A learner high-order thinking multi-granularity joint representation method for an online collaborative exploration scene

Abstract

This invention proposes a multi-granular joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios, belonging to the field of online learning and intelligent education. This method integrates multi-source data such as learner profiles, behavioral sequences, and collaborative discourse to construct a joint embedding space. It then uses a pre-trained language model to extract shared semantic features, achieving a unified representation of higher-order thinking. Discrete cosine transform (DCT) is introduced for frequency domain analysis, and the compression ratio is adaptively determined by the power spectrum concentration, reducing the computational complexity to O(N log N). A progressive block-based strategy is employed to extract fine, medium, and coarse-grained semantic features, and multi-layer integration is achieved through upsampling and weighted fusion. Simultaneously, a frequency-domain constrained label relationship modeling mechanism is introduced to strengthen the semantic matching of features with higher-order cognitive dimensions such as analysis, evaluation, and creation, providing technical support for personalized learning diagnosis and precise intervention.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of online learning and intelligent education, and relates to a learner high-order thinking multi-granularity joint representation method for an online collaboration exploration scene based on frequency domain feature enhancement. BACKGROUND

[0002] Existing researches have proposed various high-order thinking representation methods based on behavior sequence mining, text analysis based on deep semantic understanding, and cognitive reasoning based on knowledge graph, but there are still obvious limitations: firstly, insufficient data integration: most methods focus on the analysis of a single data source and fail to fully integrate the multi-level associated information between learner portraits, behaviors and speeches; secondly, when modeling high-order thinking labels, the model usually stays at shallow feature matching and lacks in-depth modeling of label interdependence and semantic constraints, resulting in limited representation accuracy and discrimination; thirdly, some methods lack a multi-granularity feature fusion mechanism and cannot effectively unify surface behavior features and deep semantic features at the same time, affecting the accurate identification of the high-order cognitive state of learners; fourthly, the computational efficiency is low, and traditional methods based on Transformer have O(N²) time and space complexity, which is difficult to process long sequence learning behavior data. Therefore, there is an urgent need for a high-order thinking representation method that can comprehensively integrate multi-source heterogeneous data, fuse multi-granularity features and introduce label relationship constraints to improve the accuracy and reliability of the identification of the high-order thinking state of learners in the online collaboration exploration environment. The power spectrum of the hidden state of the Transformer will gradually concentrate in the low frequency area as the network layer deepens, which means that there is a lot of compressible redundant information in the hidden state. By performing discrete cosine transform (DCT) on the hidden state, the computational complexity can be significantly reduced while maintaining the performance of the model. SUMMARY

[0003] The technical problem this invention aims to solve is that, in existing online collaborative inquiry scenarios, learners' higher-order thinking states are difficult to accurately represent, failing to provide effective support for personalized learning interventions. To address this, this invention proposes the following technical solution: a multi-granularity joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios, based on frequency domain feature enhancement. Specifically, based on existing learning behavior analysis methods, a multi-granularity information fusion mechanism is introduced to construct a joint representation structure of higher-order thinking from three levels: learner profile, temporal behavior, and inquiry discourse. First, a profile information embedding model is constructed to extract learners' individual characteristics and background attributes and form personalized feature representations. Then, temporal behavior features are extracted based on behavior sequence modeling and attention mechanisms to obtain dynamic cognitive information at the learning process level. Third, learners' inquiry discourse is semantically encoded and key discourse features are extracted. Frequency domain feature analysis and adaptive compression are used to characterize their deep higher-order cognitive performance in knowledge construction and collaborative inquiry. Finally, through the multi-granularity information relationship enhancement module, the profile features, behavioral features and discourse features are deeply integrated, and a label constraint mechanism is introduced to strengthen the relationship between them and higher-order thinking labels, thereby outputting a highly interpretable and discriminative joint representation of higher-order thinking.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A multi-granularity joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios includes the following steps:

[0006] S1: Multi-source information representation stage: Preprocess and encode learner profile data, learning behavior sequence data and collaborative discourse text in online collaborative inquiry scenarios. Map learner profile information, behavior sequence information and discourse text information with higher-order thinking labels to the same vector space. Input the generated joint embedding sequence into the pre-trained language model, use its multi-layer Transformer structure to learn and output shared feature representations.

[0007] S2: Frequency domain feature analysis stage: Discrete cosine transform (DCT) is performed on the feature vectors of learner profiles, behavior sequences and corresponding utterances in the shared feature representation to obtain the frequency domain representation and analyze the power spectrum distribution. The cutoff ratio of each feature source is adaptively determined based on the power spectrum concentration.

[0008] S3: Adaptive multi-granularity compression stage: High-frequency components of the frequency domain representation are truncated according to the adaptive truncation ratio, and the compressed time domain features are recovered by inverse discrete cosine transform (IDCT).

[0009] S4: Progressive multi-granularity fusion stage: A block-based progressive feature extraction strategy is adopted to process the compressed features at multiple levels, and multi-granularity feature representation is obtained through upsampling and weighted fusion.

[0010] S5: Frequency domain constraint optimization stage: Introduce label relationship modeling with frequency domain constraints, learn features by constraining the predefined label relationship spectrum template, and optimize the model by combining multi-granularity classification loss;

[0011] S6: Higher-order thinking representation and output stage: Input multi-granularity feature representations into the corresponding classification modules, aggregate and fuse the output features of each module, and output the learner's representation results in the higher-order thinking dimensions of analysis, evaluation and creation through the classifier.

[0012] Furthermore, in S1, the learner profile field, learning behavior sequence field, and learner discourse text field are first cleaned and standardized. Then, the profile attributes, behavior events sorted by timestamp, and discourse text processed by word segmentation and sub-word segmentation are converted into corresponding tokens, and a unique tag token is generated for each higher-order thinking tag. Finally, the various tokens are embedded and concatenated according to preset segment identifiers and positions to form a joint input sequence, which is then input into a pre-trained language model. The shared feature representations of learner profile semantics, behavior temporal semantics, and discourse semantics are extracted through the model's multi-layer Transformer and multi-head attention mechanism.

[0013] Furthermore, the frequency domain feature analysis in S2 includes:

[0014] (1) Perform a one-dimensional discrete cosine transform on the image feature Fp to obtain the image frequency domain representation: ;

[0015] (2) Calculate the power spectrum distribution: , where k is the frequency index;

[0016] (3) The cutoff ratio is adaptively determined based on the power spectrum concentration, and its calculation method is shown in formula (1).

[0017] (1)

[0018] K90 represents the frequency threshold at which accumulated energy reaches 90%. and These are learnable parameters;

[0019] (4) Similarly, frequency domain analysis was performed on behavioral feature Fb and discourse feature Fd to obtain the truncation ratios rb and rd, respectively.

[0020] Furthermore, the adaptive multi-granularity compression in S3 includes:

[0021] (1) Truncate the high-frequency components and retain the previous ones. There are 10 frequency components, where N is the sequence length. The truncation ratio is i∈{p,b,d}, which represent the profile, behavior, and discourse features, respectively.

[0022] (2) Recover the compressed time-domain features through inverse discrete cosine transform: .

[0023] Furthermore, the progressive multi-granularity fusion in S4 employs a block-based strategy, including:

[0024] Shallow block Block1: Compression ratio r1=0.8, used to extract fine-grained semantic features;

[0025] Block 2 (middle layer): Compression ratio r2 = 0.5, used to extract medium-granularity semantic features;

[0026] Deep Block 3: Compression ratio r3=0.3, used to extract coarse-grained high-order semantic features;

[0027] The cross-layer feature fusion formula is as follows: Features at different levels are upsampled and weighted by learnable weights.

[0028]

[0029] in, This represents the multi-granularity fusion feature representation of the i-th feature source; Let J be the learnable fusion weight corresponding to the j-th granularity block; This is the multilayer perceptron corresponding to the j-th layer (shallow / medium / deep block); Upsample is the upsampling operation, which aligns features with different compression ratios to the same dimension; represents the time-domain features recovered through frequency domain compression; j is the granularity block index.

[0030] Furthermore, the frequency domain constraint optimization in S5 includes:

[0031] (1) Frequency domain constraint loss formula:

[0032] in Frequency domain constraint loss; The frequency domain representation of the label association features output by the model; The predefined label relationship spectrum template reflects the cognitive progression of "analysis → evaluation → creation";

[0033] (2) Cross-modal frequency domain consistency constraints are abandoned:

[0034] in, This represents the cross-modal frequency domain consistency constraint loss; This represents the fusion features of different modalities;

[0035] Frequency domain analysis is used to ensure the consistency of features across different modalities (such as learner profiles, behaviors, and discourse) in the frequency domain, thereby strengthening the association between multi-source information and higher-order thinking labels.

[0036] (3) Joint loss function:

[0037] in, , , These represent the classification loss based on learner profile information, learning behavior information, and learning discourse information, respectively. This represents the matching loss based on learner-label constraint relationships; , , , These are the weighting coefficients for the corresponding loss terms; + + =1.0, and the weights of each loss term are automatically adjusted through the validation set.

[0038] Furthermore, the discrete cosine transform is implemented using the Fast Fourier Transform (FFT) operator, with a computational complexity of O(N log N), where N is the sequence length; the truncation operation compresses the original N-dimensional features to... This effectively reduces the time and space complexity of subsequent Transformer layers.

[0039] Furthermore, the higher-order thinking labels include three dimensions: analysis, evaluation, and creation, corresponding to the higher-order cognitive levels in Bloom's taxonomy of cognitive objectives; the spectrum template is designed according to the laws of cognitive progression, with low-frequency components corresponding to long-term stable learning patterns and high-frequency components corresponding to short-term cognitive fluctuations.

[0040] The beneficial effects of this invention are that it can effectively reduce the computational complexity of the model to the O(N log N) level, while significantly improving the accuracy and discriminative power of higher-order thinking representations, providing effective technical support for personalized learning diagnosis and intervention in online collaborative inquiry environments. Attached Figure Description

[0041] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0042] Figure 1 This is a framework diagram of a multi-granular joint representation method for higher-order thinking of learners in online collaborative inquiry scenarios, as described in an embodiment of the present invention.

[0043] Figure 2This is a schematic diagram illustrating the overall process framework of a multi-granularity joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios, as described in an embodiment of the present invention. The diagram shows the entire process of learner multi-source data input, joint representation of learners and higher-order thinking labels, enhancement of multi-granularity information relationships, and output of higher-order thinking representation results. Detailed Implementation

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0045] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0046] To make the technical solution, advantages and objectives of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] As shown in the figure, this invention provides a multi-granularity joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios. The method specifically includes the following steps:

[0048] S1: Learner-Label Joint Embedding Stage. In order to obtain semantic representations that can characterize learners' higher-order thinking features and fully explore the fine-grained relationship between learner information and higher-order thinking labels, this step jointly embeds learners' multi-source information and higher-order thinking labels into a pre-trained language model for representation learning.

[0049] Using a learner-tag joint embedding approach, each tag is abstracted into a tag-unique mapping token. For any tag... Generate a unique corresponding one Learners are characterized from three dimensions: learner profile (p), learning behavior (b), and discourse (d). The profile reflects the learner's personal characteristics, the behavior reflects learning activities, and the discourse reflects cognitive level; these three interact to generate higher-order thinking. Profile information mainly includes learner background information, knowledge and abilities, cognitive level, learning style, and learning goals. Learning behavior mainly includes explicit behavioral data exported from system logs or collected directly through browser plugins. Discourse information mainly refers to interactive content such as discussions, Q&A, exchanges, and evaluations conducted by learners on the online platform.

[0050] For any training sample generate Using token embeddings as input to pre-trained language models allows BERT pre-trained models to learn fine-grained relationships between learners and labels during training.

[0051] S2: Shared Feature Representation Stage. To enable learner profiles, behaviors, utterances, and higher-order thinking labels to fully interact in a unified vector space and obtain shared feature representations containing multi-source semantic relationships, this step will jointly perform multi-layer feature modeling using the embedded sequence input pre-trained language model.

[0052] First, we plan to generate a token embedding by combining learner X and label Y through learner-label joint embedding, and sum this token embedding with the segment embedding to obtain the input of the BERT pre-trained language model. As shown in formula (2):

[0053] (2)

[0054] Then, in the BERT pre-trained language model, the first Transformer layer will be based on the initial input. Generate 3 weight matrices: query matrix Key matrix Sum matrix And using three matrices respectively in the input Multiplication yields the corresponding vector Q (0) K (0) V (0) As shown in formula (3), where:

[0055] Q (0) W q K (0) W k V (0) W v (3)

[0056] The output of the current Transformer layer is then calculated using the attention mechanism. And use it as the input to the next layer of Transformer. To participate in the new round of calculations. The process is shown in formulas (4) and (5):

[0057] (4)

[0058] (5)

[0059] in, , , Both are trainable parameter matrices, d k It is a matrix The dimension of the multiplication layer is used to prevent the result of the dot product from being too large, which would cause the gradient to be too small. Thanks to the 12-layer Transformer structure in BERT, each piece of information in the input fully interacts with the other information units. That is, the shallow relationships between learner profiles, behaviors, utterances and labels are all processed through the multi-head attention mechanism in the Transformer structure to obtain rich feature representations after interaction.

[0060] S3: Frequency Domain Feature Analysis and Adaptive Compression Stage

[0061] First, extract the feature vectors of learner profiles, behaviors, and discourse at corresponding locations: F p =[F p1 ,…,F pd ],Fb=[F b1 ,…,F bd ], F d =[F d1 ,…,F dd Next, a one-dimensional DCT is performed on the eigenvectors to obtain the frequency domain representation and analyze the power spectrum distribution, as shown in formulas (6) and (7).

[0062] (6)

[0063] (7)

[0064] The cutoff ratio is adaptively determined based on the power spectrum concentration, as shown in formula (8):

[0065] (8)

[0066] High-frequency components are truncated and recovered using IDCT, as shown in formulas (9) and (10):

[0067] (9)

[0068] (10)

[0069] S4: Progressive multi-granularity information fusion stage. For shallow blocks (Block1): r1≈0.8 is fine-grained semantic feature, for medium blocks (Block2): r2≈0.5 is medium-grained semantic feature, and for deep blocks (Block3): r3≈0.3 is coarse-grained. Finally, high-order semantic features are fused across layers, as shown in formula (11):

[0070] (11)

[0071] S5: Frequency domain constraint optimization and label relationship modeling stage: Introduce frequency domain constraint loss and cross-modal consistency constraint, and combine multi-granularity classification loss for model optimization.

[0072] S6: Multi-granularity information extraction and higher-order thinking representation stage. In order to characterize learners' higher-order cognitive features from different information granularities and form a distinctive joint representation of higher-order thinking, this step extracts profile, behavioral, and discourse features separately based on shared feature representations.

[0073] Based on shared feature representations, feature vectors for learner profiles, behaviors, and utterances are extracted at corresponding locations: F p =[F p1 ,…,F pd ],Fb=[F b1 ,…,F bd ], F d =[F d1 ,…,F dd The above feature vectors are aggregated and processed separately, and then multi-granularity feature representations are generated using a multi-layer perceptron (MLP): Here, agg(∙) refers to the aggregation function, including summation aggregation, join aggregation, max pooling aggregation, or LSTM (Long Short-Term Memory Neural Network), etc. Similarly, the aim is to obtain a representation of learner behavior using the above methods. and the representation of learners' discourse .

[0074] S7: Higher-order thinking prediction and loss optimization stage. In order to supervise and constrain the multi-granularity joint representation of learners' higher-order thinking and optimize it collaboratively, and output prediction results that can represent learners' higher-order thinking state, this step predicts the learners' higher-order thinking type based on the profile feature representation, behavioral feature representation and discourse feature representation obtained in step S6, and completes model training and inference output.

[0075] Specifically, the multi-granularity feature representations obtained in step S6 are input into the corresponding classification modules to perform multi-label prediction of learners' higher-order thinking in online collaborative inquiry scenarios. The binary cross-entropy loss function is used as the optimization objective in each classification module, and its calculation method is shown in formula (12):

[0076] (12)

[0077] Based on this, the loss terms corresponding to the portrait features, behavioral features, discourse features, and label constraint relationships are weighted and fused to obtain the overall loss function, as shown in formula (13):

[0078] (13)

[0079] in, , , These represent the classification loss based on learner profile information, learning behavior information, and learning discourse information, respectively. This represents the matching loss based on learner-label constraint relationships; , , , These are the weighting coefficients for the corresponding loss terms, used to balance the contributions of features of different granularities in the overall optimization process.

[0080] By jointly optimizing the model parameters using the aforementioned loss function, the model gradually learns a stable mapping relationship between learners' multi-source information and higher-order thinking labels during the training phase. During the model inference phase, it outputs the learners' higher-order thinking category prediction results and their corresponding multi-granularity joint representation vectors, which are used to represent the learners' higher-order thinking states in online collaborative inquiry scenarios.

[0081] Example

[0082] The high-order thinking multi-granularity joint representation method in this embodiment consists of a learner-label joint embedding module, a shared feature representation module, a multi-granularity information extraction module, and a joint optimization module. First, to comprehensively represent the individual characteristics, behavioral process characteristics, and inquiry discourse characteristics of learners in the online collaborative inquiry process, and to strengthen the association between these characteristics and high-order thinking labels, this embodiment performs joint embedding modeling of learner profiles, temporal behaviors, and collaborative discourse with high-order thinking labels, thereby providing a unified semantic representation foundation for high-order thinking representation. Next, to achieve the interaction and fusion of multi-source information in the same feature space, this embodiment inputs the aforementioned joint input sequence into a pre-trained language model. Through a multi-layer Transformer structure and attention mechanism, it learns and obtains shared feature representations containing the interaction information between learner profiles, behaviors, discourse, and high-order thinking labels. Then, to extract key information of different granularities from the shared features, this embodiment performs aggregation mapping on the vectors corresponding to the learner profile position, behavior position, and discourse position in the shared feature representations, constructing learner profile layer feature representations, behavior layer feature representations, and discourse layer feature representations, forming a multi-granularity feature foundation. Based on this, this embodiment introduces a discrete cosine transform (DCT) frequency domain analysis mechanism to perform frequency domain transformation on profile features, behavioral features, and discourse features respectively. By calculating the power spectrum distribution and adaptively determining the truncation ratio of each feature source based on the power spectrum concentration, the high-frequency redundant components are accurately removed. Then, the compressed time domain features are recovered by inverse discrete cosine transform (IDCT), which optimizes the model computational complexity to the O(N log N) level while retaining the core semantic information.

[0083] To further explore the deep correlations among multi-granular features, this embodiment adopts a block-based progressive fusion strategy. It extracts fine-grained semantic features from shallow blocks, medium-granular features from mid-level blocks, and coarse-grained high-order semantic features from deep blocks. Then, it uses upsampling and learnable weights for weighted fusion to generate multi-granular fused features that consider both surface details and deep logic. Simultaneously, it introduces frequency-domain constrained label relationship modeling. Based on the cognitive progression of "analysis → evaluation → creation," it designs a preset spectral template. Frequency-domain constraint loss and cross-modal consistency constraints strengthen the semantic matching of features and labels. A combined multi-granularity classification loss from profiles, behaviors, and discourse is used to construct an overall optimization objective, achieving synergistic optimization of model parameters. Finally, through feature aggregation and classifier inference, it outputs accurate representations of learners in higher-order thinking dimensions such as analysis, evaluation, and creation, providing technical support with both discriminative power and interpretability for personalized learning diagnosis and intervention in online collaborative inquiry scenarios. Finally, by introducing a multi-module collaborative optimization mechanism, comprehensively considering various multi-granular features and conducting supervised training on the corresponding classification modules, this embodiment outputs learner higher-order thinking joint representation results with high discriminativeness and strong interpretability, providing effective support for the diagnosis and intervention of higher-order thinking in online collaborative inquiry scenarios.

[0084] Step 1: Preprocessing of Multi-Source Learner Data and Construction of a Higher-Order Thinking Tagging System. Learner profile information, learning behavior logs, and discourse data collected from the online collaborative inquiry platform undergo unified preprocessing, including outlier removal, missing value imputation, and text segmentation. Profile information includes learners' background attributes, knowledge level, learning goals, and preferences; behavior data includes operation sequences, resource access records, and collaborative interaction details; discourse data includes text communication content from discussion and collaboration areas. After cleaning, a higher-order thinking tagging system is constructed based on educational expert annotations and cognitive theories. Higher-order thinking is categorized into analysis, evaluation, and creation, and encoded to provide standardized monitoring signals for subsequent representation.

[0085] Step Two: Learner-Label Joint Embedding Modeling. To strengthen the association between learner multi-source information and higher-order thinking labels at the representation level, learner profiles, learning behaviors, and inquiry discourses are encoded into corresponding token sequences, and higher-order thinking labels are abstracted into label tokens with unique mapping relationships. By concatenating learner multi-source tokens and label tokens in a predetermined order to form a joint input sequence, this sequence is fed into a pre-trained language model. During training, the model simultaneously learns the internal structure of learner features and the fine-grained semantic relationships between learner features and higher-order thinking labels, thereby obtaining a preliminary joint embedding representation.

[0086] Step 3: Shared Feature Representation and Cross-Modal Interaction Modeling. To achieve full interaction between profile features, behavioral features, discourse features, and tag information in the same semantic space, the jointly embedded token sequence is input into a multi-layer Transformer structure of the pre-trained language model. Through a multi-head attention mechanism, the interdependencies between different information units are modeled, enabling shared feature representations to be formed between learner profiles, behaviors, discourse, and higher-order thinking tags. Through layer-by-layer mapping of the multi-layer Transformer, shared feature vectors containing multi-source information interaction relationships are obtained. Simultaneously, through frequency domain feature analysis and adaptive compression, DCT transformation is performed on the shared feature representation to calculate the power spectrum distribution. The truncation ratio is adaptively determined based on the power spectrum concentration, truncating high-frequency components and recovering the compressed temporal features through IDCT. Next, a block-based progressive feature extraction strategy is adopted, upsampling features at different levels and fusing them through learnable weights. Finally, by introducing frequency domain constraint loss and cross-modal consistency constraint, combined with multi-granularity classification loss, the model is optimized, providing a unified feature foundation for multi-granularity information extraction.

[0087] Step Four: Multi-granularity Feature Extraction and Construction of Higher-Order Thinking Representations. Based on shared feature representations, feature vectors are extracted from the learner's profile, behavior, and discourse locations. Through feature aggregation operations and a multilayer perceptron network, information at different granularities is mapped to form profile-level, behavior-level, and discourse-level feature representations. Furthermore, an attention mechanism is introduced to weightedly fuse these multi-granular features, constructing a joint higher-order thinking representation that simultaneously reflects the learner's individual characteristics, process behavior characteristics, and deep cognitive characteristics, thereby enhancing the discriminative power and interpretability of higher-order thinking at the feature level.

[0088] Step 5: Joint Optimization and Higher-Order Thinking Output. Classifiers are built and trained for different granularities of features. Cross-entropy loss is used as the optimization objective for classifiers of learner profiles, behavior, and discourse. The overall loss function is constructed by weighted summation of each loss term. The model parameters are jointly optimized using the backpropagation algorithm, enhancing the matching degree between the model and higher-order thinking labels while maintaining consistency across multi-source features. Finally, the representation results and corresponding predicted categories of learners in higher-order thinking dimensions such as analysis, evaluation, and creation are output, providing a basis for learning diagnosis and personalized support in online collaborative inquiry environments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-granularity joint representation method for learners' higher-order thinking in online collaborative inquiry scenarios, characterized in that, Includes the following steps: S1: Multi-source information representation stage: Preprocess and encode learner profile data, learning behavior sequence data and collaborative discourse text in online collaborative inquiry scenarios. Map learner profile information, behavior sequence information and discourse text information with higher-order thinking labels to the same vector space. Input the generated joint embedding sequence into the pre-trained language model, use its multi-layer Transformer structure to learn and output shared feature representations. S2: Frequency domain feature analysis stage: Discrete cosine transform (DCT) is performed on the feature vectors of learner profiles, behavior sequences and corresponding utterances in the shared feature representation to obtain the frequency domain representation and analyze the power spectrum distribution. The cutoff ratio of each feature source is adaptively determined based on the power spectrum concentration. S3: Adaptive multi-granularity compression stage: High-frequency components of the frequency domain representation are truncated according to the adaptive truncation ratio, and the compressed time domain features are recovered by inverse discrete cosine transform (IDCT). S4: Progressive multi-granularity fusion stage: A block-based progressive feature extraction strategy is adopted to process the compressed features at multiple levels, and multi-granularity feature representation is obtained through upsampling and weighted fusion. S5: Frequency domain constraint optimization stage: Introduce label relationship modeling with frequency domain constraints, learn features by constraining the predefined label relationship spectrum template, and optimize the model by combining multi-granularity classification loss; S6: Higher-order thinking representation and output stage: Input multi-granularity feature representations into the corresponding classification modules, aggregate and fuse the output features of each module, and output the learner's representation results in the higher-order thinking dimensions of analysis, evaluation and creation through the classifier.

2. According to the method described in claim 1, in step S1, the learner profile field, the learning behavior sequence field, and the learner discourse text field are first cleaned and standardized; then, the profile attributes, the behavior events sorted by timestamp, and the discourse text after word segmentation and sub-word segmentation are converted into corresponding tokens, and a unique tag token is generated for each higher-order thinking tag; finally, the various tokens are embedded and concatenated according to preset segment identifiers and positions to form a joint input sequence, and input into a pre-trained language model, through the model's multi-layer Transformer and multi-head attention mechanism to extract the shared feature representations of learner profile semantics, behavior temporal semantics, and discourse semantics.

3. The method according to claim 1, characterized in that, The frequency domain feature analysis in S2 includes: (1) Perform a one-dimensional discrete cosine transform on the image feature Fp to obtain the image frequency domain representation: ; (2) Calculate the power spectrum distribution: , where k is the frequency index; (3) The cutoff ratio is adaptively determined based on the power spectrum concentration, and its calculation method is shown in formula (1). （1） K90 represents the frequency threshold at which accumulated energy reaches 90%. and These are learnable parameters; (4) Similarly, frequency domain analysis was performed on behavioral feature Fb and discourse feature Fd to obtain the truncation ratios rb and rd, respectively.

4. The method according to claim 3, characterized in that, The adaptive multi-granularity compression in S3 includes: (1) Truncate the high-frequency components and retain the previous ones. There are 10 frequency components, where N is the sequence length. The truncation ratio is i∈{p,b,d}, which represent the profile, behavior, and discourse features, respectively. (2) Recover the compressed time-domain features through inverse discrete cosine transform: .

5. The method according to claim 4, characterized in that, The progressive multi-granularity fusion in S4 employs a block-based strategy, including: Shallow block Block1: Compression ratio r1=0.8, used to extract fine-grained semantic features; Block 2 (middle layer): Compression ratio r2 = 0.5, used to extract medium-granularity semantic features; Deep Block 3: Compression ratio r3=0.3, used to extract coarse-grained high-order semantic features; The cross-layer feature fusion formula is as follows: Features at different levels are upsampled and weighted by learnable weights. in, This represents the multi-granularity fusion feature representation of the i-th feature source; Let J be the learnable fusion weight corresponding to the j-th granularity block; Let j be the multilayer perceptron corresponding to the j-th layer; Upsample is the upsampling operation, which aligns features with different compression ratios to the same dimension; represents the time-domain features recovered through frequency domain compression; j is the granularity block index.

6. The method according to claim 5, characterized in that, The frequency domain in S5 is approximately Beam optimization includes: (1) Frequency domain constraint loss formula: in Frequency domain constraint loss; The frequency domain representation of the label association features output by the model; The predefined label relationship spectrum template reflects the cognitive progression of "analysis → evaluation → creation"; (2) Cross-modal frequency domain consistency constraints are abandoned: in, This represents the cross-modal frequency domain consistency constraint loss; This provides a fusion feature representation for different modalities; frequency domain analysis is used to ensure the consistency of features across different modalities in the frequency domain, thereby strengthening the association between multi-source information and higher-order thinking labels. (3) Joint loss function: in, , , These represent the classification loss based on learner profile information, learning behavior information, and learning discourse information, respectively. This represents the matching loss based on learner-label constraint relationships; , , , These are the weighting coefficients for the corresponding loss terms; + + =1.0, and the weights of each loss term are automatically adjusted through the validation set.

7. The method according to claim 1, characterized in that, The discrete cosine transform is implemented using the Fast Fourier Transform (FFT) operator, with a computational complexity of O(N log N), where N is the sequence length; the truncation operation compresses the original N-dimensional features to... This effectively reduces the time and space complexity of subsequent Transformer layers.

8. The method according to claim 1, characterized in that, The higher-order thinking labels include three dimensions: analysis, evaluation, and creation, corresponding to the higher-order cognitive levels in Bloom's taxonomy of cognitive objectives; the spectrum template is designed according to the laws of cognitive progression, with low-frequency components corresponding to long-term stable learning patterns and high-frequency components corresponding to short-term cognitive fluctuations.

Images