A learning resource dynamic pushing management and control system based on a multi-dimension learning situation analysis model

Abstract

The present application relates to the technical field of intelligent education and learning resource pushing, and particularly relates to a learning resource dynamic pushing control system based on a multi-dimensional learning situation analysis model, comprising the following steps: at a pushing decision node, a multi-dimensional learning situation analysis model is used to retain multiple candidate learning situation branches capable of explaining the same learning interaction result and having different resource requirements, forming a learning situation ambiguity domain; a target learning resource is selected which can make each candidate learning situation branch produce different predicted response marks, a branch continuation relationship between each predicted response mark and a corresponding subsequent fragment is established, an ambiguity resolution pushing unit is formed; a pre-fragment in the ambiguity resolution pushing unit is sent to a learning terminal, an actual response mark is matched with a predicted response mark, a corresponding candidate learning situation branch is selected and only a subsequent fragment is sent, and a learning situation state is confirmed. The present application synchronously completes resource presentation and learning situation disambiguation, and reduces resource mismatch caused by misjudgment of learning situation.

Description

Technical Field

[0001] This invention relates to the field of intelligent education and learning resource delivery technology, and in particular to a dynamic delivery and management system for learning resources based on a multi-dimensional learning analysis model. Background Technology

[0002] With the popularization of online education platforms, smart learning terminals, and digital teaching resources, learning resource recommendation systems have gradually evolved from a unified recommendation method based on course progress or fixed rules to a personalized recommendation method based on learners' knowledge mastery, ability level, learning preferences, and other learning information. Existing systems typically determine the learner's current learning status using rule models or machine learning models based on learning interaction results such as correctness of answers, answer time, number of prompt calls, and resource browsing progress, and then select appropriate learning resources from the resource library based on this learning status.

[0003] However, there is no one-to-one correspondence between learning interaction results and actual learning status. The same incorrect answer may be caused by different reasons, such as incomplete knowledge transfer, insufficient knowledge transfer ability, or inappropriate resource presentation; similarly, the same quick response may correspond to sufficient knowledge mastery, guessing, or a question with low difficulty. While these different learning statuses can produce the same or similar learning interaction results, their subsequent resource needs may differ. Existing learning analysis methods typically select the single learning status with the highest probability as the analysis result, which can easily lead to premature convergence of learning statuses, thus mistaking a candidate status explaining the surface interaction result for the learner's true state.

[0004] To mitigate the aforementioned risk of misjudgment, some existing technologies supplement the judgment criteria by increasing the number of test questions, extending the interaction process, or repeatedly collecting learning behavior data. However, these supplementary contents typically do not address the specific differences between candidate learning states, potentially increasing redundant interactions that cannot effectively distinguish between them. Furthermore, some resource delivery systems send complete learning resources directly before the learner's learning status is clear, then adjust the delivery strategy based on the learner's subsequent performance. This not only increases resource transmission volume and learning time but may also expose learners to teaching content that does not match their actual needs prematurely, interfering with subsequent learning status assessments.

[0005] Furthermore, existing systems typically treat learning status identification and resource delivery as two relatively independent processes. The delivered resources are primarily used for knowledge explanation or practice, with little utilization of the resource fragments themselves to actively differentiate between multiple candidate learning statuses. The actual response generated after resource delivery also lacks a unique continuation mechanism between the candidate learning status branches and subsequent resource fragments. In cases of network retransmission, repeated terminal submissions, or concurrent delivery, multiple mutually exclusive resource fragments may be repeatedly released. Even if an interaction has resolved learning status ambiguity, its confirmation result is often not used as the baseline state for the next delivery decision node, leading to repeated inferences in subsequent analysis or the prolonged reliance on a single identification result, failing to reflect changes in learning status in a timely manner. Summary of the Invention

[0006] This invention provides a dynamic learning resource push and control system based on a multi-dimensional learning situation analysis model. It retains multiple reasonable candidate learning situation states, selects learning resource segments that can generate differentiated responses from different candidate learning situation states, uses the actual interactive response of the segment to eliminate learning situation ambiguity, and continuously connects the disambiguation result with the unique subsequent resource and the next push decision.

[0007] A dynamic learning resource push and control system based on a multi-dimensional learning analysis model, the system being configured to perform the following steps: At the push decision node, a multi-dimensional learning analysis model is used to retain multiple candidate learning branches that can explain the same learning interaction results but have different resource requirements, and these multiple candidate learning branches form a learning ambiguity domain. Based on the aforementioned learning ambiguity domain, target learning resources that enable each candidate learning branch to generate different predicted response labels are selected from the candidate learning resources. The branch continuation relationship between each predicted response label and the corresponding subsequent segment is established to form an ambiguity resolution push order. Send the preceding segment of the ambiguity resolution push order to the learning terminal, match the actual response marker returned by the learning terminal with the predicted response marker, select the corresponding candidate learning branch and send only the subsequent segment corresponding to the candidate learning branch, confirm the selected candidate learning branch as the confirmed learning status, and use the confirmed learning status as the input for the next push decision node to form the learning ambiguity domain.

[0008] Optionally, the multi-dimensional learning analysis model includes a shared temporal encoding layer, a knowledge mastery branch, an ability level branch, a learning style branch, and a joint state output layer. The shared temporal encoding layer consists of two layers of gated loop units connected in sequence, used to encode the model input sequence arranged according to the learning interaction order and output a shared interaction representation. The knowledge mastery branch, ability level branch, and learning style branch output the state probability of the corresponding dimension based on the shared interaction representation. The joint state output layer is used to concatenate the shared interaction representation with the state probabilities of each dimension and output the joint state probability corresponding to each dimension state combination, so as to determine the state combination of each dimension as the candidate learning state.

[0009] Optionally, retaining multiple candidate learning branches that can explain the same learning interaction result but have different resource requirements specifically includes: Obtain the current learning interaction result, use the multi-dimensional learning analysis model to infer multiple candidate learning states that can generate the current learning interaction result based on the combination of states of each dimension, and calculate the confidence level of each candidate learning state. Candidate learning statuses with confidence levels higher than a preset threshold and whose required push resource types are different from each other, as determined by preset resource mapping rules, are selected as candidate learning status branches.

[0010] Optionally, all the selected candidate learning branches can be constructed into a set of states representing the uncertainty of the learner's current learning situation, and the set of states can be marked as the learning situation ambiguity domain.

[0011] Optionally, the learning ambiguity domain is obtained, and for each candidate learning branch therein, the multi-dimensional learning analysis model is used to predict the response of the candidate learning branch to each resource segment in the candidate learning resource library, generating candidate predicted response tags; resource segments that can generate different predicted response tags for different candidate learning branches are selected as the preceding segments, and the complete learning resource including the preceding segments is determined as the target learning resource.

[0012] Optionally, based on the resource requirements of each candidate learning branch, a subsequent segment that matches the candidate learning branch and is logically and content-connected with the preceding segment is determined from the target learning resources, and a unique mapping is established between each predicted response tag and the corresponding subsequent segment to form the branch continuation relationship.

[0013] Optionally, the preceding fragment, all predicted response tags and their respective mapped subsequent fragments are organized according to the branch continuation relationship to form the ambiguity resolution push order.

[0014] Optionally, the preceding segment in the ambiguity resolution push notification is sent to the learning terminal, and the learning terminal receives the actual response markers generated by the learner's interaction behavior with the preceding segment; the actual response markers are compared with the predicted response markers carried in the ambiguity resolution push notification to determine the predicted response markers that match the actual response markers.

[0015] Optionally, the branch continuation relationship is obtained, and according to the branch continuation relationship, the candidate learning branch uniquely mapped by the matching predicted response tag is selected, and only the subsequent segments corresponding to the candidate learning branch are sent to the learning terminal.

[0016] Optionally, the selected candidate learning branch is confirmed as a learning status. At the next push decision node, the confirmed learning status is used as the current learning baseline status input of the multi-dimensional learning analysis model. Combined with the newly acquired learning interaction results, the process of forming a learning ambiguity domain is re-executed to generate a new learning ambiguity domain.

[0017] The beneficial effects of this invention are: This invention addresses the issue of the hidden nature of a single learning interaction result potentially corresponding to multiple learning states, each requiring different resources. It preserves candidate learning branches with varying resource requirements, forming a learning ambiguity domain, thus avoiding premature fixation of a single learning conclusion. Furthermore, it selects a preceding segment that enables each candidate learning branch to generate distinct predictive responses, using this segment as an active discrimination carrier, and establishing a unique continuation relationship between the predictive response marker, candidate learning branches, and subsequent segments. This allows for the simultaneous completion of resource presentation and learning disambiguation using shorter learning content, reducing resource mismatches caused by misjudgments of learning states, and minimizing the learning burden and transmission overhead caused by repeatedly pushing complete resources to verify learning states.

[0018] This invention performs a step-by-step matching of the learner's actual response markers to preceding segments with each predicted response marker. Only subsequent segments corresponding to matching candidate learning branches are released. Locking, unique release, and idempotent control are used to prevent concurrent requests or network retransmissions from causing multiple mutually exclusive subsequent segments to be pushed simultaneously, improving the accuracy and controllability of the dynamic push process. Simultaneously, the confirmed learning status obtained after disambiguation is used as the current learning baseline status for the next push decision node, allowing subsequent analysis to inherit confirmed information. When a new learning interaction result conflicts with this baseline status, a new learning ambiguity domain can be formed, thus balancing continuous tracking of learning evolution and identification of abnormal changes, avoiding long-term erroneous pushes caused by a single accidental interaction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the execution steps of S2 in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] This invention does not simply recommend resources, but rather withholds each subsequent segment first, resolves ambiguities, and then selectively releases them, thereby reducing the complete and erroneous resource push caused by misjudgment due to learning difficulties.

[0023] like Figure 1 As shown, a dynamic learning resource push and control system based on a multi-dimensional learning analysis model runs on a learning resource push server and communicates with learning terminals. The system is configured to perform the following steps: S1. At the push decision node, multiple candidate learning branches that can explain the same learning interaction results but have different resource requirements are retained through a multi-dimensional learning analysis model, and multiple candidate learning branches form a learning ambiguity domain.

[0024] S11. Construct the input structure and basic data table for the multi-dimensional learning analysis model.

[0025] A multi-dimensional learning analysis model should include at least three dimensions: knowledge mastery, ability level, and learning style. The knowledge mastery dimension represents the learner's understanding and retention of the target knowledge content; the ability level dimension represents the learner's ability to independently complete learning tasks given the difficulty of the resources; and the learning style dimension represents the learner's relatively effective response tendency under different resource presentation methods.

[0026] Before building the model, resource attribute records for each learning resource are obtained from the learning resource management platform, and a learning resource attribute table is established. The learning resource attribute table includes at least the resource identifier, target knowledge content identifier, resource difficulty level, resource presentation method, resource function type, and resource version identifier.

[0027] The learning resource attribute table is shown in Table 1.

[0028] Table 1 Examples of Learning Resource Attributes

[0029] The difficulty level of a resource is initially labeled by the resource provider when the resource is added to their database, and then corrected using the independent completion performance of calibration learners. Calibration learners are those whose knowledge mastery and ability level have been determined through pre-testing. The resource difficulty correction value is obtained according to the following formula: ; in, This is the resource difficulty correction value for learning resources, with a range of values. The larger the value, the higher the difficulty of obtaining the resource; Identify learning resources; To calibrate the number of times learners correctly complete the corresponding learning task without triggering prompts; To calibrate the total number of times learners perform the corresponding learning task.

[0030] Learning resources with difficulty correction values ​​in the top 33% are marked as basic difficulty, those in the middle 34% as standard difficulty, and those in the bottom 33% as advanced difficulty. These boundaries are determined based on the resource library's own difficulty distribution; when the resource library is updated, the boundaries are recalculated according to the updated difficulty correction values.

[0031] Obtain the current learning interaction result and merge it with the corresponding resource attribute record in the learning resource attribute table to form the current interaction vector: ; in, This is the current interaction vector corresponding to the current learning interaction result; For the learning interaction order; The result is 1 for correct completion, 0.5 for partial completion, and 0 for incorrect completion. To standardize the completion time; This represents the number of valid operations during the current learning interaction process; The number of times the learner actively invokes the prompt message; This is a flag indicating mid-game exit; it is set to 1 if the game exits midway, and 0 otherwise. This is the resource difficulty correction value for the current learning resource; To encode the resource presentation method, one-hot encoding is used for text, image and text, video, audio and interactive resources respectively.

[0032] Partial completion refers to a learning task having multiple pre-defined scoring points by the resource provider, where the learner completes at least one scoring point but not all of them. The value 0.5 is only used to represent a discrete state between correct and incorrect completion; when the learning platform has already provided the actual score percentage, the actual score percentage is used instead of 0.5.

[0033] The standardized completion time is obtained using the following formula: ; in, This refers to the actual completion time of the current learning interaction; This represents the average completion time of the corresponding learning resources in the training data; This represents the standard deviation of the completion time of the corresponding learning resources in the training data.

[0034] When the current learning resource is a newly added resource and there are not enough completion time records, the average completion time and standard deviation of completion time of resources with the same target knowledge content identifier, resource difficulty level, and resource presentation method are used. This avoids misjudging normal long-term learning as insufficient ability due to the inherent difference in the duration of different learning resources.

[0035] Arrange the most recent current interaction vectors into the model input sequence according to the learning interaction order: ; in, The input sequence for the model formed at the end of the current learning interaction; The length of the input sequence.

[0036] Input sequence length During the model pre-training phase, four candidate values ​​(10, 20, 30, and 40) are selected. The model is trained using each candidate value, with the macro-average F1 score of the candidate learning state recognition set as the first evaluation metric and the single inference time as the second. Among the candidate values ​​where the difference between the macro-average F1 score and the highest value does not exceed 1 percentage point, the candidate value with the shortest inference time is selected. In this embodiment, L=20 is selected. When the learner has fewer than 20 learning interactions, zeros are padded at the beginning of the input sequence, and a mask is set to prevent the padded positions from participating in model state updates.

[0037] S12. Construct and train a multi-dimensional learning analysis model.

[0038] The multi-dimensional learning analysis model adopts a structure of shared temporal coding layer, dimensional analysis branch, and joint state output layer.

[0039] The shared temporal coding layer consists of two layers of gated recurrent units connected sequentially, used to process the model input sequence according to the learning interaction order. The output dimension of the first layer of gated recurrent units is 64, and the output dimension of the second layer of gated recurrent units is 64. The terminal hidden state of the second layer serves as the shared interaction representation.

[0040] Shared interactive representations are obtained according to the following formula: ; in, This represents the shared interaction representation corresponding to the current learning interaction. This is the first-level gated loop unit; This is the second-level gated loop unit.

[0041] The two-layer, 64-dimensional structure was chosen because, during the model pre-training phase, single-layer 32-dimensional, single-layer 64-dimensional, two-layer 64-dimensional, and two-layer 128-dimensional structures were validated. The two-layer 64-dimensional structure, compared to the single-layer structure, improved the macro-average F1 score for candidate learning state recognition, while its single inference time was shorter than that of the two-layer 128-dimensional structure. In this embodiment, measurements on the learning resource push server showed that the two-layer 64-dimensional structure could meet the real-time processing requirements of the push decision node.

[0042] The dimensional analysis branches include the knowledge mastery branch, the ability level branch, and the learning style branch. The dimensions are divided as follows: the knowledge mastery dimension is divided into three states: to be established, in formation, and mastered; the ability level dimension is divided into three states: basic, standard, and extended; and the learning style dimension is divided into four states: text-dominated, audiovisual-dominated, interaction-dominated, and hybrid response.

[0043] The knowledge mastery branch and the ability level branch each consist of a fully connected layer with 32 neurons and a classification layer with 3 output nodes, respectively; the learning style branch consists of a fully connected layer with 32 neurons and a classification layer with 4 output nodes. Each fully connected layer uses a modified linear unit as the activation function.

[0044] The outputs from the three dimensions, along with the shared interactive representation, are fed into the joint state output layer: ; in, The joint state logic value vector generated for the joint state output layer; The connection weight matrix of the joint state output layer; The state probability vector output by the knowledge acquisition branch; The state probability vector output by the capability level branch; The state probability vector output by the learning style branch; This is the bias vector for the joint state output layer; the semicolon in square brackets indicates vector concatenation.

[0045] The three knowledge mastery states, three ability level states, and four learning style states form 36 possible state combinations. Therefore, the joint state output layer has 36 output nodes, each corresponding to a candidate learning state: ; in, The candidate learning status corresponding to the current number; This represents the knowledge mastery status of the candidate learning status corresponding to the current number; This refers to the ability level status of the candidate learning status corresponding to the current number; The learning style status corresponding to the current number of candidate learning status; Number the combination of dimensional states.

[0046] The training samples consist of continuous learning interaction records and the results of subsequent verification tasks. The knowledge mastery status label is determined based on the learner's completion of the same target knowledge content verification task after spaced learning without invoking prompts; the ability level status label is determined based on the highest resource difficulty level that the learner can stably and independently complete; the learning style status label is determined based on the learner's correct completion rate and standardized completion time on resources with the same knowledge content and difficulty level but different presentation methods.

[0047] When the correct completion rate of a certain presentation method is not lower than that of other presentation methods, and its average standardized completion time is the shortest, the learning style state corresponding to that presentation method is used as the label. When the difference in the correct completion rate between two or more presentation methods does not exceed 5 percentage points, and the difference in the average standardized completion time does not exceed one-tenth of the standard deviation of completion time in the training data, the learning style state is labeled as a mixed response. The 5 percentage point and one-tenth were obtained by cross-validation on the training data from three candidate ranges: 3, 5, and 8 percentage points, and one-twentieth, one-tenth, and one-fifth, respectively. The highest repetition recognition consistency rate of the learning style state on the independent validation set was used as the selection criterion.

[0048] The training loss is determined by the following formula: ; in, The total training loss of the multi-dimensional learning analysis model; The cross-entropy loss is the output layer of the joint state. Cross-entropy loss for knowledge acquisition branches; The cross-entropy loss of the capability-level branch; The cross-entropy loss is used to learn the style branch.

[0049] Using an equal-weighted averaging method can prevent any dimension analysis branch from dominating the shared temporal coding layer during training, without the need to set additional subjective weight coefficients.

[0050] Training data is divided into training, validation, and test sets based on learner identifiers. Records from the same learner are only included in one dataset to avoid the same learner's interaction patterns appearing in both the training and testing phases. The proportions of the training, validation, and test sets are 70%, 15%, and 15%, respectively.

[0051] The model was trained using the Adam optimizer. Initial learning rates were selected from 0.0001, 0.0005, and 0.001, and batch sizes were selected from 64, 128, and 256, with the highest macro-average F1 score for joint state recognition on the validation set being the selection criterion. In this embodiment, the initial learning rate was 0.001, and the batch size was 128. The maximum number of training epochs was set to 100. If the total training loss on the validation set did not decrease for 8 consecutive epochs, training was stopped, and the model parameters were restored to the level where the total training loss on the validation set was lowest. The 8 consecutive epochs were obtained by validating four candidate values: 4, 6, 8, and 10 epochs. Eight epochs avoided premature stopping due to short-term loss fluctuations and did not result in significant overfitting.

[0052] After model training is complete, the structural parameters of the shared temporal coding layer, the three-dimensional analysis branches, and the joint state output layer are deployed to the learning resource push server. At each push decision node, the model input sequence formed in step S11 is input into the multi-dimensional learning analysis model, and the model is not retrained during the push decision process.

[0053] S13. Infer multiple candidate learning statuses and calculate confidence levels.

[0054] At the push decision node, the model input sequence is input into the trained multi-dimensional learning analysis model to obtain the joint state logic value corresponding to 36 candidate learning states.

[0055] To reduce the impact of excessive concentration in the output probabilities of the neural network, the joint state logic value is calibrated using a validation set, and the confidence level of each candidate learning state is calculated according to the following formula: ; in, The confidence level of the candidate learning status corresponding to the current number; This refers to the joint state logic value in the joint state logic value vector that corresponds to the current candidate learning state. This is the temperature calibration coefficient; This is the joint state number used for summation.

[0056] The temperature calibration coefficient is obtained by fixing other parameters of the multi-dimensional learning analysis model and minimizing the joint state negative log-likelihood loss on the validation set. ; in, These are the temperature calibration coefficients determined and actually used after training is completed; To determine the number of samples in the validation set; To verify the sample number; This represents the joint state logical value of the current verification sample in the true joint state. The true joint state number of the current validation sample; This is the joint state logical value of the current verification sample in the corresponding joint state.

[0057] In one embodiment, the temperature calibration coefficient obtained after validation set optimization is 1.37. This value is obtained from the validation results of the corresponding training model; after changing the training data or retraining the model, the temperature calibration is re-performed, and the value of 1.37 is not directly used.

[0058] The confidence score is compared with a preset confidence threshold. The preset confidence threshold is set between 0.50 and 0.90, and candidate values ​​are validated at 0.05 intervals. The recognition accuracy of the retained state and the proportion of learning interaction results that retain two or more states are calculated for each candidate value. The candidate value with the highest recognition accuracy of at least 85% and the highest proportion of results that retain two or more states is selected as the preset confidence threshold.

[0059] In this embodiment, the preset confidence threshold is: ; in, A preset threshold is set for the confidence level.

[0060] The required 85% recognition accuracy is determined based on the learning resource recommendation system's allowed false positive rate cap of 15%, which is set by the resource recommendation manager based on trial operation data before system deployment. If the system is used in learning resource scenarios with high intervention intensity, the allowed false positive rate cap can be set even lower, and the confidence threshold can be re-obtained according to the above verification process.

[0061] Add the candidate learning status that satisfies the following formula to the initial candidate status set: ; in, This is the initial candidate state set corresponding to the current learning interaction.

[0062] Therefore, the multi-dimensional learning analysis model does not only output the single state with the highest confidence, but retains multiple candidate learning states with credible explanatory power. For example, the same incorrect answer may correspond to knowledge yet to be established, ability standard, and text-driven, or it may correspond to knowledge formation in progress, basic ability, and audiovisual-driven, thus retaining different learning reasons behind the same learning interaction result.

[0063] S14. Filter candidate learning branches according to the preset resource mapping rules.

[0064] A resource mapping rule table is pre-established to determine the types of resources to be pushed to each candidate learning status. The resource mapping rule table should at least include fields for knowledge mastery status, ability level status, learning style status, resource function, resource difficulty, and resource presentation method.

[0065] Resource mapping rules are pre-built according to the following deterministic rules: 1. When the knowledge acquisition status is "to be established" and the ability level status is "based," the resource function is identified as the guiding example; 2. When the knowledge mastery status is "to be established" and the ability level status is "standard" or "expanded", the resource function is determined to be concept explanation; 3. When the knowledge acquisition status is in the process of formation and the ability level is at the basic level, the resource function is determined as strategy demonstration; 4. When the knowledge acquisition status is in the process of formation and the ability level is at the standard or extended level, the resource function is determined to be transfer practice; 5. When the knowledge mastery status is "already mastered" and the ability level status is "basic," the resource function is determined as step correction; 6. When the knowledge mastery status is "mastered" and the ability level status is "standard" or "expanded", the resource function is defined as an expansion task; 7. The resource difficulty field corresponds to the ability level status, with basic, standard, and extended abilities corresponding to basic, standard, and extended difficulty, respectively; 8. The resource presentation mode field corresponds to the learning style status: text-dominant corresponds to text resources, audiovisual-dominant corresponds to graphic or video resources, interactive-dominant corresponds to interactive resources, and mixed response corresponds to resources with two or more presentation modes that can be switched.

[0066] The above rules are first established by teaching resource management personnel based on the teaching functions of the resources, and then verified item by item using the completed learning resource attribute table. Only when the resource function, resource difficulty, and resource presentation method can be found in the learning resource attribute table can the rule be activated; otherwise, the rule is marked as a resource rule to be supplemented and will not be used for the current candidate learning information branch selection.

[0067] The resource mapping rules are shown in Table 2.

[0068] Table 2 Examples of Resource Mapping Rules

[0069] The combination of resource function, resource difficulty, and resource presentation method is encoded as the required push resource type. The same resource function is considered as different required push resource types when the resource difficulty or resource presentation method is different.

[0070] For each candidate learning status in the initial candidate status set, query the resource mapping rule table to obtain the corresponding required push resource type: ; in, The required resource type corresponding to the current candidate learning status; This is a pre-defined resource mapping function. The resource mapping function maps the three-dimensional status to the required push resource type by querying the resource mapping rule table.

[0071] In practical application, the candidate learning statuses in the initial candidate status set are first arranged in descending order of confidence level; the candidate learning status at the top of the list is retained, and then the other candidate learning statuses are read in sequence. If the required push resource type corresponding to the currently read candidate learning status has not been used by the retained status, the current candidate learning status is retained; if the required push resource type has been used, the current candidate learning status is discarded.

[0072] The screening results meet the following requirements: ; in, The required push resource types corresponding to the currently retained candidate learning status; The number is for the candidate learning status that has been retained.

[0073] The candidate learning statuses retained after screening are designated as candidate learning status branches. The above method of screening in order of confidence level can retain candidate learning statuses with higher credibility when resource needs overlap, while ensuring that the required resource types for any two candidate learning status branches are different from each other.

[0074] If only one candidate learning information branch remains after filtering, it indicates that the current learning interaction results have not created any ambiguity in learning information at the resource demand level, and the system will proceed to the normal resource matching process according to that candidate learning information branch; if more than two candidate learning information branches remain after filtering, then step S15 will be executed.

[0075] S15. Construct and label the learning ambiguity domain.

[0076] All selected candidate learning outcomes branches are constructed into a set of states representing the uncertainty of the learner's current learning situation. To enable comparison of the confidence levels of different candidate learning outcomes branches within this set of states, the retained confidence levels are normalized within the set: ; in, The normalized confidence level of the current candidate learning branch within the state set; The confidence level of the current candidate learning branch obtained in step S13; The total number of candidate learning information branches retained after screening; The candidate learning branch number used when performing normalized summation.

[0077] Write the candidate learning status, normalized confidence score, and required push resource type into the status set: ; in, This is the set of states corresponding to the current learning interaction.

[0078] The set of states is marked as the learning ambiguity domain, and learner identifier, push decision node identifier, target knowledge content identifier, and generation time are set for the learning ambiguity domain. An example of the record for the learning ambiguity domain is shown in Table 3.

[0079] Table 3 Examples of Learning Learning Ambiguity Domain Records

[0080] The learning ambiguity domain is output to the subsequent step S2 as input for filtering target learning resources and constructing predicted response labels for each candidate learning branch. After the subsequent steps determine that the actual response label matches a candidate learning branch, the learning ambiguity domain is replaced with the corresponding confirmed learning status for use by the next push decision node.

[0081] The multi-dimensional learning analysis model does not preemptively eliminate multiple possible interpretations of the same learning interaction result. Instead, it retains candidate learning states that meet the confidence requirements and will elicit different resource needs. Using the difference in resource needs as a necessary condition for retaining candidate learning state branches ensures that the resulting learning state ambiguity domain not only represents statistical uncertainty but also directly reflects decision-making conflicts between different resource delivery choices. Therefore, the subsequent system can use pre-existing fragments of learning resources to specifically differentiate candidate learning state branches, avoiding the direct adoption of the highest probability state and the sending of complete but unsuitable learning resources to the learning terminal.

[0082] S2. Based on the ambiguity domain of learning information, select target learning resources from the candidate learning resources whose preceding segments can generate different predicted response labels for each candidate learning information branch, establish the branch continuation relationship between each predicted response label and the corresponding subsequent segments, and form an ambiguity resolution push order.

[0083] S21. Predict the response of candidate learning branches to each resource segment.

[0084] Obtain the learning ambiguity domain output in step S1. For each candidate learning branch, use a multi-dimensional learning analysis model to predict the response of that candidate learning branch to each resource segment in the candidate learning resource library, and generate candidate predicted response labels.

[0085] S211, Pre-construction of candidate learning resource libraries.

[0086] Each complete learning resource is pre-divided into multiple resource fragments according to the smallest content unit that can be presented independently and form an effective learning interaction.

[0087] Each resource segment should have at least one interactive checkpoint, which can be a selection operation, input operation, playback control operation, step confirmation operation, or segment end confirmation operation.

[0088] The candidate learning resource library sets the resource fragment attribute table.

[0089] The resource fragment attribute table should include at least the resource fragment identifier, the identifier of the complete learning resource to which it belongs, the target knowledge content identifier, the resource function type, the resource difficulty level, the resource presentation method, the estimated completion time, the number of interaction checkpoints, the fragment position type, the preceding fragment identifier, the allowed subsequent fragment identifier, and the fragment version identifier.

[0090] Among them, the segment position type is divided into pre-position type, middle type and end type.

[0091] Only resource fragments that can be displayed independently, have interactive checkpoints, and have two or more allowed subsequent fragments can be marked as preceding resource fragments.

[0092] An example of a resource fragment attribute table is shown in Table 4.

[0093] Table 4 Examples of Resource Fragment Attribute Representation

[0094] Resource segments in the candidate learning resource library should not be segmented solely based on a fixed word count or fixed playback duration. Instead, segmentation should be based on the semantic integrity of the content and the ability to independently record the interaction results. For example, for video resources, a segment should be defined as the period from when a question is posed to when it is first explained; for interactive exercise resources, a segment should be defined as the period from when the task is described to when the first submission is made; and for text resources, a concept description and its corresponding confirmation question should be defined as a segment.

[0095] Resource fragments are not only transmission or display units, but also serve as the smallest unit of action for predicting the differences in responses to different candidate learning branches, thereby avoiding sending complete learning resources to the learning terminal in order to resolve learning ambiguity.

[0096] S212. Construct candidate learning information branch codes.

[0097] The knowledge mastery status, ability level status, and learning style status of each candidate learning branch are converted into one-hot codes, and then concatenated to obtain the candidate learning branch codes: ; in, Code the candidate learning information branch for the current candidate learning information branch; This is a one-hot encoding function. Its execution process is to establish the same number of encoding bits as the number of states according to the preset state order of the corresponding dimension, set the encoding position corresponding to the current state to 1, and set the remaining encoding positions to 0.

[0098] Each of the knowledge mastery dimension and the ability level dimension has three status coding bits, and the learning style dimension has four status coding bits. Therefore, the candidate learning branch coding has ten coding bits.

[0099] S213, Construct a resource fragment encoder.

[0100] For each resource fragment in the candidate learning resource library, the resource function type, difficulty level, presentation method, estimated completion time, and number of interaction checkpoints are read from the resource fragment attribute table. Content features are then extracted from the text, subtitles, or interactive prompts of the resource fragment to form the resource fragment input vector. ; in, The resource fragment input vector is the current resource fragment. Number the resource fragments; One-hot encoding for resource function types; One-hot encoding for resource difficulty level; One-hot encoding for resource presentation methods; The estimated completion time has been standardized. The number of standardized interaction checkpoints; This is the content feature vector of the resource fragment.

[0101] Content feature vectors are obtained as follows: The text, subtitles, or interactive prompts of resource fragments are segmented into words; the term frequency-inverse document frequency (IF-IVF) of each term in the candidate learning resource library is calculated; and the resulting term vectors are reduced to 64 dimensions using singular value decomposition. For audio or video resources, the subtitle text generated when the resource is added to the library is read first; for resources without subtitle text, the speech recognition module generates subtitle text during the addition process, which is then confirmed by the resource management personnel.

[0102] The expected completion time and number of interaction checkpoints are standardized using the mean and standard deviation of the candidate learning resource library, respectively, without setting a fixed upper limit. The resource fragment encoder consists of two sequentially connected fully connected layers. The first fully connected layer has 128 neurons, and the second fully connected layer has 64 neurons. Both fully connected layers use modified linear units as the activation function. The resource fragment is encoded as follows: ; in, Encode the resource fragment for the current resource fragment; This is the connection weight matrix of the first fully connected layer; This is the connection weight matrix for the second fully connected layer; This is the bias vector of the first fully connected layer; This is the bias vector for the second fully connected layer; To correct the linear unit function, the execution process involves comparing the input value with zero and outputting the larger value.

[0103] The selection criteria for 128 neurons and 64 neurons are as follows: During the model training phase, the 64-32, 128-64, and 256-128 structures were validated respectively. The accuracy of predicting response tags was used as the primary evaluation metric, and the time required to complete one prediction for all candidate learning branches and resource segments was used as the secondary evaluation metric. Among the structures whose accuracy differed from the highest value by no more than 1 percentage point, the structure with the shortest prediction time was selected. This embodiment uses the 128-64 structure.

[0104] S214. Construct the response prediction head.

[0105] The multi-dimensional learning analysis model further sets up a response prediction head based on the shared temporal coding layer, dimensional analysis branches, and joint state output layer in step S1. The response prediction head receives candidate learning branch codes and resource fragment codes, and outputs the probabilities of four response fields: completion status, completion duration status, prompt call status, and checkpoint result.

[0106] Each response field is divided into three response categories: completion status is divided into complete, partially complete, and exited midway; completion duration status is divided into short time, normal, and long time; prompt call status is divided into not called, called once, and called more than twice; checkpoint result is divided into all correct, partially correct, and all incorrect.

[0107] Short and long durations are determined based on the completion duration distribution of the same resource segments in the training data: segments not higher than the first quartile are marked as short, segments higher than the first quartile but lower than the third quartile are marked as normal, and segments not lower than the third quartile are marked as long. When a resource segment is a newly added segment, the completion duration quantiles of other resource segments with the same resource function type, difficulty level, and presentation method are used.

[0108] No. The candidate learning branch for the first The resource fragment in the first The probability of the response category on each response field is: ; in, To correspond to the candidate learning branch, the corresponding resource segment is in A probability vector of response categories on each response field; The response field numbers correspond sequentially to the completion status, completion duration status, prompt call status, and checkpoint result. This is the output weight matrix for the corresponding response field; This is the bias vector for the corresponding response field; The normalization exponential function is executed by calculating the exponent of each output value separately and then normalizing it by summing the exponents of all output values, so that the sum of the probabilities of each response class equals 1.

[0109] S215, Training of the Response Prediction Head.

[0110] The training samples for the response prediction head are formed by the confirmed learning status, the presented resource segments and their actual completion status, completion duration, prompt call status, and checkpoint results. The actual response is uploaded by the learning terminal according to the resource segment identifier and interaction checkpoint identifier, and associated with the confirmed learning status during training. The training loss of the response prediction head is the weighted average of the cross-entropy loss of the four response fields. ; in, In response to the training loss of the prediction head; This represents the cross-entropy loss for the corresponding response field.

[0111] Using an equal-weighted averaging method can prevent the completion time status or completion status from dominating model training due to a large number of samples. The response prediction head and the multi-dimensional learning analysis model in step S1 use the same training set, validation set, and test set partitioning method, and are jointly trained using the optimizer, learning rate, training batch size, and early stopping condition determined in step S1.

[0112] During joint training, the shared temporal coding layer and dimensionality analysis branch are initially fixed, and only the resource fragment encoder and response prediction head are trained. After the response prediction accuracy on the validation set no longer improves for two consecutive training epochs, the fixed state of the shared temporal coding layer is released, and all model parameters are jointly fine-tuned with one-tenth of the original learning rate. The basis for setting one-tenth of the original learning rate is: after verifying the original learning rate, half of the original learning rate, and one-tenth of the original learning rate, it was found that one-tenth of the original learning rate can improve the response prediction accuracy while reducing the macro-average F1 value of joint state recognition in step S1 by no more than 0.5 percentage points.

[0113] S216. Generate candidate predicted response labels.

[0114] For each response field, select the response category with the highest probability, and arrange them in the order of completion status, completion duration status, prompt call status, and checkpoint result to generate candidate predicted response tags: ; in, The candidate predicted response labels are used to mark the corresponding resource segments for the corresponding candidate learning branches; Number the response category within a response field; This represents the predicted probability for the corresponding response category; This is a function that takes the response category number that maximizes the target value.

[0115] For example, candidate predicted response labels of "Completed—Short-time—No Call Hint—All Correct" and "Partially Completed—Long-time—More Than Two Calls—Partially Correct" are distinct predicted response labels. To exclude candidate predicted response labels where the model is clearly uncertain about a certain response field, the label confidence is calculated based on the minimum of the highest class probabilities among the four response fields: ; in, The label confidence level for the corresponding candidate predicted response label; This is a function that takes the minimum value among the results of each response field. This is a function that takes the highest class probability from a response field.

[0116] The preset threshold for label confidence is: ; in, A preset threshold is set for the confidence level of the label.

[0117] The preset confidence threshold for the label is obtained as follows: On the validation set, candidate values ​​from 0.50 to 0.80 are validated one by one at intervals of 0.05. The proportion of candidate predicted response labels higher than each candidate value that match the actual response labels is calculated. The candidate value with a consistency ratio of not less than 80% that retains the most resource fragments is selected. In this embodiment, 0.65 is obtained. When the candidate learning resource library or learner group is updated, the above validation is re-executed, and the original threshold is not directly reused.

[0118] S22. Select preceding segments and determine target learning resources.

[0119] For the same resource segment, compare the candidate predicted response tags corresponding to different candidate learning branches. To avoid making a rough judgment based solely on whether the tag text is exactly the same, calculate the branch response difference based on the proportion of different fields in the four response fields: ; in, For the first The candidate learning information branches and the first The candidate learning branch is for the first The branch response difference of each resource fragment; This refers to the corresponding response field in the previous candidate predicted response tag; This refers to the corresponding response field in the subsequent candidate predicted response tag; This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise.

[0120] For each resource segment, calculate the minimum branch response difference between any two candidate learning branches: ; in, For the first Minimum branch response difference of each resource fragment.

[0121] When the minimum branch response difference is greater than zero, it indicates that the candidate predicted response labels generated by any two candidate learning branches for this resource segment are different, satisfying the requirement of mutual difference. Resource segments that simultaneously satisfy the condition of a minimum branch response difference greater than zero and a label confidence level of all candidate predicted response labels not lower than a preset label confidence level threshold are added to the discriminable segment set. ; in, The set of discriminable segments corresponding to the current push decision node; This means that both screening conditions are met simultaneously.

[0122] When multiple resource segments exist in the discriminable segment set, they are selected in the following order: First, the resource segment with the largest minimum branch response difference is selected; if the minimum branch response differences are the same, the resource segment with the shorter expected completion time is selected; if the expected completion time is still the same, the resource segment with the smaller amount of transmitted data is selected; if the amount of transmitted data is still the same, the resource segment with the newest resource version update time is selected. This eliminates the need to manually set weighting coefficients for discriminative ability, learning time, and transmitted data volume, thus avoiding instability in the initial segment selection results caused by changes in weights.

[0123] The final selected resource segment will be used as the preceding segment: ; in, This is a pre-production segment; These are the resource fragments ultimately selected according to the above filtering order; The resource fragment number for the selected resource fragment.

[0124] The candidate learning resource repository specifies that each resource fragment belongs to only one complete learning resource. By querying the identifier of the complete learning resource to which it belongs in the resource fragment attribute table, the complete learning resource containing the preceding fragment is identified as the target learning resource. ; in, Learning resources for the target audience; The function is a complete learning resource query function. Its execution process involves reading the identifier of the complete learning resource to which the preceding segment belongs and retrieving the corresponding complete learning resource from the candidate learning resource library.

[0125] If the set of discriminable segments is empty, then there are no resource segments in the current candidate learning resource library that can reliably distinguish all candidate learning branches. In this case, the system does not force the formation of an ambiguity resolution push order, but waits for new learning interaction results to enter step S1, or expands the candidate learning resource library from resources with different resource function types than the existing candidate resource segments.

[0126] S23. Determine the subsequent segments and establish branch continuation relationships.

[0127] S231, Pre-built resource fragment connection table.

[0128] When a complete learning resource is added to the database, a resource fragment connection table is created. The resource fragment connection table should include at least the complete learning resource identifier, the preceding fragment identifier, the subsequent fragment identifier, the entry content anchor, the exit content anchor, the target knowledge content identifier, the subsequent fragment resource type, the release order, and the valid version identifier.

[0129] Entry and exit content anchors are set by the resource creators during segmentation. Exit content anchors identify concepts, problems, or tasks that have been raised but not yet completed at the end of the preceding segment; entry content anchors identify the concepts, problems, or tasks undertaken at the beginning of the subsequent segment. Only two resource segments with identical entry and exit content anchors are considered content-connected. An example of a resource segment connection table is shown in Table 5.

[0130] Table 5 Examples of Resource Fragment Connections

[0131] The resource fragment connection table is validated as follows: the preceding fragment and the subsequent fragment must belong to the same target learning resource; the preceding fragment identifier of the subsequent fragment must point to the preceding fragment; the exit content anchor of the preceding fragment must be consistent with the entry content anchor of the subsequent fragment; the target knowledge content identifiers of the preceding fragment and the subsequent fragment must be consistent; no two subsequent fragments with the same subsequent fragment resource type and the same release order may be set after the same preceding fragment.

[0132] The final verification step ensures that each type of required resource corresponds to at most one subsequent segment after the same preceding segment, guaranteeing the uniqueness of branch continuation relationships at the resource repository level. Complete learning resources that fail the verification are not allowed to enter the candidate learning resource repository.

[0133] S232. Determine the subsequent segments according to the candidate learning branches.

[0134] For each candidate learning branch, read the required resource type and select resource segments from the resource segment connection table corresponding to the target learning resource that meet the following conditions: the complete learning resource identifier corresponds to the target learning resource; the preceding segment identifier corresponds to the preceding segment; the resource type of the subsequent segment is consistent with the required resource type of the current candidate learning branch; the entry content anchor is consistent with the exit content anchor of the preceding segment; and the target knowledge content identifier is consistent with the target knowledge content identifier of the preceding segment.

[0135] The selected resource segments are identified as subsequent segments of the corresponding candidate learning information branches. Subsequent segments are represented as follows: ; in, This refers to the subsequent segment corresponding to the current candidate learning information branch; This is a function for querying subsequent segments. Its execution process involves using the target learning resource, the preceding segment, and the type of resource to be pushed as query conditions to read a unique subsequent segment from the resource segment connection table.

[0136] If a candidate learning branch does not have a subsequent segment that meets the conditions, the current complete learning resource cannot be identified as the target learning resource. The system returns to step S22 and selects the next resource segment from the set of discriminable segments until the identified target learning resource can provide a subsequent segment for each candidate learning branch.

[0137] S233. Establish a unique mapping.

[0138] Since step S22 has ensured that different candidate learning branches generate distinct candidate prediction response tags, and step S1 has ensured that the required push resource types corresponding to each candidate learning branch are distinct from each other, and the resource segment connection table has restricted that one required push resource type corresponds to only one subsequent segment, a unique mapping between candidate prediction response tags and subsequent segments can be established: ; in, The branch continuation mapping function executes by querying the corresponding candidate learning information branch with the candidate prediction response marker, and then reading the unique subsequent segment corresponding to the candidate learning information branch.

[0139] All unique mappings are collectively marked as branch continuation relationships: ; in, This defines the branch continuation relationship corresponding to the current push decision node. This represents a unique correspondence between candidate predicted response labels and subsequent segments.

[0140] Instead of directly selecting subsequent segments based on the confidence level of candidate learning branches, this branch continuation relationship first waits for the preceding segment to generate an actual response, and then the subsequent steps use the actual response marker to trigger the corresponding mapping, thereby avoiding high-confidence but incorrect candidate learning branches from directly obtaining full resource push permissions.

[0141] S24. Form an ambiguity resolution push order.

[0142] The preceding segments, all candidate predicted response tags, and their respective mapped subsequent segments are organized into an ambiguity resolution push order according to the branch continuation relationship. The ambiguity resolution push order includes a push order header, a preceding segment area, a branch continuation area, and an execution control area.

[0143] The push notification header records the push notification identifier, learner identifier, target learning resource identifier, target knowledge content identifier, model version identifier, and generation time; the preceding segment area records the preceding segment identifier, preceding segment version identifier, preceding segment data address, and integrity verification value; the branch continuation area records the candidate prediction response marker, corresponding subsequent segment identifier, subsequent segment version identifier, subsequent segment data address, and integrity verification value item by item; the execution control area records the preceding segment priority release identifier, subsequent segment lock identifier, and push notification failure condition.

[0144] The data structure for disambiguation push notifications is represented as follows: ; in, To resolve ambiguities in the push notification generated at the current push decision node; For push order identification; This serves as the integrity check value for the corresponding subsequent segments; The version identifier of the multi-dimensional learning analysis model used when generating candidate predicted response labels; This is the condition for the push order to expire.

[0145] The integrity check value is obtained using the SHA-256 digest value of the resource fragment file. When reading a resource fragment, the learning terminal or resource push server recalculates the digest value and compares it with the integrity check value in the push order; if they are inconsistent, the corresponding resource fragment is not executed to prevent inconsistencies between the predicted response marker and the actual resource content caused by resource version updates or transmission anomalies.

[0146] The conditions for push notification failure are determined according to the following rules: the current learning task ends, the version of the target learning resource changes, the version of the multi-dimensional learning analysis model changes, or a new learning ambiguity domain is formed in step S1. If any of these conditions are met, the current ambiguity resolution push notification fails. Therefore, there is no need to set a fixed failure duration, avoiding the problem of a uniform time threshold being inapplicable for different course durations.

[0147] When forming an ambiguity resolution push order, only the release lock of the preceding segment is released; all subsequent segments remain locked. After sending the preceding segment to the learning terminal in step S3, the branch continuation relationship is queried based on the actual response mark returned by the learning terminal. Only one subsequent segment corresponding to the matching candidate predicted response mark is released, while other subsequent segments remain locked and are canceled when the push order expires. An example of an ambiguity resolution push order is shown in Table 6.

[0148] Table 6 Examples of Branch Continuation Areas in Ambiguity Resolution Push Orders

[0149] Existing learning resource delivery methods typically select a learning status first, then directly send complete learning resources matching that status. This implementation, however, retains multiple candidate learning status branches and uses a multi-dimensional learning status analysis model to predict the differences in responses to resource segments for each branch. It selects preceding segments that can actively distinguish candidate learning status branches and establishes a unique mapping between the different predicted responses of the candidates and the corresponding subsequent segments. Thus, the preceding segments simultaneously serve the functions of presenting learning content and resolving learning status ambiguities; subsequent segments are only released when the actual response supports the corresponding candidate learning status branch. The technical advantages include reducing the amount of erroneous complete resources transmitted, decreasing the number of invalid downloads and display switching on learning terminals, avoiding continuous erroneous resource delivery caused by a single high-probability learning status judgment error, and allowing learning status analysis results to directly participate in the release control of resource segments.

[0150] S3. Send the preceding segment of the ambiguity resolution push order to the learning terminal, match the actual response mark returned by the learning terminal with the predicted response mark, select the corresponding candidate learning branch and send only the subsequent segment corresponding to the candidate learning branch, confirm the selected candidate learning branch as the confirmed learning status, and use the confirmed learning status as the input for the next push decision node to form the learning ambiguity domain.

[0151] S31. Send the pre-fragment and generate the actual response marker.

[0152] The resource push server reads the ambiguity resolution push order and sequentially verifies the push order identifier, learner identifier, target learning resource version, preceding segment version, and expiration date. Only when the identifier and version match the current record in the candidate learning resource library, the ambiguity resolution push order has not expired, and the SHA-256 digest of the preceding segment file matches the digest in the library, will the preceding segment be unlocked and sent to the learning terminal.

[0153] The resource push server sends the ambiguity resolution push order identifier, the preceding segment identifier, the segment version identifier, the interaction checkpoint configuration, and the actual response marker generation rules along with the preceding segment, but does not send subsequent segments that are in a locked state. This allows the learning terminal to independently collect the interaction behavior of the preceding segment, while preventing the learning terminal from prematurely accessing subsequent segments that it should not have entered.

[0154] The learning terminal establishes a segment interaction session using the disambiguation push order identifier, the preceding segment identifier, and the local session sequence number. It records events such as start presentation, interaction checkpoint operations, prompt invocation, segment completion, mid-session exit, and abnormal closure. Each event is written with a monotonically increasing local sequence number. The resource push server restores the event order according to the local sequence number, not by network arrival time, thus avoiding distortion of the checkpoint operation sequence caused by retransmissions in weak networks. An example of a segment interaction event log table is shown in Table 7.

[0155] Table 7 Examples of Fragment Interaction Event Records

[0156] The actual response markers are generated as follows: After the segmented interactive session ends, the learning terminal discretizes the actual interactive behavior into four response fields—completion status, completion duration status, prompt call status, and checkpoint result—in the same field order as the training samples in step S21, and generates an actual response marker: ; in, The learner's actual response to the preceding segment is labeled; For completion status field; For completion duration status field; To prompt the call status field; This is the checkpoint result field.

[0157] The completion status field is determined by the segment end event: it is marked as complete when a segment completion event exists; it is marked as partially complete when at least one interactive checkpoint has been completed but no segment completion event exists; and it is marked as exited midway when no interactive checkpoints have been completed and the segment has been closed abnormally. The completion duration status uses the terminal monotonic clock difference from the start of presentation to the last valid event, and is divided according to the first quartile and the third quartile determined in step S21 for the same resource segment: it is marked as short when it is not higher than the first quartile; it is marked as normal when it is higher than the first quartile and lower than the third quartile; and it is marked as long when it is not lower than the third quartile.

[0158] The prompt invocation status is divided into not invoked, invoked once, and invoked more than twice, based on the number of valid prompt invocation events. The checkpoint results only count the interactive checkpoints that have been reached: if all checkpoints are correct, it is marked as all correct; if at least one checkpoint is correct or partially correct and at least one checkpoint is incorrect, it is marked as partially correct; if all reached checkpoints are incorrect, it is marked as all incorrect; if no reached checkpoints exist, it is marked as not formed based on the completion status, and the not formed field is used in the distance calculation in step S32.

[0159] When an abnormal shutdown occurs, the learning terminal uses the lifecycle callback of the preceding segment to write an exit event and saves the events that have not yet been uploaded in a local persistent queue. After the network is restored, the ambiguity resolution push order identifier and local sequence number are still uploaded. After receiving the actual response flag, the resource push server re-verifies the ambiguity resolution push order identifier, the preceding segment identifier, and the segment version identifier. If the verification is inconsistent, step S32 is not entered to avoid cross-version interaction records being incorrectly used for branch selection.

[0160] S32. Match the actual response markers with the candidate predicted response markers.

[0161] The resource push server reads all candidate predicted response tags corresponding to the preceding fragment from the ambiguity resolution push order and compares them item by item according to the four response fields. The actual response tag is compared with the... The actual-predicted response distance between the candidate predicted response labels corresponding to each candidate learning branch is: ; in, For the actual response marker and the first The actual-predicted response distance between the candidate predicted response labels corresponding to each candidate learning branch; For the first in the actual response marker One response field; For the first The candidate predicted response tags for the preceding segments in the candidate learning branches are the first one. One response field; This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise.

[0162] Since both the actual response marker and the candidate predicted response marker have four response fields, the possible values ​​for the actual-predicted response distance are 0, 0.25, 0.50, 0.75, and 1. A smaller distance indicates that the actual response is closer to the predicted response of the corresponding candidate learning branch. For unformed fields in the actual response marker, they are considered to be different from any formed predicted field to prevent records with insufficient interaction information from being mistakenly identified as exact matches.

[0163] First matching process: If only one candidate predicted response marker has an actual-predicted response distance of 0, it is directly identified as the matching predicted response marker. If no candidate predicted response marker has a distance of 0, the minimum actual-predicted response distance is selected, and a matching distance threshold is used to determine whether to allow further matching. The matching distance threshold is set as follows: ; in, This is the matching distance threshold.

[0164] The matching distance threshold is obtained through a validation set: four candidate values ​​of 0, 0.25, 0.50, and 0.75 are validated sequentially. The proportion of candidate learning branches whose distances are not higher than the candidate values ​​that match the manually reviewed learning status is calculated, and the maximum value is selected from the candidate values ​​whose proportions are not less than 85%. In this embodiment, 0.25 is obtained, meaning that at most one of the four response fields is allowed to be inconsistent. After the candidate learning resource library, response category classification rules, or learner group are updated, the above validation is re-executed. If the minimum actual-predicted response distance is greater than the matching distance threshold, subsequent segments are not released, and the learning terminal returns to the incomplete interaction checkpoint; if the interaction cannot be supplemented, it waits for new learning interaction results to proceed to step S1.

[0165] Processing of candidate predicted response tags at the same distance: If two or more candidate predicted response tags achieve the same minimum actual-predicted response distance, then the response category probability already output in the response prediction header and saved with the disambiguation push order in step S21 is invoked, and the average negative log-likelihood is calculated for the four response fields of the actual response tag: ; in, For the first The average negative log-likelihood of each candidate learning branch; Mark the actual response as number 1 The response category number for each response field; The natural logarithm function is used to read the probability of the actual response category in the corresponding response category probability vector and calculate the natural logarithm.

[0166] The smaller the mean negative log-likelihood, the higher the joint support the model assigns to the actual response label. The system first selects the candidate predicted response label with the smallest mean negative log-likelihood; If they are still the same, select the candidate learning branch with the higher normalized confidence in step S1; if the normalized confidence is still the same, do not release any subsequent segments, and mark the current ambiguity resolution push order as an interaction to be supplemented.

[0167] This step-by-step processing does not use manual weighting coefficients, which can avoid fluctuations in matching results caused by differences in distance, probability, and learning confidence metrics.

[0168] ; in, Label the predicted responses that match; The candidate learning branch number is determined after being screened step by step using the actual-predicted response distance, average negative log-likelihood, and normalized confidence.

[0169] S33. Select the candidate learning information branch and send the corresponding subsequent segment.

[0170] The resource push server uses the matching predicted response marker as input to the branch continuation mapping function, and reads the uniquely mapped candidate learning information branches and corresponding subsequent segments based on the branch continuation relationship established in step S23: ; in, To allow subsequent segments to be sent to the learning terminal.

[0171] The resource push server performs the following checks before sending subsequent segments: 1. The ambiguity resolution push notification remains valid; 2. The resource fragment identifier and version identifier of the subsequent fragments to be sent are consistent with the ambiguity resolution push order; 3. The SHA-256 digest value of the subsequent fragment to be sent is consistent with the integrity check value in the ambiguity resolution push order; 4. The subsequent segments to be sent are currently locked; 5. The same push notification has not yet released the subsequent segments corresponding to other candidate learning branches.

[0172] After all verifications pass, the resource push server changes the status of the subsequent segments to be sent from locked to released using an atomic state update method, and changes the status of the subsequent segments corresponding to other candidate learning branches from locked to prohibited release.

[0173] The atomic state update method means that the above state changes are completed in the same transaction. If any state change fails, all states are restored to the state before the change.

[0174] Subsequently, only the subsequent segments to be sent are sent to the learning terminal, and the subsequent segments corresponding to other candidate learning branches are not sent.

[0175] An example of a branch release status table is shown in Table 8.

[0176] Table 8. Examples of Branch Release Status

[0177] Subsequent segment transmission requests use the push order identifier and the subsequent segment identifier to form an idempotent transmission identifier. When the learning terminal does not return a receipt confirmation and the resource push server needs to resend, only subsequent segments with the same idempotent transmission identifier are allowed to be resent; branch selection or resending other subsequent segments is not permitted.

[0178] After receiving the subsequent segment, the learning terminal recalculates its SHA-256 digest value; if the verification matches, it returns a success message; if the verification does not match, it deletes the locally received data and requests the resource push server to resend the same subsequent segment.

[0179] This step uses lock-release control to convert the determination result of the candidate learning branch into the permission to send resource fragments, which can avoid releasing two or more mutually exclusive subsequent fragments due to request retries, concurrent processing, or repeated submission of actual response flags by the terminal.

[0180] S34. Generate and confirm the learning status.

[0181] Select the candidate learning status branch uniquely corresponding to the matching predicted response tag as the confirmed learning status of the current push decision node: ; in, Push decision node for the current time The corresponding confirmation of learning progress status; The candidate learning status is represented by the selected candidate learning status branch.

[0182] The system updates the branch status of the selected candidate learning information branch to "confirmed," updates other candidate learning information branches within the same learning information ambiguity domain to "unconfirmed," and establishes a confirmation record. The confirmation record includes at least the learner identifier, learning information ambiguity domain identifier, confirmed learning information status, actual response marker, actual-predicted response distance, preceding segment identifier, subsequent segment identifier, model version, resource version, confirmation time, and status. An example of a confirmation record is shown in Table 9.

[0183] Table 9. Examples of Records Confirming Learning Status

[0184] The confirmed learning status represents the resolution result of the previous segment interaction on the ambiguous domain of learning, rather than an immutable permanent learner label. If subsequent learning interactions are inconsistent with the confirmed learning status, the multi-dimensional learning analysis model can still regenerate multiple candidate learning branches, thereby avoiding the long-term solidification of an erroneous learning status due to a single accidental operation.

[0185] S35. Generate a new ambiguous domain of learning status based on the confirmed learning status.

[0186] At the next push decision node, the system reads the most recent valid confirmation record, uses the confirmed learning status as the current learning baseline status input for the multi-dimensional learning analysis model, and uses the newly acquired learning interaction results since the confirmation time as the incremental interaction sequence input.

[0187] S351, Current Learning Baseline Status Input Interface.

[0188] To use the confirmed learning status without changing the shared temporal coding layer structure in step S1, two baseline state projection layers are set up to initialize the hidden states of the two-layer gated loop units respectively: ; in, For the next push decision node The initial hidden state of the layer-gated loop unit; This refers to the gated loop unit layer number; For the first The connection weight matrix of each baseline state projection layer; To confirm the candidate learning status branch codes; For the first The bias vector of each reference state projection layer; It is a hyperbolic tangent activation function, which compresses the projection result to a numerical range of -1 to 1.

[0189] The output dimensions of the two baseline projection layers are the same as the hidden state dimensions of the corresponding gated recurrent unit layers. The projection results only replace the original zero initial state and do not change the connection relationships of the shared temporal coding layer, knowledge mastery dimension branch, ability level dimension branch, learning style dimension branch, and joint state output layer. This allows the new learning interaction results to be interpreted within the context of the confirmed learning status.

[0190] S352, Training of the baseline state initialization interface.

[0191] The training samples consist of historically confirmed learning statuses, learning interaction sequences after the confirmation time, and verified learning statuses obtained through subsequent assessments or teacher review. During training, historically confirmed learning statuses are converted into corresponding candidate learning branch codes, and initial hidden states are generated through two baseline state projection layers. Subsequent learning interaction sequences are then processed by a multi-dimensional learning analysis model. The baseline state projection layer is trained using the loss between the model output and the verified learning statuses. The model parameters are jointly updated with the same loss function, optimizer, learning rate, training batch size, and early stopping condition as in step S1, without additional manual loss weights.

[0192] To prevent the model from over-relying on confirmed learning status and ignoring new learning interactions, each training batch is configured with samples having historical confirmed learning status and samples without such status in a 5:1 ratio; for the latter, the initial hidden state of the gated recurrent unit is set to zero. The 5:1 ratio was determined by comparing 3:1, 5:1, and 10:1 on the validation set: selecting the proportion where the joint state recognition macro-average F1 score is highest, and the macro-average F1 score of samples without confirmed learning status decreases by no more than 1 percentage point relative to the full training. In this embodiment, 5:1 satisfies this condition.

[0193] S353, the execution process of the next push decision node.

[0194] When the next push decision node arrives, the system organizes the newly acquired learning interaction results according to the event occurrence time and local sequence number, removes events that were used before the confirmation record was established, and calls the complete processing procedure of step S1: ; in, A new learning ambiguity domain is generated for the next push decision node; This refers to newly acquired learning interaction results after the confirmation time. The function for generating the learning ambiguity domain corresponding to step S1 is executed as follows: it initializes the multi-dimensional learning analysis model by confirming the learning status, infers candidate learning status by combining the newly acquired learning interaction results, calculates the confidence level, filters candidate learning branches with different resource needs, and forms a new learning ambiguity domain.

[0195] When the newly acquired learning interaction result is consistent with the confirmed learning status, the current learning status baseline state allows the model to prioritize inference within its neighboring state range, thereby reducing unnecessary candidate learning status branches; when the newly acquired learning interaction result conflicts with the confirmed learning status, step S1 still retains multiple candidate learning status branches according to the differences in confidence and resource requirements, without forcing the use of the original confirmed learning status.

[0196] This invention first selects a unique candidate learning branch by marking the actual response of the preceding segment, then releases only the corresponding subsequent segments in a transactional manner, and finally sends the confirmed learning status back to the next push decision node as the model baseline. The same learning interaction result may be generated by multiple learning states with different resource requirements, and the preceding segments with mutually exclusive predictive responses can provide proactive discrimination information for this ambiguity; therefore, it not only reduces the exposure of erroneous resource segments, but also enables subsequent learning analysis to inherit the already resolved state information, while retaining the ability to reform the learning ambiguity domain when new evidence appears.

[0197] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0198] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic learning resource push and control system based on a multi-dimensional learning analysis model, running on a learning resource push server and communicating with learning terminals, characterized in that, The system is configured to perform the following steps: At the push decision node, a multi-dimensional learning analysis model is used to retain multiple candidate learning branches that can explain the same learning interaction results but have different resource requirements, and these multiple candidate learning branches form a learning ambiguity domain. Based on the aforementioned learning ambiguity domain, target learning resources that enable each candidate learning branch to generate different predicted response labels are selected from the candidate learning resources. The branch continuation relationship between each predicted response label and the corresponding subsequent segment is established to form an ambiguity resolution push order. Send the preceding segment of the ambiguity resolution push order to the learning terminal, match the actual response marker returned by the learning terminal with the predicted response marker, select the corresponding candidate learning branch and send only the subsequent segment corresponding to the candidate learning branch, confirm the selected candidate learning branch as the confirmed learning status, and use the confirmed learning status as the input for the next push decision node to form the learning ambiguity domain.

2. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 1, characterized in that, The multi-dimensional learning analysis model includes a shared temporal encoding layer, a knowledge mastery branch, an ability level branch, a learning style branch, and a joint state output layer. The shared temporal encoding layer consists of two layers of gated loop units connected in sequence, used to encode the model input sequence arranged according to the learning interaction order and output a shared interaction representation. The knowledge mastery branch, ability level branch, and learning style branch output the state probability of the corresponding dimension based on the shared interaction representation. The joint state output layer is used to concatenate the shared interaction representation with the state probabilities of each dimension and output the joint state probability corresponding to each dimension state combination, so as to determine the state combination of each dimension as the candidate learning state.

3. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 2, characterized in that, The retention of multiple candidate learning branches that can explain the same learning interaction result but have different resource requirements specifically includes: Obtain the current learning interaction result, use the multi-dimensional learning analysis model to infer multiple candidate learning states that can generate the current learning interaction result based on the combination of states of each dimension, and calculate the confidence level of each candidate learning state. Candidate learning statuses with confidence levels higher than a preset threshold and whose required push resource types are different from each other, as determined by preset resource mapping rules, are selected as candidate learning status branches.

4. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 3, characterized in that, All the selected candidate learning branches are constructed into a set of states representing the uncertainty of the learner's current learning situation, and the set of states is marked as the learning situation ambiguity domain.

5. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 1, characterized in that, Obtain the learning ambiguity domain, and for each candidate learning branch therein, use the multi-dimensional learning analysis model to predict the response of the candidate learning branch to each resource segment in the candidate learning resource library, and generate candidate predicted response tags; select resource segments that can generate different predicted response tags for different candidate learning branches as the pre-segment segments, and determine the complete learning resource including the pre-segment segments as the target learning resource.

6. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 5, characterized in that, Based on the resource requirements of each candidate learning branch, a subsequent segment that matches the candidate learning branch and is logically and content-connected with the preceding segment is determined from the target learning resources. A unique mapping is established between each predicted response tag and the corresponding subsequent segment, forming the branch continuation relationship.

7. A learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 6, characterized in that, The preceding fragment, all predicted response markers, and their respective mapped subsequent fragments are organized according to the branch continuation relationship to form the ambiguity resolution push order.

8. The learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 1, characterized in that, The preceding segment in the ambiguity resolution push order is sent to the learning terminal, and the learning terminal receives the actual response markers generated by the learner's interactive behavior in response to the preceding segment. The actual response marker is compared with each predicted response marker carried in the ambiguity resolution push order to determine the predicted response marker that matches the actual response marker.

9. A learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 8, characterized in that, Obtain the branch continuation relationship, and based on the branch continuation relationship, select the candidate learning branch uniquely mapped by the matching predicted response tag, and send only the subsequent segments corresponding to the candidate learning branch to the learning terminal.

10. A learning resource dynamic push and control system based on a multi-dimensional learning analysis model according to claim 9, characterized in that, The selected candidate learning status branch is confirmed as the learning status status. At the next push decision node, the confirmed learning status is used as the current learning status baseline input of the multi-dimensional learning analysis model. Combined with the newly acquired learning interaction results, the process of forming the learning status ambiguity domain is re-executed to generate a new learning status ambiguity domain.

Images