestamp data of the previous operation record from the modification timestamp data of the candidate modification behavior record, and use the original time difference value as the initial modification dwell time data of this candidate modification behavior record; S1.4.4: Perform a validity check on the initial modified dwell time data, following the order below: If the initial modification dwell time data is less than 0 milliseconds or greater than the preset reasonable upper limit threshold, the modification dwell time data of this candidate modification behavior record will be marked as abnormal and discarded. If the initial modification dwell time data is less than the preset minimum threshold, the modification dwell time data of this candidate modification behavior record will be marked as invalid and discarded. The initial modified stay duration data that passes the above two checks will be used as the valid modified stay duration data, and this valid modified stay duration data will be associated and stored with the corresponding candidate modified behavior record; Preferably, the preset reasonable upper limit threshold is determined based on the average answering time per question in English test-taking scenarios. According to the question type field obtained in S1.1, different upper limit thresholds are set for different question types such as multiple choice, fill-in-the-blank, and reading comprehension. For example, the upper limit threshold for multiple choice is set to 120 seconds, and the upper limit threshold for reading comprehension is set to 300 seconds. The preset minimum threshold is set as follows: if the answering terminal is detected as a touch screen tablet, the preset minimum threshold is set to 0.8 seconds; if the answering terminal is detected as a physical keyboard computer, the preset minimum threshold is set to 0.3 seconds; if the terminal type cannot be detected, the default value is set to 0.5 seconds. Each candidate modification behavior record and its corresponding valid modification duration data are used as input data for step S1.5; if all candidate modification behavior records are marked as invalid during a certain question-answering process, the subsequent analysis of that question-answering process is skipped. S1.5: Map the content data before modification and the content data after modification to the preset English knowledge point tag library respectively to obtain the first knowledge point tag corresponding to the content before modification and the second knowledge point tag corresponding to the content after modification. Furthermore, the content data before and after modification are mapped to a preset English knowledge point tag library to obtain the first knowledge point tag corresponding to the content before modification and the second knowledge point tag corresponding to the content after modification; the mapping is implemented using a fuzzy matching algorithm based on edit distance. S1.5.1: Obtain the pre-modification content data and post-modification content data corresponding to each candidate modification behavior record, where the pre-modification content data is the answer text before modification; the post-modification content data is the answer text after modification. S1.5.2: Read all knowledge point tags and their corresponding standard expressions from the preset English knowledge point tag library; Furthermore, the pre-defined English knowledge point tag library contains multiple knowledge point tags. Each knowledge point tag consists of a standard tag name, at least one synonym variant string, and an identifier of the grammar category to which the knowledge point belongs. The English knowledge point tag library is stored in the form of a structured data table, containing the following fields: unique tag identifier, standard tag name, list of synonym variants, and grammar category. The standard tag name adopts the standardized terminology in the teaching syllabus, such as the simple present tense, present perfect tense, passive voice, relative clauses, modal verbs can / could, transitive verbs, and present participles of non-finite verbs, etc. Specifically, each knowledge point tag corresponds to at least one synonym variant string to support fuzzy matching based on edit distance in S1.5.3; for example: for the standard tag name present simple tense, its synonym variant list includes: simplepresent, presentsimple, do / does, third person singular, and present simple tense; for the standard tag name present perfect tense, its synonym variant list includes: presentperfect, have / hasdone, have / has + past participle, present perfect tense, and already / ever / never; for the standard tag name passive voice, its synonym variant list includes: passivevoice, be + past participle, passive voice, and passive voice; for the standard tag name relative clause, its synonym variant list includes: relativeclause, attributeclause, and clauses introduced by that / which / who; for the standard tag name modal verb, its synonym variant list includes: modalverb, can / could / may / might / must, and modal verb usage; Furthermore, grammatical category identifiers are used to classify knowledge points in a coarse-grained manner. Grammatical categories include, but are not limited to, tense, voice, clause, non-finite verb, morphology, and sentence structure. For example, the simple present tense and the present perfect tense are both classified as tenses; the passive voice is classified as a voice; and relative clauses and noun clauses are classified as clauses. In an optional embodiment, the English knowledge point tag library also supports dynamic expansion: when fuzzy matching based on edit distance fails to match a high-frequency content data before or after modification to any existing knowledge point tag multiple times in a row, this high-frequency content data is added as a new synonym variant to the corresponding knowledge point tag, or the administrator reviews and adds a new knowledge point tag. S1.5.3: The fuzzy matching algorithm based on edit distance is used to calculate the edit distance between the original content data and the standard expression of each knowledge point tag in the preset English knowledge point tag library in turn; Preferably, the edit distance is the Levenstein distance, which calculates the minimum number of single-character editing operations required to convert one string into another; single-character editing operations include insertion, deletion, and replacement; to accommodate the differences in capitalization in English expressions, the standard expressions of the original content data and knowledge point tags are converted to lowercase before calculating the edit distance. S1.5.4: The knowledge point label corresponding to the minimum value of all edit distances is used as the candidate matching knowledge point label of the content data before modification, and the normalized edit distance value is recorded as the matching confidence. The normalized edit distance value is equal to the edit distance divided by the maximum value of the length of the standard expression string of the content data before modification and the knowledge point label. Furthermore, if the ratio between the minimum value and the string length of the content data before modification is less than the preset matching threshold, the match is considered successful; otherwise, the match is considered unsuccessful. The knowledge point tag corresponding to this content data before modification is marked as unknown, and manual annotation or subsequent semantic model supplementary matching is triggered. The preset matching threshold is set to 0.7 by default. S1.5.5: Following the same method as S1.5.3 and S1.5.4, perform fuzzy matching based on edit distance on the modified content data to obtain candidate matching knowledge point labels and corresponding matching confidence scores for the modified content data; S1.5.6: The candidate matching knowledge point tags obtained by successfully matching the content data before modification are used as the first knowledge point tags, and the candidate matching knowledge point tags obtained by successfully matching the content data after modification are used as the second knowledge point tags. Preferably, when the content data before or after modification matches multiple knowledge point tags and the minimum edit distance is the same, the knowledge point tag with the highest frequency in the preset English knowledge point tag library is selected as the final matching result. S1.5.7: Associate and store the first knowledge point tag and the second knowledge point tag with the modification timestamp data and modification dwell time data in the corresponding candidate modification behavior record, as a component of the encapsulated time-series answer behavior data entries in step S1.6; In an optional embodiment, for matching results with a matching confidence level lower than a preset high confidence threshold (e.g., 0.9), the original content data before or after modification is retained as text features for semantic enhancement processing by the bidirectional long short-term memory network in the subsequent S2 step.

[0017] S1.6: Encapsulate the modification timestamp data, the content data before modification, the content data after modification, the modification dwell time data, the first knowledge point tag, and the second knowledge point tag into a time-seriesd answer behavior data entry; S1.7: Sort all the answer behavior data entries generated by the same target learner in the same answer session according to the time sequence of the modified timestamp data to form the modification behavior sequence of this target learner.

[0018] In this embodiment of the application, step S2 includes: S2.1: Construct and train the basic architecture of the bidirectional long short-term memory network, which includes a forward long short-term memory layer, a backward long short-term memory layer, and a splicing layer; Specifically, the forward long short-term memory layer reads the input sequence in forward chronological order and outputs the forward hidden state vector at each time step; the backward long short-term memory layer reads the same input sequence in reverse chronological order and outputs the backward hidden state vector; the concatenation layer concatenates the forward hidden state vector and the backward hidden state vector at the same time step along the feature dimension to form a bidirectional fused feature vector for this time step. S2.1.1: Obtain the training dataset of the bidirectional long short-term memory network, wherein the training dataset consists of multiple training samples, each training sample including a sequence of modified behavior vectors and the pattern label corresponding to each time step in the sequence; It should be noted that the pattern labels include hesitant pattern labels, repeated modification pattern labels, self-correction pattern labels, and ordinary behavior labels; the rules for constructing the training dataset are as follows: When the modified dwell time data exceeds the preset hesitation time threshold and the knowledge point tags corresponding to the content data before modification are the same as those corresponding to the content data after modification, it is marked as a hesitation mode tag. When three or more consecutive modification behaviors are under the same unique identifier for the same question, and the knowledge point tags corresponding to the content data before modification and the knowledge point tags corresponding to the content data after modification switch back and forth between two fixed tags, then each modification behavior in this group of modification behaviors is marked as a repeated modification mode tag. When the knowledge point tag corresponding to the content data before modification is not equal to the knowledge point tag corresponding to the correct answer of the question, and the knowledge point tag corresponding to the content data after modification is equal to the knowledge point tag corresponding to the correct answer of the question, and the modification dwell time data is within the preset reasonable time range, it is marked as a self-correction mode tag. Other situations are labeled as ordinary behavior. The preset hesitation duration threshold is determined based on the median of all modified dwell times in historical data, with a default value of 3 seconds. The lower limit of the preset reasonable duration range is consistent with the preset minimum threshold in S1.4.4 (dynamically determined according to terminal type, with a default value of 0.5 seconds), and the upper limit is consistent with the preset reasonable upper limit threshold in S1.4.4 (with a default value of 120 seconds, adjustable according to question type). S2.1.2: Train the bidirectional long short-term memory network by sequentially inputting the sequence of modified behavior vectors from the training dataset into the bidirectional long short-term memory network to obtain the bidirectional fused feature vector at each time step, and then sending it into the classification output layer. Furthermore, the classification output layer uses the Sigmoid activation function to output the probability value of belonging to each type of pattern label at each time step. The categories are independent of each other, and no normalization constraint is applied. The bidirectional fused feature vector of each time step output by the bidirectional long short-term memory network is used as the input to the classification output layer, which is a fully connected layer with four output nodes. The four output nodes correspond to the ordinary behavior category, the hesitant mode category, the repeated modification mode category, and the self-correction mode category, respectively. The fully connected layer performs a linear transformation on the input bidirectional fused feature vector: multiplying the bidirectional fused feature vector by a pre-stored weight matrix and adding a bias vector to obtain four original output scores, each output score corresponding to a category; The four original output scores are input into the Softmax function, which performs an exponential operation on each original output score and normalizes it, so that the values ​​of the four output scores are converted into probability values ​​between zero and one and the sum of the four probability values ​​is one. It should be noted that the formula for calculating the flexible maximum transfer function is as follows: for the k-th category, its output probability is equal to the exponent of the original output score of the category with the natural constant as the base, divided by the sum of the exponents of the original output scores of the four categories; the classification output layer outputs four probability values ​​for the modification behavior at this time step, which belong to the ordinary behavior category, the hesitant mode category, the iterative modification mode category, and the self-correction mode category, respectively, as the mode label probability distribution for this time step; Furthermore, the cross-entropy loss function is used to calculate the loss value between the predicted probability distribution and the true pattern label, and the network parameters are updated using the time backpropagation algorithm. For each time step, the four probability values ​​output by the classification output layer are used as the predicted probability distribution, and the real mode label corresponding to that time step is obtained; the real mode label is converted into a four-dimensional one-hot vector, where the value of the real mode label at the corresponding position in the one-hot vector is 1, and the value of the other positions is 0. The cross-entropy loss function is used to calculate the loss value between the predicted probability distribution and the one-hot vector: for each class position, the true label value (0 or 1) is multiplied by the natural logarithm of the predicted probability, and then the product of all classes is summed and the negative value is taken to obtain the single-step loss value at that time step. The total loss value for a training sample is obtained by summing the single-step loss values ​​for all time steps in the training sample; the batch average loss value is obtained by summing the total loss values ​​of all training samples and dividing by the total number of training samples. The loss value is used to guide the update of network parameters: the gradient of the loss value with respect to each network parameter is calculated layer by layer along the time expansion direction of the bidirectional long short-term memory network using the backpropagation algorithm. The network parameters include the input gate weight matrix, forget gate weight matrix, output gate weight matrix and corresponding bias vector in the forward long short-term memory layer, the same type of parameters in the backward long short-term memory layer and the weight matrix and bias vector of the classification output layer. Based on the gradient value, the Adam optimizer is used to adjust all network parameters at a preset initial learning rate of 0.001 to reduce the loss value generated in the next forward calculation. The above process of forward calculation, loss calculation, gradient calculation and parameter update is repeated until the preset training 50 epochs are reached or the validation set loss value does not decrease for 5 consecutive epochs, at which point the parameter update stops. Specifically, after training is completed, the parameters of the classification output layer are used for classifier splitting in the subsequent step S2.1.3. During training, a four-class Softmax is used, but during inference, it is changed to three independent binary classifiers. This conversion is achieved through parameter remapping and does not affect the feature extraction capability of the trained classifier. S2.1.3: The trained classification output layer is split into three independent classifiers, which include a first classifier, a second classifier, and a third classifier; S2.2: Convert the answer behavior data entries of the modified behavior sequence into numerical vectors to obtain the modified behavior vector sequence to be identified: Normalization is performed on the modification dwell time data in the answer behavior data entries, mapping the modification dwell time data to a numerical range of 0 to 1 to obtain a normalized dwell time value. The normalization operation adopts the minimum and maximum value normalization method. The global minimum and global maximum values ​​of all modification dwell time data in the historical answer log are pre-calculated. The current modification dwell time data is subtracted from the global minimum value and then divided by the difference between the global maximum value and the global minimum value. When the global maximum value equals the global minimum value, all modification dwell time data are normalized to 0.5. In each time-series question-answering behavior data entry, the first knowledge point label is converted into a first embedding vector, and the second knowledge point label is converted into a second embedding vector. This conversion is performed using a pre-trained knowledge point embedding matrix. The dimension of the knowledge point embedding matrix is ​​the total number of knowledge point labels multiplied by a preset embedding dimension, with each row corresponding to the embedding vector of one knowledge point label. The row corresponding to the first knowledge point label is found in the knowledge point embedding matrix, and its vector is extracted as the first embedding vector. Similarly, the row vector corresponding to the second knowledge point label is extracted as the second embedding vector. The knowledge point embedding matrix is ​​optimized along with other network parameters during the training of the bidirectional long short-term memory network. The normalized dwell time value, the first embedding vector and the second embedding vector are concatenated to form a one-dimensional numerical vector, which is then used as the input feature vector for this time-series answer behavior data item. The concatenation order is as follows: first place the normalized dwell time value (scalar), then place each component of the first embedding vector in sequence, and finally place each component of the second embedding vector; if the preset embedding dimension is 64 dimensions, then the total dimension of the input feature vector is 129 dimensions, that is, 1+64+64. According to the original time order of the sequential answer behavior data items in the modified behavior sequence, the input feature vectors corresponding to each data item are arranged sequentially to form an input feature vector sequence. This input feature vector sequence is used as the modified behavior vector sequence to be identified. The sequence of modified behavior vectors to be identified is stored in a temporary buffer and passed to step S2.3 as input data for the bidirectional long short-term memory network. After generating the sequence of modified behavior vectors to be identified, the sequence length is uniformly processed: when the sequence length is less than the maximum time step preset by the bidirectional long short-term memory network, the end of the sequence is padded with all zeros to the maximum length; when the sequence length is greater than the maximum time step, a sliding window is used to truncate the nearest maximum length subsequence. S2.3: Input the sequence of modified behavior vectors to be identified into the trained bidirectional long short-term memory network, and the bidirectional long short-term memory network outputs the bidirectional fused feature vector at each time step in sequence; S2.4: Input the bidirectional fused feature vector into the classifier to identify the pattern label; S2.4.1: Use the bidirectional fused feature vector as the input feature of the current time step, and pass it to the first classifier, the second classifier and the third classifier at the same time; Specifically, after receiving the bidirectional fused feature vector, the first classifier performs the following calculations: It performs a dot product operation between the bidirectional fused feature vector and the first weight row vector stored internally by the first classifier to obtain the first dot product result; it adds the first dot product result to the first bias value stored internally by the first classifier to obtain the original score of the hesitant mode; it inputs the original score of the hesitant mode into the first flexible maximum transfer function, where the first flexible maximum transfer function calculates the exponent value of the original score of the hesitant mode with the natural constant as the base, then divides this exponent value by one and the sum of the exponent values, outputting a value between zero and one, and using the exponent value as the probability value that the modified behavior belongs to the hesitant mode; Furthermore, after receiving the bidirectional fused feature vector, the second classifier performs the following calculations: It performs a dot product operation between the bidirectional fused feature vector and the second weight row vector stored internally by the second classifier to obtain a second dot product result; it adds the second dot product result to the second bias value stored internally by the second classifier to obtain the original score of the iterative modification mode; and it inputs the original score of the iterative modification mode into the second flexible maximum transfer function, where the second flexible maximum transfer function uses the same exponential normalization calculation method as the first flexible maximum transfer function, and outputs a value between zero and one. This value is used as the probability value that the modification behavior belongs to the iterative modification mode. Furthermore, after receiving the bidirectional fused feature vector, the third classifier performs the following calculations: It performs a dot product operation between the bidirectional fused feature vector and the third weight row vector stored internally by the third classifier to obtain the third dot product result; it adds the third dot product result to the third bias value stored internally by the third classifier to obtain the original score of the self-correction mode; and it inputs the original score of the self-correction mode into the third flexible maximum transfer function, which uses the same exponential normalization calculation method and outputs a value between zero and one. This value is used as the probability value that the modification behavior belongs to the self-correction mode. It should be noted that the first flexible maximum transfer function, the second flexible maximum transfer function, and the third flexible maximum transfer function are independent of each other. Each of them only transforms the original score of the current classifier, and there is no constraint relationship between the three output probability values ​​that they add up to one. S2.4.2: The probability values ​​of the hesitation mode, the repeated modification mode, and the self-correction mode are compared with the preset mode discrimination thresholds respectively: If all three probability values ​​are less than the preset pattern discrimination threshold, the modification behavior corresponding to this time step will be identified as a normal behavior and no pattern label will be assigned. If at least one probability value is greater than or equal to the preset pattern discrimination threshold, then the pattern with the highest probability value is selected as the pattern label for the modification behavior. If two or three probability values ​​are equal and all are greater than or equal to the preset mode discrimination threshold, the mode label is determined according to the priority order of self-correction mode, repeated modification mode, and hesitation mode. That is, self-correction mode is selected first, followed by repeated modification mode, and finally hesitation mode. Preferably, the preset mode discrimination threshold is preset to 0.5. The preset mode discrimination threshold is not fixed and can be adjusted according to the false positive rate requirements of subsequent confusion identification tasks: when it is necessary to improve the coverage of confusion detection, the threshold is reduced to 0.3; when it is necessary to reduce false alarms, the threshold is increased to 0.7. S2.4.3: Associate the identified pattern label with the modification behavior corresponding to the current time step, and output the modification behavior record with the pattern label; S2.5: Associate and map the pattern label with the first knowledge point label and the second knowledge point label corresponding to this time step to generate a knowledge point pair mapping record with the pattern label; In this embodiment of the application, step S3 includes: S3.1: Extract the first knowledge point label, the second knowledge point label, and the pattern label from the knowledge point pair mapping record with pattern labels; S3.1.1: Read the storage list of knowledge point pair mapping records with pattern labels, wherein each record in the storage list contains at least four fields: modification timestamp data, first knowledge point label, second knowledge point label, and pattern label; S3.1.2: Traverse each record in the storage list, and extract the first knowledge point tag field value of the record as the first knowledge point tag, extract the second knowledge point tag field value of the record as the second knowledge point tag, and extract the pattern tag field value of the record as the pattern tag. It should be noted that the values ​​of the pattern label are limited to four types: hesitant pattern label, repeated modification pattern label, self-correction pattern label, and ordinary behavior label. S3.1.3: Determine if the pattern label is a normal behavior label: If the pattern label is equal to a normal behavior label, skip this record and do not use it for subsequent construction of the confusion entity association strength matrix; if the pattern label is any of the hesitant pattern label, repeatedly modified pattern label, or self-correcting pattern label, retain all three extracted fields of this record (first knowledge point label, second knowledge point label, pattern label) and the associated modification timestamp data. Furthermore, for each retained record, the correct answer knowledge point label is mapped to the text of the correct answer to the question obtained in S1. If the first knowledge point label is not equal to the correct answer knowledge point label and the second knowledge point label is equal to the correct answer knowledge point label, then the first knowledge point label of the record is marked as an incorrect knowledge point label and the second knowledge point label is marked as a correct knowledge point label; otherwise, no correctness labeling is performed. This labeling information is stored in association with the record and used in step S5.2 to enhance the state representation of the reinforcement learning algorithm. S3.2: Based on the type of pattern label, assign an initial confusion contribution weight value to the mapping record for each knowledge point with a pattern label; Preferably, the initial confusion contribution weight is set based on the following criteria: the self-correction mode reflects learners actively correcting incorrect answers to correct answers, which is the strongest evidence of confusion; the repeated modification mode reflects learners switching back and forth between two knowledge points, with moderate confusion uncertainty; and the hesitant mode reflects learners' hesitation on the same knowledge point but without switching knowledge points, with weaker evidence of confusion. Furthermore, obtain the pattern label for each retained record. If the pattern label is a hesitant pattern label, assign an initial obfuscation contribution weight of 0.3 to the record; if the pattern label is a repeatedly modified pattern label, assign an initial obfuscation contribution weight of 0.8 to the record; if the pattern label is a self-correcting pattern label, assign an initial obfuscation contribution weight of 1.0 to the record. In an optional embodiment, the initial confusion contribution weight value is further adjusted according to the learner's historical answer accuracy: when the learner's historical answer accuracy is higher than 90%, the initial confusion contribution weight value corresponding to the self-correction mode is reduced to 0.6; when the learner's historical answer accuracy is lower than 50%, the initial confusion contribution weight value corresponding to the hesitation mode is increased to 0.5. Specifically, the modification timestamp data of each knowledge point pair mapping record with pattern label is obtained. Based on the time interval between the modification timestamp data and the current system time, the time decay weight coefficient is calculated. The modification timestamp data is read from each record retained in step S3.1.3, and the current system time at the time of execution of step S3 is also read. The time interval between the current system time and the modification timestamp data is calculated. The time interval is in days, and the calculation formula is: the time interval (days) is equal to the date value corresponding to the current system time minus the date value corresponding to the modification timestamp data. Furthermore, the time decay weighting coefficient is calculated using an exponential decay function: the time decay weighting coefficient is equal to e raised to the power of (the negative preset decay rate parameter multiplied by the time interval). It should be noted that the date value corresponding to the current system time and the date value corresponding to the modified timestamp data are both represented by the number of days accumulated since the year 1 AD; if the calculation result is less than 1 hour, the time interval will be taken as 1 hour and uniformly converted into days for subsequent calculations; the preset decay rate parameter is set to 0.05; the base in the exponential decay function is the natural constant e, which is approximately taken as 2.71828; Furthermore, the exponential decay function is executed through the following steps: first, the product of the preset decay rate parameter and the time interval is calculated; second, the negative of the product is calculated; and finally, the exponent is raised to the power of the negative of the product with the natural constant e as the base. Specifically, the time decay weight coefficient is associated with the corresponding record and stored. If the time interval is greater than 30 days, the time decay weight coefficient is less than 0.22. In this case, the coefficient can be rounded to three decimal places before storage. In an optional embodiment, a linear decay function is used instead of an exponential decay function: the time decay weighting coefficient is equal to one minus the linear decay rate multiplied by the time interval, and takes the value of 0 when the calculation result is less than 0; the linear decay rate is set to 0.02. S3.3: Multiply the initial confusion contribution weight value by the time decay weight coefficient to obtain the final confusion contribution value of each knowledge point with pattern label to the mapping record; For each record, read the initial obfuscation contribution weight value allocated in step S3.2 and the time decay weight coefficient calculated in step S3.2; perform a multiplication operation on the initial obfuscation contribution weight value and the time decay weight coefficient, and use the product as the final obfuscation contribution value for that record; associate and store the final obfuscation contribution value with the first knowledge point tag, the second knowledge point tag, and the modification timestamp data in the corresponding record, and use it as the input data for step S3.4; It should be noted that the final confusion contribution value ranges from zero to one; when the time decay weighting coefficient is zero, the final confusion contribution value is also zero. S3.4: Using the first knowledge point label and the second knowledge point label as row index and column index, construct a confusion entity association strength matrix with an initial value of zero, traverse all knowledge point pair mapping records with pattern labels, and accumulate the final confusion contribution value of each record to the corresponding position in the confusion entity association strength matrix; S3.4.1: Obtain the total number of all knowledge point tags in the preset English knowledge point tag library, denoted as the total number of tags N, create an n-row n-column confusion entity association strength matrix M, and initialize all elements in the matrix to zero; Specifically, the row and column indices of the confusion entity association strength matrix M are arranged in ascending order according to the unique identifiers of the knowledge point tags in the preset English knowledge point tag library; S3.4.2: Obtain all records output from step S3.3, find the corresponding unique identifier in the preset English knowledge point tag library based on the first knowledge point tag, and use this unique identifier as the row index i; find the corresponding unique identifier based on the second knowledge point tag, and use this unique identifier as the column index j; S3.4.3: The final confusion contribution value recorded is added to the element in the i-th row and j-th column of the confusion entity association strength matrix M, that is, the element in the i-th row and j-th column is equal to the original value of the element plus the final confusion contribution value; It should be noted that the accumulation operation supports multiple records being repeatedly accumulated at the same index position. The value of the element in the i-th row and j-th column after accumulation represents the total confusion intensity from the first knowledge point label i to the second knowledge point label j. When the first knowledge point label and the second knowledge point label in a record are the same, the accumulation operation is accumulated to the element in the i-th row and i-th column, that is, the diagonal element. S3.4.4: After traversing all records, store the confusion entity association strength matrix M in memory or persistent storage; S3.5: Normalize the confusion entity association strength matrix M to generate a normalized confusion association strength matrix; S3.5.1: Copy the dimension information N and all element values ​​of the obfuscated entity association strength matrix M, create a normalized obfuscated association strength matrix with the same dimension, and initialize all elements of the normalized obfuscated association strength matrix to zero; S3.5.2: For each row i of the confusion entity association strength matrix, calculate the sum of all elements in this row, denoted as the sum of elements in the i-th row S_i; S3.5.3: Determine if the sum of elements in the i-th row, S_i, is greater than zero: If the sum of elements in the i-th row, S_i, is greater than zero, then for each column j in that row, calculate that the element in the i-th row and j-th column of the normalized confusion association strength matrix is ​​equal to the element in the i-th row and j-th column of the confusion entity association strength matrix divided by the sum of elements in the i-th row, S_i; if the sum of elements in the i-th row, S_i, is equal to zero, then keep all elements in this row as zero; It should be noted that row normalization makes the sum of the element values ​​in each row equal to one (when the sum of the elements in the i-th row is greater than zero), and the element values ​​of the normalized confusion association strength matrix are in the range of zero to one. S3.5.4: Store the calculated normalized confusion association strength matrix, where the elements in the normalized confusion association strength matrix represent the proportion of learners who changed the first knowledge point label i to other knowledge points and changed it to the second knowledge point label j among all the modification behaviors. S3.6: Based on the normalized confusion association strength matrix, calculate the cognitive association degree between any two different knowledge point tags; S3.6.1: Obtain the normalized confusion association strength matrix and the total number of knowledge point tags in the preset English knowledge point tag library; Specifically, for all knowledge point label pairs that satisfy i < j, where i and j represent unique identifiers for two distinct knowledge point labels, the following calculation is performed: Read the element in the i-th row and j-th column from the normalized confusion association strength matrix as the directed confusion strength from knowledge point label i to knowledge point label j, and read the element in the j-th row and i-th column as the directed confusion strength from knowledge point label j to knowledge point label i. Determine whether the directed confusion intensity from knowledge point label i to knowledge point label j and the directed confusion intensity from knowledge point label j to knowledge point label i are both greater than zero: If both are greater than zero, then the cognitive relevance is calculated as two times (the directed confusion intensity from knowledge point label i to knowledge point label j multiplied by the directed confusion intensity from knowledge point label j to knowledge point label i) divided by (the directed confusion intensity from knowledge point label i to knowledge point label j plus the directed confusion intensity from knowledge point label j to knowledge point label i). If only one of the two directed confusion intensities is greater than zero while the other is equal to zero, then the cognitive association degree is set to zero. If both directional confusion intensities are equal to zero, then the cognitive association degree is set to zero. It should be noted that cognitive relevance is calculated using the harmonic mean formula, and its value ranges from zero to one; cognitive relevance is greater than zero only when the directional confusion intensity in both directions is non-zero. Preferably, in order to suppress low-association noise, a square root transformation is performed on the calculated cognitive association degree: the transformed cognitive association degree is equal to the square root of the cognitive association degree; In an optional embodiment, the cognitive relevance is calculated using the Jaccard similarity coefficient: the feature vector of each knowledge point label i is defined as the i-th row of the normalized confusion association strength matrix, and the Jaccard similarity coefficient of the two feature vectors is calculated as the cognitive relevance. It should be noted that each pair of knowledge point tags and their corresponding cognitive relevance are stored as a relevance record; S3.7: Construct a confusing entity knowledge graph using knowledge point tags as nodes and cognitive relevance as edge weights; S3.7.1: Create an empty set of nodes, an empty set of edges, and an empty set of weights. Add all the knowledge point tags from the preset English knowledge point tag library as nodes to the node set. S3.7.2: Traverse the correlation records and preset the correlation threshold; Preferably, the preset association threshold is determined based on the distribution of all non-zero cognitive associations, and is set to the upper quartile of all non-zero cognitive associations. Specifically, if the cognitive relevance of the relevance record is greater than the preset relevance threshold, then an undirected edge is added between the node corresponding to identifier i and the node corresponding to identifier j in the node set, this edge is added to the edge set, and the weight value of this edge is set as the cognitive relevance and added to the weight set. S3.7.3: Combine the node set, edge set, and weight set into a confused entity knowledge graph, and store the confused entity knowledge graph in the form of an adjacency list; Furthermore, the adjacency list takes the form of: for each node in the node set, a list of neighboring nodes is stored, where the list of neighboring nodes contains the identifiers of all neighboring nodes connected to this node through the edge set and the corresponding edge weight values. Furthermore, each node in the obfuscated entity knowledge graph is accompanied by additional information: the frequency of occurrence of the sub-knowledge point tag obtained from the statistics of step S1, and the average modification dwell time of this knowledge point tag obtained from the sum of the normalized dwell time values ​​in step S2. It should be noted that the obfuscated entity knowledge graph is used as input to the community discovery algorithm in step S4, where only edges with cognitive relevance greater than a preset relevance threshold are retained. In this embodiment of the application, step S4 includes: S4.1: Obtain the set of nodes, the set of edges, and the weight value of each edge in the obfuscated entity knowledge graph; Specifically, the node set and edge set are read from the stored adjacency list data. Each node in the node set corresponds to a knowledge point label, and each node is accompanied by a unique identifier, standard name, frequency of occurrence, and average modification dwell time for that knowledge point label. Each edge in the edge set connects two nodes, and each edge is accompanied by a weight value, which is the cognitive association degree calculated in step S3.7. The edges in the confused entity knowledge graph are undirected edges, and the weight value of the edge represents the strength of the bidirectional cognitive association between the two knowledge point labels. S4.2: The Leuven community discovery algorithm is used to initialize the confused entity knowledge graph, assigning an independent community identifier to each node and treating each node as an initial community; Furthermore, obtain the total number of nodes in the node set, denoted as node number K; assign initial community identifiers to the nodes in the node set according to the order of their unique identifiers, where the initial community identifier is the same as the node's unique identifier, that is, each node constitutes a community on its own. Furthermore, the initial modularity of the confused entity knowledge graph is calculated, where the initial modularity is calculated as follows: the modularity is equal to the sum of [(the sum of the weights of the edges inside community i divided by the sum of the weights of all edges) minus (the sum of the degrees of the nodes inside community i divided by the square of twice the sum of the weights of all edges)] of all communities i; the sum of the weights of all edges is the sum of the weight values ​​of all edges in the edge set. S4.3: Perform the first phase of community optimization traversal, and try to move each node to the community where its neighboring nodes are located. Calculate the modularity gain before and after the move. If the modularity gain is positive, then perform the move operation. Specifically, for the current node in the node set, obtain all neighboring nodes of the current node, where neighboring nodes are other nodes directly connected to the current node through the edge set; obtain the first community identifier to which the current node currently belongs, and the second community identifier to which each neighboring node belongs; for the second community identifier, if the second community identifier is different from the first community identifier, simulate moving the current node from the first community to the second community, calculate the difference between the modularity after the move and the modularity before the move, and record it as the modularity gain; Furthermore, the formula for calculating the modularity gain is: Modularity gain equals [(the sum of the weights of edges within the target community plus the sum of the weights of edges between the current node and nodes in the target community) divided by twice the sum of the weights of all edges minus ((the square of the total degree of nodes in the target community plus the degree of the current node) divided by four times the square of the sum of the weights of all edges)] minus [(the sum of the weights of edges within the target community divided by twice the sum of the weights of all edges) minus ((the square of the total degree of nodes in the target community divided by four times the square of the sum of the weights of all edges) minus (the degree of the current node divided by twice the square of the sum of the weights of all edges))]; where the sum of the weights of all edges is the sum of the weight values ​​of all edges in the edge set; It should be noted that the degree of the current node is the sum of the weights of all edges connected to the current node; Furthermore, among all the second communities corresponding to neighboring nodes, the second community with the largest modularity gain is selected as the candidate target community; if the largest modularity gain is greater than zero, the community identifier of the current node is updated to the candidate target community identifier; if the largest modularity gain is less than or equal to zero, the community identifier of the current node remains unchanged. Specifically, the above steps are repeated in the order of the nodes in the node set until all nodes have been traversed and no community identifier has changed in a complete traversal, then the first phase of community optimization traversal ends. S4.4: Perform the second phase of community aggregation operation, merge all nodes with the same community identifier after the first phase optimization into super nodes, and construct a compressed obfuscated entity knowledge graph; S4.4.1: Obtain the final community identifier of each node after the first stage of optimization, and divide nodes with the same final community identifier into the same node group, with each node group corresponding to a super node; S4.4.2: Assign a new supernode identifier to the supernode, where the supernode identifiers are numbered sequentially starting from 1; S4.4.3: Calculate the weight of the internal edge of the supernode in the compressed confused entity knowledge graph: For each supernode, sum up the original edge weights between all nodes inside it to obtain the weight of the supernode's self-loop edge, where the weight of the self-loop edge represents the degree of cohesion within this community. S4.4.4: Calculate the edge weights between supernodes in the compressed confused entity knowledge graph: For any two different supernodes, sum the original edge weights between any node in the first supernode and any node in the second supernode to obtain the edge weights between the two supernodes. It should be noted that if there is no existing edge between two supernodes, the edge weight between these two supernodes is zero. S4.5: Determine if the iteration termination condition is met: If the module degree increment of the compressed obfuscated entity knowledge graph relative to the module degree of the previous iteration is less than the preset convergence threshold, then terminate the iteration and output the final community partitioning result; otherwise, use the compressed obfuscated entity knowledge graph as the new obfuscated entity knowledge graph and repeat steps S4.3 and S4.4. S4.5.1: Calculate the modularity of the compressed obfuscated entity knowledge graph in the current round, where the method for calculating the modularity is the same as step S4.2; S4.5.2: Calculate the difference between the modularity of the current round and the modularity of the previous round, and record it as the modularity increment; S4.5.3: Compare the module degree increment with the preset convergence threshold: If the module degree increment is less than the preset convergence threshold, terminate the iteration; if the module degree increment is greater than or equal to the preset convergence threshold, use the compressed obfuscated entity knowledge graph as the new obfuscated entity knowledge graph, return to step S4.3 and re-execute the community optimization traversal of the first stage. Preferably, the preset convergence threshold is determined based on the convergence accuracy requirement of the modularity increment after multiple iterations. That is, when the change in modularity of community division is less than one ten-thousandth, the clustering result is considered to be stable, which can effectively avoid invalid iterations. It is set to 0.0001. S4.6: Based on the compressed obfuscated entity knowledge graph, extract the knowledge point tags corresponding to all nodes in the community to form obfuscated entity groups, and calculate the intra-group cohesion and inter-group separation of each obfuscated entity group. S4.6.1: Obtain the final community identifier of the node after the iteration terminates, and collect the knowledge point tags corresponding to all nodes with the same final community identifier into the same set, with each set as a group of confused entities; S4.6.2: For a group of confused entities, calculate the cohesion within the group: the cohesion within the group is equal to the sum of the weights of all edges within the group divided by the maximum number of edges that can exist in all nodes within the group; the maximum number of edges is equal to the number of nodes in the group multiplied by the number of nodes in the group minus one, and then divided by two. S4.6.3: For a confused entity group, calculate the inter-group separation degree between the group and other external groups: the inter-group separation degree is equal to the sum of the weights of all edges between the nodes in the group and the nodes in the external groups, divided by the maximum number of possible edges between the nodes in the group and all external nodes. Furthermore, when the cohesion within a group is greater than a first preset threshold and the separation between groups is less than a second preset threshold, the group of confused entities is marked as a strongly confused bundled group; otherwise, it is marked as a weakly confused bundled group. It should be noted that higher cohesion within a group and lower separation between groups indicate a stronger internal cognitive association and a sparser external association within the confused entity group. The first preset threshold is determined based on the empirical distribution of cohesion within a large number of learners' confused entity groups—in actual data, the internal edge density of a strong confused group is usually not less than 0.3; if it is lower than this value, the internal association within the group is weak. The second preset threshold is determined based on the empirical distribution of separation between groups, that is, the connection strength between a strong confused group and external nodes is usually less than 0.1; if it is higher than this value, the group boundary is blurred and it is not suitable as an independent bundled set. S4.7: Sort the obfuscated entity groups in descending order of the number of nodes within the group to generate an obfuscated entity bundle set; Preferably, the obfuscated entity bundle includes the group identifier, knowledge point tag list, intra-group cohesion, and inter-group separation for each obfuscated entity group; S4.7.1: Assign a unique group identifier to each obfuscated entity group, with group identifiers numbered sequentially starting from 1; S4.7.2: For each obfuscated entity group, extract the standard names of the knowledge point tags of all nodes in the group, arrange them in order of the unique identifiers of the knowledge point tags, and form a knowledge point tag list; S4.7.3: Encapsulate the group identifier, knowledge point tag list, group cohesion, group separation, and the identifiers of strongly obfuscated or weakly obfuscated bundled groups into a single obfuscated entity bundled record. S4.7.4: Sort all obfuscated entity bundle records in descending order of the number of nodes in the group, and use the sorted bundle record set as the obfuscated entity bundle set; It should be noted that confusing entity groups with a large number of nodes indicate that learners have a wide range of confused understandings of the knowledge points in that group, and should be given priority in the recommended bundle processing. In an optional embodiment, for records marked as weakly obfuscated bundles, their recommendation priority is reduced or they are split into smaller subgroups for separate processing in subsequent step S5. In this embodiment of the application, step S5 includes: S5.1: Obtain the bundled set of confusing entities and the historical learning data extracted from the learner's historical answer logs, where the historical learning data includes the learner's historical answer accuracy, historical answer time, and learning feedback on the recommended content completed for each knowledge point tag; Preferably, each record in the obfuscated entity bundle set includes a group identifier, a list of knowledge point tags, the cohesion within the group, the separation between groups, and a strong obfuscated bundle group identifier or a weak obfuscated bundle group identifier. Historical answer accuracy rate is the ratio of the number of questions answered correctly under this knowledge point tag to the total number of questions answered under this knowledge point tag; historical answer time is the average dwell time of all answer operations under this knowledge point tag, in seconds; learning feedback includes the learner's completion status (completed or incomplete) of the content after receiving the recommended content, as well as the accuracy rate of the immediate test after completion. S5.2: Construct the state space and action space of the reinforcement learning algorithm, and design the reward function of the reinforcement learning algorithm; It should be noted that the state space consists of the current learning state vector of each confused entity group in the confused entity bundle set; the current learning state vector of each confused entity group includes the group's intra-group cohesion, inter-group separation, the average historical answer accuracy of each knowledge point in the group, the average historical answer time of each knowledge point in the group, and the correctness index of the knowledge points in the group obtained from the statistics of the correct / incorrect knowledge tags marked in step S3.1; where, for each knowledge point tag in the group, the correctness index = (the number of times the knowledge point is marked as a correct knowledge tag) / (the number of times the knowledge point is marked as a correct knowledge tag + the number of times it is marked as an incorrect knowledge tag + 1), and is set to 0.5 by default when the total number is zero; Specifically, extract the cohesion values ​​within the confused entity groups and the separation values ​​between groups from the confused entity bundle set; based on the knowledge point tag list of the confused entity groups, extract the historical answer accuracy rate of each knowledge point tag within this confused entity group from the historical answer accuracy rate of the knowledge points obtained in step S5.1, calculate the arithmetic mean of the historical answer accuracy rates, and record it as the group average accuracy rate; extract the historical answer time of each knowledge point tag within this group from the historical answer time of the knowledge points, calculate the arithmetic mean of the historical answer time, and record it as the group average time. Furthermore, the four values ​​of intra-group cohesion, inter-group separation, average group accuracy, and average group time are concatenated in a fixed order to form the current learning state vector of this confused entity group; the current learning state vectors of all confused entity groups are combined according to the group identifier order to form the current state of the reinforcement learning algorithm. It should be noted that the action space includes actions to adjust the recommended order of each group of obfuscated entities in the obfuscated entity bundle set and actions to adjust the difficulty gradient of knowledge points within each group of obfuscated entities: The recommended order adjustment action is defined as follows: based on the current sorting result, perform a swap operation on two adjacent confused entity groups. Each action only affects the local sorting structure, thereby constraining the original factorial action space to a linear scale. To reduce the complexity of the action space, a hierarchical action strategy is adopted: select the preferred group identifier, select the secondary recommended group identifier, and so on. Each action only adjusts the relative order of two adjacent groups. The difficulty gradient adjustment action is defined as follows: For each group of confused entities, three difficulty gradients are set according to the cognitive complexity of the knowledge point tags from low to high. The three difficulty gradients are the basic difficulty gradient, the medium difficulty gradient, and the advanced difficulty gradient. The cognitive complexity is determined based on the average correlation between the knowledge point and other knowledge points in the cognitive correlation calculated by S3.6. The lower the average correlation, the lower the complexity. Each difficulty gradient corresponds to a set of recommended content presentation parameters: the recommended content presentation order for the basic difficulty gradient is knowledge point definition and typical example sentences; the recommended content presentation order for the medium difficulty gradient is comparison and analysis and fill-in-the-blank exercises; and the recommended content presentation order for the advanced difficulty gradient is contextual application and error correction test. Furthermore, a reward function for the reinforcement learning algorithm is designed, wherein the reward function calculates the instant reward value based on the learner's instant test accuracy of the recommended content, the time taken to complete the recommended content, and the change in cohesion within the confused entity group. Specifically, obtain the learner's immediate test accuracy after completing the recommended content in the current round, denoted as immediate test accuracy, where the value of immediate test accuracy ranges from zero to one; obtain the total time spent by the learner to complete the recommended content in the current round, denoted as recommendation completion time; The time taken to complete the recommendation is compared with the preset standard reference time, and the time coefficient is calculated: Time coefficient = Standard reference time / Recommendation completion time, and the upper limit of the standard reference time coefficient is one when the recommendation completion time is less than the standard reference time. The standard reference time is 300 seconds by default. The change in the cohesion of each obfuscated entity group in the obfuscated entity bundle set before and after the execution of the current recommendation action is obtained. The average value of the change in the cohesion of all obfuscated entity groups is calculated and recorded as the average cohesion gain. The change in cohesion is obtained by comparing the cohesion of the group calculated in step S4 before the current round of recommendation with the cohesion of the group calculated in step S4 before the previous round of recommendation. If there is no data for the previous round, the gain is zero. The instant reward value is calculated: Instant reward value = Instant test accuracy × Time coefficient + Average cohesion gain. The higher the instant test accuracy, the shorter the recommendation completion time, and the greater the average cohesion gain, the higher the instant reward value. S5.3: A deep Q-network is used as the core decision model of the reinforcement learning algorithm. The current state is used as the input of the deep Q-network, and the estimated cumulative reward value corresponding to each action is output. The action with the maximum estimated cumulative reward value is selected as the current recommendation strategy. S5.3.1: Construct a deep Q-network, which includes an input layer, a fully connected hidden layer, and an output layer. The number of neurons in the input layer is equal to the dimension of the current state, and the number of neurons in the output layer is equal to the size of the action space. S5.3.2: Input the current state into the deep Q network, where the deep Q network outputs the estimated cumulative reward value for each action in the action space after forward propagation; The ε-greedy strategy is used to select actions: an action is randomly selected from the action space with probability ε, and the action with the largest estimated cumulative reward value is selected with probability 1 minus ε. The initial value of ε is set to 1.0, and it gradually decreases to 0.01 in an exponential decay manner as the number of training rounds increases. S5.3.4: The selected action is used as the current recommendation strategy, where the current recommendation strategy includes the specific values ​​of the action to adjust the recommendation order and the action to adjust the difficulty gradient; S5.4: Based on the current recommendation strategy, retrieve the corresponding learning content from the recommendation content library, adjust the actions according to the recommendation order to determine the presentation order of the confused entity group, adjust the actions according to the difficulty gradient to determine the corresponding recommended content difficulty gradient for the knowledge points within the confused entity group, and generate a learning path recommendation scheme; S5.4.1: Based on the recommended order adjustment action, obtain the order list of the first obfuscated entity group identifier, the second obfuscated entity group identifier, and the subsequent group identifiers that need to be presented first; S5.4.2: Based on the confused entity group, determine the difficulty gradient level (basic difficulty gradient, medium difficulty gradient, or advanced difficulty gradient) of the knowledge point tag list corresponding to the group according to the difficulty gradient adjustment action. S5.4.3: Retrieve learning content from the recommended content library that meets the following conditions: the knowledge point tags of the learning content belong to the knowledge point tag list of the current confused entity group, and the difficulty level of the learning content matches the difficulty gradient level specified by the difficulty gradient adjustment action; S5.4.4: Arrange the retrieved learning content in a sequential list to form a learning path recommendation scheme; the learning path recommendation scheme is stored in the form of structured data, including a recommendation order list, a confusion entity group identifier corresponding to each recommendation position, a list of knowledge point tags in the group, and a list of corresponding learning content identifiers; S5.5: The system pushes the learning path recommendation scheme to the learner's terminal device, records the learner's completion status and real-time test accuracy for each recommended content in the learning path recommendation scheme, and stores the completion status and real-time test accuracy as new historical learning data in the historical learning database; the system updates the historical learning data immediately after each learning session to ensure that the reinforcement learning algorithm status at the start of the next session reflects the latest learning progress; the system sets a periodic recommendation interval (default is 24 hours), and unless the learner actively requests it, the next round of recommendations will be carried out according to the preset interval; Specifically, the learning path recommendation scheme is displayed on the learner's terminal device in the form of an interactive interface. The interactive interface presents the learning content corresponding to each obfuscated entity group in the order of recommendation. The learner's interaction with each recommended content is monitored. If the learner completes the learning of a recommended content, the completion status of the recommended content is recorded as "completed" and an instant test question pops up. If the learner skips or does not complete the recommended content, the completion status is recorded as "incomplete". Furthermore, the learner obtains the instant test accuracy after completing each recommended content, and associates and stores the completion status and instant test accuracy with the corresponding knowledge point tags and recommendation timestamps to form new historical learning data; the new historical learning data is written into the historical learning database, and the historical answer accuracy and historical answer time of each knowledge point tag maintained in step S5.1 are updated so that the status of the next round of reinforcement learning algorithm reflects the learner's latest cognitive level; In an optional embodiment, if the learner's instant test accuracy for a certain confusing entity group is higher than 90% in three consecutive rounds of recommendation, then this confusing entity group is removed from the confusing entity bundle set and will no longer participate in subsequent recommendations; if the learner's instant test accuracy for a certain confusing entity group is lower than 40% in three consecutive rounds of recommendation, then the difficulty gradient level of the confusing entity group is reduced and the recommendation frequency of the group is increased.

[0019] In summary, this invention collects learners' modification behavior sequences during the question-answering process and transforms them into temporal-sequential question-answering behavior data, preserving crucial time and sequence information for subsequent cognitive diagnosis. A bidirectional long short-term memory network is used to automatically identify cognitive patterns such as hesitation, repeated modifications, and self-correction from this sequence and map them to specific knowledge points, achieving precise tracing of the causes of confusion from behavior to the knowledge level. Based on this, a confusion entity association strength matrix with time decay weights is constructed, forming a confusion entity knowledge graph that dynamically reflects the current cognitive state, making the estimation of confusion associations both cumulative and timely. A community detection algorithm is used to cluster this graph, automatically generating a bundle of confusion entities with inherent cognitive associations, reducing the dimensionality of the complex network to an interventionizable cognitive group. A reinforcement learning algorithm is used to dynamically adjust the recommendation order and difficulty gradient based on the bundles and historical learning data, generating an optimized learning path recommendation scheme, significantly improving the accuracy of English confusion knowledge point identification and the adaptability and efficiency of learning path recommendation.

[0020] 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 scope of the claims of the present invention.

Claims

1. A method for identifying and recommending English knowledge-confused entity pairs based on question-answering modification behavior sequences, characterized in that, include: Collect learners' modification behavior sequences during the English test-taking process to form time-series test-taking behavior data; A bidirectional long short-term memory network is used to extract features from the modification behavior sequence, identify the hesitation pattern, repeated modification pattern and self-correction pattern in the modification behavior, and establish an association mapping with preset English knowledge points; Based on the extracted modification behavior features, a confusion entity association strength matrix is ​​constructed. By setting a time decay weight coefficient, the cognitive association degree between confusion entity pairs is calculated, forming a confusion entity knowledge graph. The community detection algorithm is used to perform cluster analysis on the confused entity knowledge graph to identify confused entity groups with inherent cognitive relationships and generate confused entity bundle sets; Based on the learner's confused entity bundle set and historical learning data, a reinforcement learning algorithm is used to dynamically adjust the presentation order and difficulty gradient of recommended content to form an optimal learning path recommendation scheme.

2. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 1, characterized in that, The formation of the optimized learning path recommendation scheme includes: Obtain the obfuscated entity bundle set and historical learning data extracted from learners' historical answer logs; Construct the state space and action space of the reinforcement learning algorithm, and design the reward function of the reinforcement learning algorithm; A deep Q-network is used as the core decision model of the reinforcement learning algorithm. The current state is used as the input of the deep Q-network, and the estimated cumulative reward value corresponding to each action is output. The action with the maximum estimated cumulative reward value is selected as the current recommendation strategy. Based on the current recommendation strategy, the corresponding learning content is retrieved from the recommendation content library, the actions are adjusted according to the recommendation order to determine the presentation order of the confused entity groups, the actions are adjusted according to the difficulty gradient to determine the corresponding recommendation content difficulty gradient for the knowledge points within the confused entity groups, and a learning path recommendation scheme is generated.

3. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 2, characterized in that, The state space consists of the current learning state vector of each group of confused entities in the confused entity bundle set; the action space includes actions to adjust the recommended order of each group of confused entities in the confused entity bundle set and actions to adjust the difficulty gradient of knowledge points within each group of confused entities.

4. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 3, characterized in that, The obfuscated entity bundle includes: Obtain the set of nodes, the set of edges, and the weight value of each edge in the obfuscated entity knowledge graph; The Leuven community detection algorithm is used to initialize the confused entity knowledge graph, assigning an independent community identifier to each node and treating each node as an initial community. The first phase of community optimization traversal is performed, in which each node is attempted to move to the community where its neighboring nodes are located, and the modularity gain before and after the move is calculated. If the modularity gain is positive, the move operation is performed. The second phase of community aggregation is performed, merging all nodes with the same community identifier after the first phase of optimization into super nodes, and constructing a compressed obfuscated entity knowledge graph. Based on the compressed obfuscated entity knowledge graph, the knowledge point tags corresponding to all nodes in the community are extracted to form obfuscated entity groups, and the intra-group cohesion and inter-group separation of each obfuscated entity group are calculated. The obfuscated entity groups are sorted from largest to smallest according to the number of nodes within the group to generate an obfuscated entity bundle set.

5. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 4, characterized in that, The compressed obfuscated entity knowledge graph includes: If the module degree increment of the compressed obfuscated entity knowledge graph relative to the module degree increment of the previous iteration is less than a preset convergence threshold, the iteration is terminated and the final community partitioning result is output; otherwise, the compressed obfuscated entity knowledge graph is used as a new obfuscated entity knowledge graph.

6. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 5, characterized in that, The obfuscated entity knowledge graph includes: Extract the first knowledge point label, the second knowledge point label, and the pattern label from the knowledge point pair mapping record with pattern labels; Based on the type of the pattern label, an initial confusion contribution weight value is assigned to the mapping record for each knowledge point with the pattern label; Multiply the initial confusion contribution weight value by the time decay weight coefficient to obtain the final confusion contribution value of each knowledge point with pattern label to the mapping record; Using the first knowledge point label and the second knowledge point label as row index and column index, a confusion entity association strength matrix with an initial value of zero is constructed. All knowledge point pair mapping records with pattern labels are traversed, and the final confusion contribution value of each record is accumulated to the corresponding position in the confusion entity association strength matrix. The confusion entity association strength matrix M is normalized to generate a normalized confusion association strength matrix; Based on the normalized confusion association strength matrix, calculate the cognitive association degree between any two different knowledge point tags; Using knowledge point tags as nodes and the cognitive relevance as the edge weights, a confusing entity knowledge graph is constructed.

7. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 6, characterized in that, The knowledge point pair mapping record with pattern labels includes: Construct the basic architecture for training a bidirectional long short-term memory network; Convert the answer behavior data entries of the modified behavior sequence into numerical vectors to obtain the modified behavior vector sequence to be identified. The sequence of modified behavior vectors to be identified is input into a trained bidirectional long short-term memory network, and the bidirectional long short-term memory network outputs the bidirectional fused feature vector for each time step in sequence. The bidirectional fused feature vector is input into the classifier to identify the pattern label; The pattern label is associated and mapped with the first knowledge point label and the second knowledge point label corresponding to this time step to generate a knowledge point pair mapping record with the pattern label.

8. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 7, characterized in that, The bidirectional long short-term memory network includes a forward long short-term memory layer, a backward long short-term memory layer, and a concatenation layer. The forward long short-term memory layer reads the input sequence in forward chronological order and outputs the forward hidden state vector at each time step. The backward long short-term memory layer reads the same input sequence in reverse chronological order and outputs the backward hidden state vector. The concatenation layer concatenates the forward hidden state vector and the backward hidden state vector at the same time step along the feature dimension to form a bidirectional fused feature vector for that time step.

9. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 7, characterized in that, The modified behavior vector sequence includes: Obtain the original operation log data of the target learner during the English test-taking process from the behavior log database of the learning terminal; Based on the original operation log data, filter out the original operation log data with the operation type of modifying the answer, and take each modified answer operation log data obtained after filtering as a candidate modification behavior record; Extract the operation timestamp field of the candidate modification behavior record as the modification timestamp data corresponding to this modification behavior, the answer text before modification in the operation content field as the content data before modification, and the answer text after modification in the operation content field as the content data after modification. The difference between the modification timestamp data corresponding to the candidate modification behavior record and the timestamp data of the previous operation record in the same question-answering process is calculated, and the difference is used as the modification dwell time data corresponding to this modification behavior. The original content data and the modified content data are respectively mapped to a preset English knowledge point tag library to obtain the first knowledge point tag corresponding to the original content and the second knowledge point tag corresponding to the modified content. The modification timestamp data, the content data before modification, the content data after modification, the modification dwell time data, the first knowledge point tag, and the second knowledge point tag are collectively encapsulated into a time-seriesd question-answering behavior data entry; Based on the chronological order of the modified timestamp data, all the question-answering behavior data entries generated by the same target learner within the same question-answering session are sorted to form a sequence of modified behaviors for this target learner.

10. The method for identifying and bundling English knowledge-confusing entity pairs based on answer modification behavior sequences as described in claim 9, characterized in that, The original operation log data includes at least the operation type field, operation content field, operation timestamp field, question type field, question unique identifier field, and question correct answer text field.

Images