Learning state data processing method and system

Abstract

The invention discloses a learning state data processing method and system. The method comprises the steps of obtaining a learning state inference result based on multi-source heterogeneous original data, and converting the learning state inference result into a learning state object; aggregating a plurality of learning state objects of the same type, which are continuous in time sequence and meet a semantic coherence preset condition, into a learning state change object representing a learning state change process; aggregating a plurality of learning state change objects having a causal relationship or a situational dependency relationship through statistical verification into a learning state change chain object representing a learning state evolution path; and combined analysis of learning state change chain objects in different dimensions and tracing between learning state inference results and original evidences can be carried out. According to the method, hierarchical and semantic organization of mass learning state data is realized, and learning state evolution rules and key turning points are deeply mined, so that'learning process understanding 'is realized.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to a method and system for processing learning state data. Background Technology

[0002] With the construction and popularization of smart campuses, online learning platforms, smart terminals, and multimodal interaction tools are able to continuously collect large amounts of diverse and heterogeneous student learning process data, providing a rich data foundation for comprehensively understanding students' learning status. However, current student learning process data management systems generally still focus on raw data collection and classified storage, adopting a traditional model of classification and storage according to data source and format. This leads to the formation of "data silos" and repetitive analysis of diverse and heterogeneous learning process data (such as text, audio and video, operation logs, etc.), making it difficult to achieve cross-modal integration and unified management, and failing to support higher-level cognitive needs such as analysis of learning status change trends. Therefore, a new data organization paradigm is urgently needed to break through the traditional "data-centric" management model and shift to a new learning process data management model "centered on learning status." Summary of the Invention

[0003] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a learning state data processing method and system that systematically depicts the process and evolution path of student state changes, thereby enhancing the semantic expression capability and cross-modal fusion efficiency of the learning analysis system. This provides data support for achieving intelligent educational applications that "understand the learning process" rather than merely "record learning behavior."

[0004] According to one aspect of this application, a learning state data processing method of the present invention includes the following steps: Collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects, which include multiple different types. Multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions are aggregated into a learning state change object that represents the learning state change process. The learning state change object includes a variety of different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type. Multiple learning state change objects that have been verified to have causal or context-dependent relationships are aggregated into a learning state change chain object that represents the learning state evolution path, based on the causal or context-dependent relationships.

[0005] Furthermore, the learning state data processing method also includes the following steps: The learning state change chain object includes multiple different types. Time calibration and semantic alignment are performed on different types of learning state change chain objects to extract multiple different types of learning state change chain objects that are in the same period or causally adjacent, and the dynamic interaction relationship between multiple different types of learning state change chain objects is analyzed.

[0006] Furthermore, the learning state data processing method also includes the following steps: The multi-source heterogeneous raw data used for learning state inference is transformed into structured evidence units, each of which has a unique evidence anchor identifier. Based on the utilization relationship between the learning state inference results and the multi-source heterogeneous original data, a mapping relationship between structured evidence units and learning state inference results is constructed. Based on the correspondence between the learning state inference results and learning state objects, a mapping relationship between the learning state inference results and learning state objects is constructed.

[0007] Furthermore, learning state objects include at least three types: cognitive state objects, emotional state objects, and behavioral state objects; The objects of learning state changes include at least three types: objects of cognitive state changes, objects of emotional state changes, and objects of behavioral state changes; Learning state change chain objects include at least four types: cognitive development learning state change chain objects, emotional evolution learning state change chain objects, behavioral performance learning state change chain objects, and mixed learning state change chain objects.

[0008] Furthermore, determining the object type of a learning state change chain includes the following steps: For each state change object in a learning state change chain, it is assigned to the cognitive, emotional, or behavioral dimension according to its learning state change object type. The learning state change objects of the same dimension are sorted in chronological order to form sub-sequences of learning state change chains for each dimension. The dominance score of each sub-sequence of learning state change chains is calculated. Based on the dominance scores of each dimension, when the ratio of a certain dimension score to the scores of other dimensions is higher than a preset threshold, it is determined to be a single dimension type. When the ratio of each dimension score to the scores of other dimensions is not higher than the preset threshold and all show a complete evolutionary process, it is determined to be a mixed type.

[0009] Furthermore, the step of obtaining the learning state inference result based on multi-source heterogeneous raw data includes the following steps: Multiple different models are used to infer the multi-source heterogeneous raw data, and the learning state inference results of each model are output. When multiple different models provide the same or complementary inferences about the learning state of the same student within the same time period, the overall confidence of the inferences is increased, and the confidence of the learning state of the learning state object is consistent with the overall confidence of its corresponding inference results. If multiple different models provide conflicting inferences about the learning state of the same student within the same time period, an anomaly warning is triggered, and the fusion of learning state inference results or manual adjustment of learning state inference results is initiated.

[0010] Furthermore, determining the causal relationship includes the following steps: Pattern matching is performed based on a pre-defined causal knowledge base of the learning process. This causal knowledge base includes multiple causal rule templates. Each causal rule template defines a premise pattern and a target pattern that conform to the causal relationship. The "premise pattern" is a specific sequence or combination of one or more learning state change objects, and the "target pattern" is a specific sequence or combination of one or more other learning state change objects. The learning state change object that occurs earlier in time is structurally aligned with the "premise pattern" to compare whether the state type, state characteristics, and evolution direction are consistent within the tolerance range. The learning state change object that occurs later in time is structurally aligned with the "target pattern" to compare whether the state type, state characteristics, and evolution direction are consistent within the tolerance range. If both conditions are met, it is determined that the learning state change object that occurs earlier in time and the learning state change object that occurs later in time satisfy the causal relationship candidate type. The learning state change objects that meet the causal relationship candidate type are verified for causal relationship based on the constructed Bayesian network model. The learning state change objects that are verified to have causal relationship are determined to have causal relationship.

[0011] Furthermore, the determination of the causal relationship or contextual dependency includes the following steps: extracting multidimensional contextual variables during the learning process, introducing these contextual variables as moderating variables into the same dynamic Bayesian network model, constructing an extended Bayesian network model, inputting the learning state change object into the constructed extended Bayesian network model for contextual dependency verification, and determining that the learning state change object that is verified to have a contextual dependency relationship.

[0012] According to another aspect of this application, a learning state data processing system of the present invention includes the following steps: The learning state object acquisition unit is used to collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects. The learning state objects include multiple different types. The learning state change object aggregation unit is used to aggregate multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions into a learning state change object representing the learning state change process. The learning state change object includes a variety of different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type. The learning state change object aggregation unit aggregates multiple learning state change objects that have been verified to have causal or context-dependent relationships into a learning state change chain object that represents the learning state evolution path, based on the causal or context-dependent relationships.

[0013] Furthermore, the learning state change chain object includes several different types, and the learning state data processing system also includes: The dynamic interaction relationship analysis unit is used to perform time calibration and semantic alignment on different types of learning state change chain objects, extract multiple different types of learning state change chain objects that are in the same period or causally adjacent, and analyze the dynamic interaction relationships between multiple different types of learning state change chain objects.

[0014] The evidence index unit is used to transform the multi-source heterogeneous raw data used for learning state inference into structured evidence units. Each structured evidence unit has a unique evidence anchor identifier. Based on the utilization relationship between the learning state inference result and the multi-source heterogeneous raw data, a mapping relationship between the structured evidence unit and the learning state inference result is constructed. Based on the correspondence between the learning state inference result and the learning state object, a mapping relationship between the learning state inference result and the learning state object is constructed.

[0015] In summary, the learning state data processing method and system provided in this application have beneficial effects compared with the prior art: (1) To achieve a systematic characterization of the process and evolution path of student state changes, and to realize the association and organization from raw learning behavior data to higher-order learning state semantics, effectively meeting the cognitive and analytical needs of the learning state evolution law "from point to line, from static to dynamic, from result to attribution". By performing structured modeling and hierarchical organization of massive, cross-modal, and multi-temporal granular learning state information, it supports multi-dimensional tracing of key learning state turning points, and improves the value density, reusability and intelligent service level of learning process data.

[0016] (2) By organizing the structured evidence units and the learning state inference results, standardized evidence data blocks that can be directly called and are accurately located are provided for the upper-level learning state inference, and the traceability from the learning state inference results to all supporting evidence can be realized. Attached Figure Description

[0017] Figure 1 This is a flowchart of the learning state data processing method provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description, in conjunction with the accompanying drawings and examples, further clarifies this application. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0019] like Figure 1 As shown, an embodiment of the present invention provides a learning state data processing method, which includes the following steps: Step 1: Collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects, which include multiple different types.

[0020] By deploying multimodal sensors and a data management system in a smart teaching environment, raw data including video streams, audio streams, text interaction data, operation logs, environmental perception data, and physiological data are collected. Based on the data modality, a pre-trained large model tailored to that modality is adaptively invoked to process the raw data from each modality, outputting learning state inference results containing information such as learning state type, learning state label, inference result confidence level, and evidence anchor points. Through semantic alignment and confidence fusion mechanisms, the multi-source inference results of the learning state are normalized into learning state objects conforming to a predefined structure, serving as the basic data unit for subsequently constructing the learning state evolution path.

[0021] Learning status represents a snapshot of a student's performance at a certain moment or in a short window, covering dimensions such as cognition, emotion, and behavior. For example, cognitive status can include descriptions of comprehension level (such as superficial comprehension, deep comprehension, conceptual misunderstanding, etc.) and thinking activities (such as exploration, reasoning, etc.). Emotional status can include descriptions of emotional expression (such as confidence, anxiety, curiosity, etc.) and motivational tendency (such as intrinsic motivation, extrinsic stimulation, etc.). Behavioral status can include descriptions of participation mode (such as active participation, passive observation, etc.) and interaction mode (such as asking questions, using tools, etc.).

[0022] The key processing steps in step 1 are explained in detail below.

[0023] (1) Structured processing of learning state inference results The specific implementation process of the structured processing of student learning status inference results is as follows: establish semantic description rules for standard data of learning status, and convert the learning status inference results into standard data of learning status with predefined semantic labels according to the semantic description rules of standard data of learning status.

[0024] In one embodiment, the semantic description rules for standard learning state data are as follows: Learning State (semantic label "LS") includes two main categories: Learning State Semantics (semantic label "LSSemantics") and Learning State Metadata (semantic label "LSMetadata"). Learning State Semantics includes Learning State Subject (semantic label "LSSubject"), Learning State Type (semantic label "LSType"), Learning State Label (semantic label "LSLabel"), and Learning State Confidence (semantic label "LSConfidence"). The Learning State Subject is the individual student, uniquely identifying their learning state. The Learning State Type is a semantic classification of the student's learning state, including dimensions such as Cognitive State, Affective State, and Behavioral State, and supports expansion. The Learning State Label is the student's specific state within a particular learning state dimension; for example, in the Cognitive State, it can be further subdivided into "Confused" and "Understanding." The learning state confidence score represents the reliability of the learning state identification result and can be numerical or enumerated. Learning state metadata includes the learning state identifier (LSIdentifier, semantic label "LSIdentifier") and the learning state time (LSTime, semantic label "LSTime"). The learning state identifier is a unique code for the instance of the learning state. The learning state time is the moment of occurrence or duration of the learning state. The semantic description rules for learning states are shown in Table 1.

[0025] Table 1. Semantic Description Rules of Learning States ; (2) Convert the standard data of learning status into learning status objects. A learning state object is created. Based on the semantic description rules described in step (1), learning state object attributes are created, and values ​​are assigned to the learning state object attributes according to the learning state standard data. The learning state object belongs to the meta-object of the learning state, that is, the object that describes the learning state at the smallest granularity. The learning state object includes three major categories: basic learning state attributes, feature attributes, and confidence attributes. Basic learning state attributes include learning state identifier and learning state description. Feature attributes include learning state subject, learning state type, learning state label, and learning state occurrence time. Confidence attributes include learning state confidence, source of judgment basis, and state judgment time. Among them, learning state confidence represents the reliability of the learning state recognition result.

[0026] The attribute definitions of the learning state object are shown in Table 2.

[0027] Table 2. Definition of Learning State Object Attributes ; In one embodiment, semantic alignment and collaborative verification of multi-model inference results can also be constructed. To improve the accuracy and robustness of learning state inference, semantic alignment and collaborative verification of multi-model inference results are supported. Multiple different models are used to infer multi-source heterogeneous raw data, and the learning state inference result corresponding to each model is output. Based on a unified semantic label space for learning state inference results, spatiotemporal alignment and semantic normalization are performed on the learning state inference results output by different models for the same student within the same time period to identify cross-modal consistency or contradictory signals. When multiple models point to the same or complementary learning state inference results from different data sources, the overall confidence of the learning state inference result is improved through a consensus reinforcement mechanism, and the learning state confidence of the learning state object corresponding to the learning state inference result is consistent with the overall confidence of the learning state inference result. If there is a conflict between the learning state inference results pointed to by multiple models from different data sources, an anomaly warning is triggered and a fusion strategy (such as weighted voting, Bayesian inference) is initiated to fuse multiple learning state inference results or manually adjust the learning state inference results. The learning state confidence attribute also includes an extended attribute, which is used to adjust the weight when fusing multiple learning state inference results. By fusing the inference results of multiple models, we can transform from "black box inference" of a single model to "white box collaboration" of multiple models, achieve cross-verification and collaborative error correction of learning state inference results, and improve the credibility and interpretability of learning state inference results.

[0028] Step 2: Aggregate multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions into a learning state change object representing the learning state change process. The learning state change object includes multiple different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type.

[0029] Learning state changes are used to characterize the dynamic transition process between adjacent learning states, focusing on describing the driving mechanism of the change.

[0030] The key processing steps are explained in detail below.

[0031] (1) Define the semantic description rules for changes in learning state Learning State Change (LSChange) records the process by which a student transitions from one basic or composite learning state to another within a specific time period. In one embodiment, its semantic description includes the Learning Process Starting State (LPStartingState), the Learning Process Intermediate State (LPIntermediateState), the Learning Process Target State (LPTargetState), the Learning State Change Type (LSChangeType), the Learning State Change Time Window (LSChangeTimeWindow), the Learning State Change Intensity (LSChangeIntensity), and the Learning State Trigger Event (LSTriggerEvent). The Learning Process Starting State and Target State characterize the beginning and end points of the learning state transition and both reference unique identifiers of existing learning state instances. The learning state change time window is the start and end time period of the learning state change. The intensity of the learning state change is used to quantify the magnitude of the change and can be numerical or enumerated. The learning state triggering event describes the external variables that cause the learning state to change, including event type, source, timestamp, and other information. The semantic description rules for learning state changes are shown in Table 3.

[0032] Table 3. Semantic Description Rules for Learning State Changes ; (2) Create a learning state object A learning state object is created, and its attributes are created based on the semantic description rules in (1). The attributes of the learning state change object include basic attributes of learning state change, process attributes of learning state change, and related attributes of learning state change. The basic attributes of learning state change include change object identifier, change object description, change object type, and change object subtype. The process attributes of learning state change include starting state, intermediate state, target state, change start time, change end time, change duration, and change intensity. The related attributes of learning state change include trigger event identifier, predecessor change object, successor change object, and evidence chain identifier list. The attribute definitions of the learning state change object are shown in Table 4.

[0033] Table 4. Definition of Attributes for Learning State Change Objects ; (3) Aggregate and assign values ​​to the properties of the learning state change object. The process of transforming learning state objects into learning state change objects uses aggregation mapping, which can be implemented using a linked list of mapping relationships. The aggregation mapping relationship is characterized by vertical hierarchical integration. First, by comparing and analyzing the temporally consecutive learning state objects for thematic consistency, semantic coherence, etc., and when the aggregation threshold condition is met, they are integrated into learning state change objects, recording their initial state, intermediate state, target state, and change process. Thematic consistency means that the learning state objects are of the same type. Learning state changes can be categorized into cognitive state changes, emotional state changes, and behavioral state changes based on dimensions. Each category can be further subdivided into various subtypes based on the type of change, such as positive leaps, negative regressions, cyclical fluctuations, and stagnation. Positive leaps refer to learning state changes that show a trend towards a better or higher level, and can include descriptions of cognitively driven and emotionally induced states. Negative regressions refer to learning state changes that show a trend towards a worse or lower level. Stagnation refers to learning state changes that do not show a clear directional evolution and linger at a certain level or pattern for a long time. Negative regression / stagnation can include descriptions of resource depletion and motivation decay. Fluctuations refer to learning state changes that show non-monotonic ups and downs rather than a clear unidirectional trend, and can include descriptions of constructive fluctuations (such as strategy exploration, motivation internalization, and stress adaptation) and destructive fluctuations (such as emotional vortexes and goal erosion). Semantic coherence preset conditions are defined for each type of learning state change object. If multiple sequentially occurring learning state objects satisfy the semantic coherence preset conditions corresponding to one type of learning state change object, then these multiple sequentially occurring learning state objects are aggregated into a learning state change object of the corresponding type. The steps for obtaining the semantic coherence of multiple learning state objects include: first, obtaining the semantic features of each learning state object according to the attributes of the multiple learning state objects using a preset semantic feature extraction method; and then, obtaining the semantic coherence of the multiple learning state objects according to the semantic features of the multiple learning state objects using a preset semantic coherence calculation method, such as a similarity calculation formula for semantic features.

[0034] Assign values ​​to the various attributes of the learning state change objects based on the aggregated learning state change objects.

[0035] Step 3: Based on the causal relationship, aggregate multiple learning state change objects that have been statistically verified to have causal or context-dependent relationships into a learning state change chain object that represents the learning state evolution path.

[0036] The semantic category labels of the learning state change sequence characterize the overall performance pattern and development trend of a learning process, such as positive development, abnormal deviation, external dependence, and autonomous efficiency. All pre-trained large models involved in learning state inference must map their outputs to a unified semantic label space to provide a benchmark reference for subsequent semantic alignment and collaborative verification of learning state inference results.

[0037] The key steps are explained in detail below.

[0038] (1) Establish rules for describing the relationships in a learning state change sequence. A learning state change sequence (LSChangeSequence) consists of a set of learning state change objects arranged in chronological order, each referencing a unique identifier of an existing learning state change object instance. The relationships between objects within the sequence include, but are not limited to, causal relationships (LSCausalRelation) and contextual dependencies (LSContextualDependency). Causal relationships mainly describe the causal influence of a previous learning state change on a subsequent learning state change, including the preceding change object, the target change object, the type of causal relationship (such as "cognitive drive", "emotional trigger", "action assistance", etc.), the causal strength (i.e., the score of causal strength, which can be numerical or enumerated), and the evidence chain (i.e., the reference of the multimodal evidence chain supporting the judgment). Contextual dependence primarily characterizes the semantic relationships between learning state change events and external teaching elements, including the preceding change object, the target change object, the type of contextual dependence relationship (such as "learning activity dependence," "teaching scenario dependence," "learning resource dependence," "teacher intervention dependence," etc.), the contextual dependence object, the strength of contextual dependence, and the chain of evidence. The rules for describing the association relationships of learning state change sequences are shown in Table 5.

[0039] Table 5. Description Rules of the Correlation between Learning State Change Sequences ; (2) Create learning state change chain object attributes. Based on the semantic description rules in (1), create learning state change chain object attributes. A learning state change chain is an ordered sequence of existing learning state change object instances arranged in chronological order. Its core value lies not only in the linear combination of multiple learning state transition events, but also in revealing the deep dynamic correlation mechanism between these learning state changes. This object model organizes discrete learning change events into cognitive development paths, emotional evolution trajectories, or behavioral performance patterns with educational interpretability by introducing causal relationships and contextual dependence. The learning state change chain object attributes include basic learning state change chain attributes, structural learning state change chain attributes, and internal correlation attributes of the learning state change chain.

[0040] The basic attributes of the learning state change chain include the change chain object identifier, change chain object description, change chain object type, as well as the student subject to which the learning state change object sequence belongs, the learning context, the sequence generation mode, and the sequence version information.

[0041] Learning state change chains can be categorized into at least four types: cognitive developmental learning state change chains, affective evolutionary learning state change chains, behavioral performance-based learning state change chains, and hybrid learning state change chains. Cognitive developmental learning state change chains refer to the evolution of learning states primarily driven by changes in psychological representations such as knowledge structure, thinking skills, or metacognitive abilities. Affective evolutionary learning state change chains refer to the evolution of learning states primarily driven by fluctuations, adjustments, or transformations in emotional states such as emotions, motivation, and attitudes. Behavioral performance-based learning state change chains refer to the evolution of learning states primarily driven by changes in explicit operations, action sequences, or interaction patterns. Hybrid learning state change chains refer to complex processes in which multiple dimensions, including cognition, emotion, and behavior, co-evolve and are deeply coupled during the evolution of learning states, making them difficult to fully represent using a single dimension.

[0042] The type of an object in a learning change chain is determined based on its overall evolutionary characteristics, and specifically includes the following steps: (a) The learning state change chain is decomposed into dimensions such as cognitive development, emotional evolution, and behavioral pattern, and the evolutionary dominance score of each dimension is calculated. For each state change object in the learning state change chain, it is assigned to the cognitive, emotional, or behavioral dimension according to its learning state change object type, and the learning state change objects of the same dimension are sorted by time to form the sub-sequence of the learning state change chain of each dimension. The dominant score is calculated by weighting three aspects: (1) the proportion of the learning state change chain sequence in the learning state change chain (quantity or duration); (2) the completeness of the internal state evolution of the learning state change chain sequence in the dimension, that is, assessing whether the learning state change chain sequence in the dimension constitutes a semantically closed path with a clear evolutionary direction, including whether there is a significant difference between the starting state and the ending state of the learning state change chain sequence in the dimension (such as cognition from "wrong" to "correct", emotion from "anxiety" to "calm"), and whether there is a reasonable transition link in the middle (such as "anxiety → awareness → relaxation → calm"). The matching degree is calculated in combination with the preset typical development template of the dimension. The higher the matching degree, the higher the completeness score; (3) the driving strength of the learning state change chain sequence in the dimension on the learning state change chain sequence in other dimensions. Based on the temporal correlation and statistical dependency analysis, if the state change of a certain dimension significantly increases the probability of the occurrence of a specific state of another dimension within a predetermined time window, it is considered that there is an influence relationship. Its strength is quantified by Granger causality test or cross-lag correlation coefficient, and the comprehensive influence score is obtained by weighting with domain knowledge. The dominant scores for each dimension are generated by weighting and integrating the three aspects.

[0043] (b) Based on the dominant scores of each dimension, when the score of a single dimension is significantly higher than that of other dimensions (e.g., the ratio of the score of a certain dimension to the scores of other dimensions is higher than a preset threshold), it is determined to be a single type; when the scores of multiple dimensions are close (e.g., the ratio of the score of a certain dimension to the scores of other dimensions is not higher than a preset threshold) and all show a complete evolution process, it is determined to be a hybrid type; finally, a learning state change chain object type is generated.

[0044] The structural attributes of the learning state change chain include the main body of the change sequence and the time window of the change sequence.

[0045] The internal attributes of the learning state change chain include causal relationship sequences and context dependency sequences. Each causal relationship includes the preceding change object, the target change object, the causal relationship type, the causal strength, and the evidence chain. Each context dependency includes the preceding change object, the target change object, the context dependency relationship type, the context dependency strength, and the evidence chain. The attribute definitions of the learning state change chain objects are shown in Table 6.

[0046] Table 6. Attribute Definitions of the Learning State Change Chain Object ; (3) Aggregate and assign values ​​to the properties of the learning state change chain object. The learning state change object to the learning state change chain object uses aggregation mapping, which can be implemented using a mapping relationship linked list.

[0047] A learning state change chain object comprises multiple learning state change objects that can be related by cause, context, or a mixture of both. By integrating prior knowledge rules from the education domain with data-driven sequence mining algorithms, causal or contextual dependencies among multiple learning state change objects are uncovered, enabling automatic identification and dynamic organization of the internal association attributes of the learning state change chain. Multiple learning state change objects that have been statistically verified to have causal or contextual dependencies are aggregated into a learning state change chain object representing the learning state evolution path, based on the temporal sequence in the causal relationship and the logical association in the contextual dependency.

[0048] The determination of the correlation includes the following steps: (a) Pattern matching based on a pre-defined causal knowledge base of the learning process. The knowledge base stores causal rule templates encapsulated in the form of "If<Prerequisite Pattern>Then<Target Pattern>With<Causal Type>". Here, "Prerequisite Pattern" is a specific sequence or combination of one or more learning state change objects; "Target Pattern" is a specific sequence or combination of one or more other learning state change objects; and "Causal Type" includes types such as cognitive drive, emotion induction, and behavioral assistance. The matching process is as follows: the preceding (i.e., the earlier in time) learning state change sequence (i.e., "preceding change object") is structurally aligned with the "Prerequisite Pattern," and the consistency of state type, state characteristics, and evolution direction is compared within the tolerance range; simultaneously, the target (i.e., the later in time) learning state change sequence (i.e., "target change object") is structurally aligned with the "Target Pattern" in the rule, and the consistency of state type, state characteristics, and evolution direction is compared within the tolerance range. If both conditions are met, the causal relationship is determined to be valid, and the corresponding causal relationship type is labeled for the internal correlation attributes of the change chain. This process supports weighted matching and multi-rule competitive adjudication to handle complex causal relationships in real-world scenarios. For example, when students experience conceptual misunderstandings or practical setbacks, they often trigger help-seeking behavior, corresponding to a cognitively driven causal relationship; timely positive feedback gradually enhances self-efficacy, while prolonged high anxiety inhibits deep cognitive processing, corresponding to an emotion-induced causal relationship; repeated practice accelerates the internalization of concepts, corresponding to a behaviorally assisted causal relationship. The identified preceding and target change objects are combined and pattern-matched with a rule base. If the preconditions are met, the corresponding causal relationship candidate type is assigned.

[0049] (b) After initially assigning causal relationship types based on a pre-defined causal knowledge rule base for the learning process, a dynamic Bayesian network model is further constructed to probabilistically and hierarchically verify candidate causal relationships. This process does not involve blind matching of any two learning state change objects. Instead, it takes the preceding learning state change objects and the target learning state change objects, which satisfy the causal relationship and are output from the causal knowledge base matching, as inputs. These are modeled as temporal nodes, and the theoretical confidence of the corresponding rules in the rule base is weighted and fused with the actual matching degree of the data to generate the prior existence probability of the directed edge in the dynamic Bayesian network pointing from the preceding change object to the target change object. The posterior existence probability of the causal path is calculated, and the predictive ability and temporal dependence strength of the preceding change on the target change are evaluated. Finally, statistical verification of the causal relationship is completed by setting a threshold. In other words, the causal relationship types obtained in the previous step based on template matching are verified using a Bayesian network model. Learning state change objects that pass verification are determined to have a causal relationship.

[0050] (c) To further characterize the applicable boundaries of causal relationships, multi-dimensional situational variables (such as "learning activity - task difficulty", "learning activity - collaboration mode", "teaching scenario - intelligence level", "teacher intervention - intervention intensity") are extracted during the learning process and introduced as moderating variables into the same dynamic Bayesian network model. Introducing these situational variables as moderating variables means that in the probabilistic inference of the dynamic Bayesian network model, the probability of the influence of the preceding changed object on the target changed object is modeled conditioned on the situational variables. For example, in a "high collaborative atmosphere", the conditional probability of "cognitive confusion" triggering "help-seeking behavior" is significantly higher than the corresponding probability in a "low collaborative atmosphere". Based on this extended structure, the system jointly infers the joint posterior distribution of each situational variable and causal path. When the posterior probability of a causal path in a specific situation is significantly higher than in other situations, the causal relationship is determined to have situation-dependent characteristics, and the corresponding situational label is automatically assigned. Furthermore, teachers or educational experts can manually annotate the observed sequences of learning state changes and their contextual situations through a visual analysis interface. This expert feedback will be used to dynamically update the prior parameters of the dynamic Bayesian network model and the upper-level causal rule base.

[0051] By creating learning state objects, learning state change objects, and learning state change chain objects and their attributes, and constructing a mapping relationship chain between different objects of the learning state "meta-state-state change-state change chain", we can achieve structured modeling and hierarchical organization of massive, cross-modal, and multi-temporal granular learning state information.

[0052] The learning state data processing method of this embodiment of the invention may further include: Step 4: The learning state change chain objects include multiple different types. Time calibration and semantic alignment are performed on the different types of learning state change chain objects. Multiple different types of learning state change chain objects that are in the same period or causally adjacent are extracted, and the dynamic interaction relationship between multiple different types of learning state change chain objects is analyzed.

[0053] Extracting multiple learning state change chains of different types that occur simultaneously or are causally adjacent can be achieved using a mapping relationship linked list, which represents a horizontal multi-dimensional fusion and supports cross-dimensional correlation analysis of learning state evolution paths.

[0054] First, time calibration and semantic alignment are performed on learning state change chain objects of different types (i.e. different dimensions, such as cognition, emotion, and behavior). Multiple different types of learning state change chain objects that are in the same period or causally adjacent (such as meeting time window constraints, task / spatial association, semantic and pattern matching, or satisfying domain statistical experience, etc.) are extracted.

[0055] Subsequently, multivariate sequence analysis was employed to calculate the dynamic interaction relationships between multiple learning state change chains of different types, such as co-occurrence probability and temporal coupling strength. Co-occurrence probability is the joint probability of multiple learning states occurring simultaneously within adjacent or specified time windows. Temporal coupling strength is the degree of dependence of a learning state in one learning state change chain on a learning state in another learning state change chain. For example, spatiotemporal alignment of "cognitive developmental learning state change chains" and "emotional evolutionary learning state change chains" was performed, and cross-analysis was used to uncover cross-dimensional influence patterns such as "cognitive confusion → heightened emotional anxiety." Based on the analysis results, the developmental trajectories of complex learning states, such as "cognitive-emotional coupling" (e.g., "high confusion-high anxiety") or "behavioral-cognitive synergy" (e.g., "trial and error-understanding"), were identified and formally described. Co-occurrence probability can be calculated using methods such as frequency statistics and weighted overlap. Temporal coupling strength can be calculated using methods such as Granger causality tests and transition entropy.

[0056] By constructing an evolutionary model of cross-dimensional learning states, we can reveal the collaborative evolutionary laws of synchronization, alternation, or mutual exclusion among multidimensional learning states.

[0057] The evolution of learning states exhibits varying focuses across different teaching scenarios. For instance, classroom teaching emphasizes the cognitive leap from "understanding to insight," while mental health monitoring focuses on emotional fluctuations from "anxiety to calm." Therefore, various learning state observation scales, covering micro, meso, and macro levels, can be set according to business needs, including time scales, cognitive / emotional / behavioral scales, and teaching scenario scales. These scales can be incorporated into a unified learning state change chain modeling framework through a unified object identification system. When a learning state change chain at a specific scale is invoked, its original state sequence, supporting evidence chain, and relevant background information can be queried, supporting multi-level, progressive analysis of the learning development process.

[0058] The learning state data processing method of this embodiment of the invention may further include: Step 5: Transform the multi-source heterogeneous raw data used for learning state inference into structured evidence units, each of which has a unique evidence anchor identifier; construct a mapping relationship between structured evidence units and learning state inference results based on the utilization relationship between learning state inference results and multi-source heterogeneous raw data; and construct a mapping relationship between learning state inference results and learning state objects based on the correspondence between learning state inference results and learning state objects.

[0059] That is, to construct a semantic label space for learning state inference results, establish a two-way mapping mechanism between learning state inference results and original data evidence, support semantic alignment and collaborative verification of multi-model inference results, and realize two-way traceability of "inference-evidence".

[0060] The key steps are explained below: (1) Construct a spatiotemporal anchor point index library for multimodal raw evidence. Transform the multi-source heterogeneous raw data supporting learning state inference into structured evidence units that can be accurately located and referenced by the system. For example, based on a unified spatiotemporal benchmark, dynamic slicing and event boundary recognition can be performed on video stream data to generate structured video segments or key video frames carrying timestamps, shot identifiers, geographic tags, event tags, etc.; voice activity detection and event segmentation can be performed on audio streams to extract audio segments with semantic value. Generate globally unique evidence anchor point identifiers for each independent data segment that records key information of the learning state, and record its type, data source, spatiotemporal range, and other core metadata, thereby providing directly callable and accurately located standardized evidence data blocks for upper-level learning state inference.

[0061] (2) Construct a bidirectional mapping relationship chain of "inference-evidence" to achieve bidirectional traceability query. Based on the utilization relationship between the learning state inference results and the multi-source heterogeneous original data, construct a mapping relationship between structured evidence units and learning state inference results. For example, when a multimodal pre-trained large model generates a learning state inference result, automatically record the inference result identifier, result semantic label, pre-trained large model category and version, the complete evidence anchor code list on which it is based, and the contribution weight of each evidence fragment. Establish a bidirectional association table of "inference-evidence" in the database. Its key fields cover inference result identifier, evidence anchor identifier, contribution score and model source, etc., thereby forming a complete mapping chain between inference and evidence at the data structure level. The bidirectional mapping relationship chain of "inference-evidence" not only supports forward tracing of all supporting evidence based on any inference identifier, that is, answering "which evidence analysis led to the learning state inference conclusion", but also supports reverse querying of all related inference conclusions based on a specific evidence fragment, that is, answering questions such as "which other inference results have cited this evidence" or "what other dimensions of inference results have been derived from this evidence". Furthermore, a mapping relationship between the learning state inference results and the learning state objects can be constructed based on the correspondence between the learning state inference results and the learning state objects, thereby forming a complete mapping relationship between structured evidence units, learning state inference results, learning state objects, learning state change objects, and learning state change chain objects. Each level can realize index query of structured evidence units.

[0062] This invention constructs a learning process data management system with structured learning states as the core object. It transforms raw data into standardized learning state units containing fields such as learning state labels, confidence levels, and evidence chains, based on the reasoning capabilities of pre-trained large models oriented to specific modalities. Then, it constructs an evolution path of key nodes in the learning process with the learning state change chain as the core, and establishes a two-way mapping mechanism of "inference-evidence" to ensure the transparency and credibility of learning state determination. This achieves a fundamental leap from "recording learning behavior" to "understanding the learning process," and comprehensively improves the value density, reusability, and intelligent service level of learning process data.

[0063] A learning state data processing system according to an embodiment of the present invention includes: The learning state object acquisition unit is used to collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects. The learning state objects include multiple different types. The learning state change object aggregation unit is used to aggregate multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions into a learning state change object representing the learning state change process. The learning state change object includes a variety of different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type. The learning state change object aggregation unit aggregates multiple learning state change objects that have been statistically verified to have causal or context-dependent relationships into a learning state change chain object that represents the learning state evolution path, based on the causal or context-dependent relationships.

[0064] The learning status data processing system may also include: The dynamic interaction relationship analysis unit is used to perform time calibration and semantic alignment on different types of learning state change chain objects, extract multiple different types of learning state change chain objects that are in the same period or causally adjacent, and analyze the dynamic interaction relationships between multiple different types of learning state change chain objects.

[0065] The evidence index unit is used to transform the multi-source heterogeneous raw data used for learning state inference into structured evidence units, construct a spatiotemporal anchor index library of structured evidence units, construct a mapping relationship between structured evidence units and learning state inference results based on the utilization relationship between learning state inference results and multi-source heterogeneous raw data, and construct a mapping relationship between learning state inference results based on the correspondence between learning state inference results and learning state objects.

[0066] The implementation principle and technical effects of the above-mentioned learning state data processing system are the same as those of the learning state data processing method, and will not be repeated here.

[0067] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing learning state data, characterized in that, Including the following steps: Collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects, which include multiple different types. Multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions are aggregated into a learning state change object that represents the learning state change process. The learning state change object includes a variety of different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type. Multiple learning state change objects that have been verified to have causal or context-dependent relationships are aggregated into a learning state change chain object that represents the learning state evolution path, based on the causal or context-dependent relationships.

2. The learning state data processing method as described in claim 1, characterized in that, It also includes the following steps: The learning state change chain object includes multiple different types. Time calibration and semantic alignment are performed on different types of learning state change chain objects to extract multiple different types of learning state change chain objects that are in the same period or causally adjacent, and the dynamic interaction relationship between multiple different types of learning state change chain objects is analyzed.

3. The learning state data processing method as described in claim 1, characterized in that, It also includes the following steps: The multi-source heterogeneous raw data used for learning state inference is transformed into structured evidence units, each of which has a unique evidence anchor identifier. Based on the utilization relationship between the learning state inference results and the multi-source heterogeneous original data, a mapping relationship between structured evidence units and learning state inference results is constructed. Based on the correspondence between the learning state inference results and learning state objects, a mapping relationship between the learning state inference results and learning state objects is constructed.

4. The learning state data processing method as described in claim 1, characterized in that, Learning state objects include at least three types: cognitive state objects, emotional state objects, and behavioral state objects; The objects of learning state changes include at least three types: objects of cognitive state changes, objects of emotional state changes, and objects of behavioral state changes; Learning state change chain objects include at least four types: cognitive development learning state change chain objects, emotional evolution learning state change chain objects, behavioral performance learning state change chain objects, and mixed learning state change chain objects.

5. The learning state data processing method as described in claim 4, characterized in that, The determination of the object type in a learning state change chain includes the following steps: For each state change object in a learning state change chain, it is assigned to the cognitive, emotional, or behavioral dimension according to its learning state change object type. The learning state change objects of the same dimension are sorted in chronological order to form sub-sequences of learning state change chains for each dimension. The dominance score of each sub-sequence of learning state change chains is calculated. Based on the dominance scores of each dimension, when the ratio of the score of a certain dimension to the scores of other dimensions is higher than a preset threshold, it is determined to be a single dimension type. When the ratio of the scores of each dimension to the scores of other dimensions is not higher than the preset threshold and all show a complete evolution process, it is determined to be a mixed type.

6. The learning state data processing method as described in claim 1, characterized in that, The steps for obtaining the learning state inference result based on multi-source heterogeneous raw data include: Multiple different models are used to infer the multi-source heterogeneous raw data, and the learning state inference results of each model are output. When multiple different models provide the same or complementary inferences about the learning state of the same student within the same time period, the overall confidence of the inferences is increased, and the confidence of the learning state of the learning state object is consistent with the overall confidence of its corresponding inference results. If multiple different models provide conflicting inferences about the learning state of the same student within the same time period, an anomaly warning is triggered, and the fusion of learning state inference results or manual adjustment of learning state inference results is initiated.

7. The learning state data processing method as described in claim 1, characterized in that, The determination of the causal relationship includes the following steps: Pattern matching is performed based on a pre-defined causal knowledge base of the learning process. This causal knowledge base includes multiple causal rule templates. Each causal rule template defines a premise pattern and a target pattern that conform to the causal relationship. The "premise pattern" is a specific sequence or combination of one or more learning state change objects, and the "target pattern" is a specific sequence or combination of one or more other learning state change objects. The learning state change object that occurs earlier in time is structurally aligned with the "premise pattern" to compare whether the state type, state characteristics, and evolution direction are consistent within the tolerance range. The learning state change object that occurs later in time is structurally aligned with the "target pattern" to compare whether the state type, state characteristics, and evolution direction are consistent within the tolerance range. If both conditions are met, it is determined that the learning state change object that occurs earlier in time and the learning state change object that occurs later in time satisfy the causal relationship candidate type. The learning state change objects that meet the causal relationship candidate type are verified for causal relationship based on the constructed Bayesian network model. The learning state change objects that are verified to have causal relationship are determined to have causal relationship.

8. The learning state data processing method as described in claim 7, characterized in that, The determination of the causal relationship or contextual dependency includes the following steps: extracting multidimensional contextual variables during the learning process, introducing these contextual variables as moderating variables into the same dynamic Bayesian network model, constructing an extended Bayesian network model, inputting the learning state change object into the constructed extended Bayesian network model for contextual dependency verification, and determining that the learning state change object that is verified to have a contextual dependency relationship.

9. A learning state data processing system, characterized in that, Including the following steps: The learning state object acquisition unit is used to collect multi-source heterogeneous raw data, obtain learning state inference results based on multi-source heterogeneous raw data, convert the learning state inference results into learning state standard data with predefined semantic labels, and convert the sequence of learning state standard data including multiple different times into multiple learning state objects. The learning state objects include multiple different types. The learning state change object aggregation unit is used to aggregate multiple learning state objects of the same type that are sequentially continuous and meet the semantic coherence preset conditions into a learning state change object representing the learning state change process. The learning state change object includes a variety of different types, and each type of learning state change object has a corresponding semantic coherence preset condition. When multiple sequentially continuous learning state objects meet the semantic coherence preset condition corresponding to one type of learning state change object, then the multiple sequentially continuous learning state objects are aggregated into a learning state change object of the corresponding type. The learning state change object aggregation unit aggregates multiple learning state change objects that have been verified to have causal or context-dependent relationships into a learning state change chain object that represents the learning state evolution path, based on the causal or context-dependent relationships.

10. A learning state data processing system as described in claim 9, characterized in that, The learning state change chain object includes several different types, and the learning state data processing system also includes: The dynamic interaction relationship analysis unit is used to perform time calibration and semantic alignment on different types of learning state change chain objects, extract multiple different types of learning state change chain objects that are in the same period or causally adjacent, and analyze the dynamic interaction relationship between multiple different types of learning state change chain objects. The evidence index unit is used to transform the multi-source heterogeneous raw data used for learning state inference into structured evidence units. Each structured evidence unit has a unique evidence anchor identifier. Based on the utilization relationship between the learning state inference result and the multi-source heterogeneous raw data, a mapping relationship between the structured evidence unit and the learning state inference result is constructed. Based on the correspondence between the learning state inference result and the learning state object, a mapping relationship between the learning state inference result and the learning state object is constructed.

Images