Bloom cognitive level constraint-based achievement-oriented education diagnosis method and system

Abstract

The application discloses a Bloom cognitive hierarchy constraint-based achievement-oriented education diagnosis method and system. The method takes course outline text and student evaluation data as input, and realizes the automatic diagnosis of course goal achievement through three structural modifications in the large language model Transformer architecture: the logarithmic value of the cognitive hierarchy transfer matrix is embedded as a learnable bias item in the attention score calculation path to realize feature coding of cognitive hierarchy perception; the continuous differentiable relaxation technique is used to establish an end-to-end differentiable inference path for the course goal achievement weight; and the Bloom partial order constraint is coded as a training loss, so that the diagnosis result meets the cognitive hierarchy progressive relationship. The three modifications are optimized by a joint loss function, which significantly improves the diagnosis accuracy and promotes the interpretability and scientificity of the large language model diagnosis in the achievement-oriented education scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and educational informatization, specifically to an outcome-oriented educational diagnostic method and system based on Bloom's hierarchy of cognitive constraints. Background Technology

[0002] Outcome-based education is a core concept in higher engineering education accreditation, requiring quantitative evaluation of the achievement of each course objective and driving continuous improvement based on the evaluation results. As the application of large language models in education expands, researchers are beginning to explore their introduction into outcome-based education evaluation scenarios to automate diagnostic report generation. However, existing solutions have limitations in the following three aspects.

[0003] At the attention computation level, the standard Transformer's self-attention mechanism uses a uniform similarity measure for each token in the sequence, failing to identify the differences between memory-level and creativity-level questions within Bloom's cognitive framework, and also unable to dynamically model the matching degree between the cognitive level of the questions and the level of the curriculum objectives. Existing research introduces a cognitive level matching coefficient. ,in, For the cognitive level of the topic, At the course objective level, It is a natural exponential function. It is a fixed attenuation hyperparameter set manually, but this coefficient exists in the data preprocessing stage, does not participate in the gradient calculation of the neural network, cannot be adaptively optimized through data-driven methods, and cannot be jointly optimized with the language model generation target.

[0004] At the semantic alignment level, existing methods identify the action verb hierarchy and knowledge domain coverage in course objective description sentences through a predefined dictionary, and output weight anchors in the form of discrete scalars. The verb identification and knowledge domain extraction processes are non-differentiable, and their outputs are passed to the achievement calculation module as discrete symbols. There is no gradient connection between weight inference and achievement calculation, making it impossible to back-optimize the semantic alignment process through achievement prediction errors. Furthermore, the fusion weights for verb hierarchy compatibility and knowledge domain coverage use fixed values, which cannot adaptively adjust the relative importance of the two in different course scenarios.

[0005] At the constraint guarantee level, existing methods use the 15 partial order constraint pairs from Bloom's six cognitive levels as data post-processing rules for constraint violation detection and correction after the language model generates a report. The detection signals of constraint violations cannot affect model parameters through backpropagation, and the model is unaware of partial order constraints during training. This separate architecture cannot provide architecture-level constraint guarantees during the inference phase, and the system's constraint satisfaction capability cannot continuously improve with the accumulation of training data.

[0006] Existing solutions combining Bloom's hierarchy of knowledge with deep learning techniques can be summarized into four categories. The first category uses Bloom's hierarchy as the output label for text classification, training a classifier to classify learned texts at cognitive levels. However, the classification model does not contain structured information about the cognitive hierarchy and cannot model the progressive relationships and transition characteristics between levels. The second category constructs Bloom's hierarchy as an external feature tensor as additional features for the model input. The hierarchical information only interacts with the model at the input end and does not participate in the core path of attention calculation within the model. The third category injects Bloom's hierarchy information into the input of a large language model in text form through prompt word templates. There is no gradient connection between the hierarchical information and the model parameters. The fourth category performs general architectural optimizations on the large language model, such as introducing a state-space model or a hybrid expert network to improve model capacity or computational efficiency. However, this type of optimization does not target the attention calculation structure itself, does not change the way attention scores are calculated, cannot introduce cognitive hierarchy bias terms into the attention scores, and cannot encode Bloom's partial order relations as backpropagable training constraints.

[0007] The common limitation of the above four types of schemes is that the structured knowledge of Bloom's cognitive level is always outside the attention calculation architecture of the model. It cannot directly participate in the calculation process of attention score in the form of learnable parameters and differentiable constraints. It cannot internalize the transfer characteristics of cognitive level, verb hierarchical mapping relationship and partial order logic constraint into learnable parameters of the model through the training process. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an outcome-oriented educational diagnostic method and system based on Bloom's hierarchy of cognitive constraints, aiming to solve the problems in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints, comprising the following steps: Step S1: Obtain the course outline text, course objective description, and student assessment response data of the course to be diagnosed; perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; construct a diagnostic input word sequence based on the course outline text, the annotated course objective description, and the student assessment response data. Step S2: Construct an improved large language model, which consists of an embedding layer, an N-layer modified Transformer layer, a Bloom partial order constraint gradient propagation layer, and an output layer connected in sequence; wherein, the modified Transformer layer includes: replacing the original multi-head self-attention module in the original Transformer layer with a cognitive level sensitive weighted attention module, and connecting the two-dimensional semantic alignment sub-network in parallel with the feedforward network path in the Transformer layer in the form of a bypass; Step S3: Input the diagnostic input word sequence into the embedding layer of the improved large language model for processing to obtain the diagnostic input embedding matrix; Step S4: The diagnostic input embedding matrix is ​​sequentially processed through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the attention output of the cognitive level perception. After residual and layer normalization, an intermediate feature representation is obtained. The intermediate feature representation is then fed into the feedforward network and the two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final attention weights of the target sensitivity of the corresponding layer. The alignment weights of the cognitive-level sensitive weighted attention module of the first Transformer layer are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers. Step S5: Input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

[0010] Furthermore, the construction of the diagnostic input embedding matrix in step S3 includes: calculating the standard word embedding, rotation position encoding, and Bloom hierarchical embedding obtained from the learnable Bloom six-level cognitive hierarchy embedding matrix for each word in the diagnostic input word sequence, and adding the three elements one by one to obtain the final input embedding vector.

[0011] Furthermore, the processing procedure of the cognitive level sensitive weighted attention module includes: Define the transition matrix of Bloom's cognitive levels; The inputs of the cognitive-level sensitive weighted attention module are projected into query vectors, key vectors, and value vectors, respectively. Based on the Bloom cognitive hierarchy labels of the lexical units, the query vector and key vector are subjected to element-wise multiplication modulation of cognitive hierarchy embedding to obtain cognitive hierarchy perceived query vector and cognitive hierarchy perceived key vector. Attention scores are calculated based on the cognitive level-aware query vector and cognitive level-aware key vector, and the logarithmic bias term obtained from the cognitive level transition matrix is ​​embedded into the addition path of the attention scores; the Bloom cognitive level corresponding to the course objectives is introduced to calculate the final attention weights that are sensitive to the objectives; the final attention weights are used to weight and sum the value vectors to obtain the output.

[0012] Furthermore, the logarithmic bias term is obtained by taking the logarithm of the transition weights in the cognitive level transition matrix, which contains learnable scalar parameters and learnable bias terms, and satisfies asymmetric constraints.

[0013] Furthermore, the processing procedure of the two-dimensional semantic alignment sub-network includes: Construct a learnable embedding dictionary and hierarchical prototype vectors for Bloom's action verbs; Calculate the Euclidean distance between the course objective action verbs and the prototype vectors at each level, and obtain the hierarchical distribution vectors of the standardized course objective description action verbs through a Gumbel-Softmax relaxation combined with a temperature annealing strategy. , is represented as: ; In the formula, Indicates the first The Bloom's six-level cognitive probability distribution vector corresponding to each verb, that is, the hierarchical distribution vector of behavioral verbs describing standardized curriculum objectives; For temperature parameters; Action verbs in the target description With the Euclidean distance between layer prototype vectors; This represents the Gumbel-Softmax relaxation method; Based on intermediate feature representations, an initial feature vector of the knowledge domain is extracted. Then, the knowledge domain feature vectors corresponding to the course objectives and the knowledge domain feature vectors corresponding to student assessment responses are obtained through graph attention network encoding. The hierarchical distribution vector of action verbs in combination with standardized curriculum objectives And the Bloom's hierarchy of cognition labels corresponding to the lexical units, calculate the alignment score for the verb dimension; Based on the knowledge domain feature vector corresponding to the course objectives The knowledge domain feature vector corresponding to student assessment answers Calculate the similarity score along the knowledge domain dimension; By using a learnable two-dimensional fusion coefficient, the alignment scores of the verb dimension and the knowledge domain dimension are fused to obtain a comprehensive alignment score; The comprehensive alignment score is passed to the final attention weight calculation step of the cognitive level sensitive weighted attention module of the next layer of the Transformer layer, and participates in the weight normalization operation together with the Bloom cognitive level corresponding to the course objectives.

[0014] Furthermore, the processing procedure of the Bloom partial order constrained gradient propagation layer includes: Construct a fixed constraint matrix based on the partial order relation of Bloom's six-level cognitive hierarchy; The final hidden state is mapped to a probability distribution of Bloom's six cognitive levels by using a prediction head; For each pair satisfied The cognitive level affects the computation of differentiable constraint violations: ; In the formula, Indicates the first The and the first The differentiable constraint violation quantity of each cognitive level; This indicates a modified linear unit activation function; , They represent the first , The probability of achieving a higher level of cognition; Represents a fixed constraint matrix The Middle Line number Column elements; when hour, This indicates a low level of cognition. The probability of achieving this is lower than at higher cognitive levels. This violates the Bloom partial order constraint; when the constraint is satisfied, , No punishment; Constructing a composite partial order constraint penalty loss function by combining linear and superlinear penalties : ; In the formula, The adjustable weight hyperparameters represent the superlinear penalty term; Indicates batch; Indicates an indicator function; The gradient of the composite partial order constraint penalty loss function is backpropagated to the model's full-link learnable parameters; the model's full-link learnable parameters include the prediction head parameters, the weights of all modified Transformer layers, the learnable parameters of the cognitive level sensitive weighted attention module, the hierarchical prototype vector of the two-dimensional semantic alignment subnetwork, and the learnable two-dimensional fusion coefficients.

[0015] Furthermore, the attention score is calculated based on the cognitive-level perceived query vector and the cognitive-level perceived key vector, and is expressed as: ; In the formula, The cognitive hierarchy perceptual attention score matrix represents the first... line, number Column elements; Indicates the first The cognitive hierarchy of query vectors for each word element; Indicates the first The cognitive level perception key vector of each word element; Indicates transpose; This represents the feature dimension of the key vector; Indicates from the cognitive level Towards Jump transition weights Indicates the first Each word corresponds to a Bloom cognitive tag. Indicates the first Bloom's cognitive tags corresponding to each word element; express The logarithm of .

[0016] An outcome-oriented educational diagnostic system based on Bloom's hierarchy of cognitive constraints includes: The diagnostic input word sequence construction module is used to obtain the course outline text, course objective description, and student assessment response data of the course to be diagnosed, and to perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; based on the course outline text, the annotated course objective description, and the student assessment response data, the diagnostic input word sequence is constructed. The model building module is used to build an improved large language model. The improved large language model consists of an embedding layer, an N-layer modified Transformer layer, a Bloom partial order constraint gradient propagation layer, and an output layer connected in sequence. The modified Transformer layer includes replacing the original multi-head self-attention module in the original Transformer layer with a cognitive level sensitive weighted attention module, and connecting the two-dimensional semantic alignment sub-network in parallel with the feedforward network path in the Transformer layer in the form of a bypass. The diagnostic input lexical sequence processing module is used to process the diagnostic input lexical sequence into the embedding layer of the improved large language model to obtain the diagnostic input embedding matrix. The diagnostic input embedding matrix processing module processes the diagnostic input embedding matrix through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the cognitive-level perceptual attention output. After residual and layer normalization, an intermediate feature representation is obtained. This intermediate feature representation is then fed in parallel into a feedforward network and a two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final attention weights for the target sensitivity of the corresponding layer. The alignment weights of the cognitive-level sensitive weighted attention module in the first Transformer layer are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers. The output module is used to input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

[0017] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints.

[0018] A non-volatile computer storage medium storing computer-executable instructions that execute an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints.

[0019] Compared with existing technologies, the present invention has the following advantages:

[0020] First, this invention directly embeds the logarithm of the Bloom cognitive hierarchy transition matrix into the addition path of the attention score matrix, placing it in the same computational graph as the query-key inner product. This makes cognitive hierarchy information an organic component of the attention computational structure, rather than an external auxiliary signal. The transition matrix parameters and learnable bias terms are directly updated through backpropagation of the attention mechanism, overcoming the shortcomings of existing technologies where the cognitive hierarchy matching coefficients are fixed in the preprocessing stage and cannot be adaptively optimized. Simultaneously, this invention performs element-wise multiplication modulation of the query vector and key vector using cognitive hierarchy embedding, forming a composite cognitive hierarchy perceptual structure together with the bias terms. This enables the model to dynamically distinguish the attention weights of memory-level and creation-level questions, significantly improving the cognitive hierarchy sensitivity of the diagnostic report generation.

[0021] Second, this invention transforms the Bloom hierarchy recognition of action verbs from a discrete lookup table operation into a computation based on continuous differentiable relaxation. Combined with a temperature annealing strategy and six learnable hierarchical prototype vectors as mapping centers, the gradient of the achievement prediction error can flow backward through the verb hierarchy recognition steps, continuously optimizing the position of the prototype vectors in the semantic space. The knowledge domain dimension is encoded using a multi-layer graph attention network to achieve adaptive aggregation of knowledge point association information. More importantly, this invention replaces the fixed constants in existing technologies with learnable fusion coefficients, automatically adjusting the relative importance of the verb dimension and the knowledge domain dimension through training, enabling the alignment score to adapt to the differentiated weight requirements of different course scenarios. The above design establishes a fully end-to-end differentiable computational path from the original course outline to the achievement weight inference, overcoming the fundamental problem of gradient breakage between verb recognition and achievement calculation in existing technologies.

[0022] Third, this invention encodes the fifteen partial order constraints of Bloom's six-level cognitive hierarchy into differentiable constraint violation quantities and constructs a composite partial order constraint penalty loss, which combines linear and superlinear penalties. The constraint violation gradient is backpropagated to all model parameters via a chain rule, including the transition matrix of the attention layer, the fusion coefficients of the aligned sub-networks, and the hierarchical prototype vector. This invention further introduces a constraint-aware diagnostic generation gating system, triggering fine-grained diagnostic generation only when the gating value exceeds a threshold, avoiding unnecessary computational overhead. This scheme enables the model to learn and generate diagnostic outputs that satisfy partial order constraints during the training phase, and the constraint satisfaction capability can continuously improve with the accumulation of training data, unlike existing technologies that treat partial order constraints as ex-post verification rules during the inference phase.

[0023] Fourth, this invention achieves physical coupling of gradients by enabling the three technological innovations to share the hidden layer parameter space of the model through a joint loss function. Cognitive-level sensitive attention provides a cognitively discriminative feature basis for semantic alignment, while the partial order constraint gradient acts simultaneously on the transition matrix and fusion coefficients. These three elements form mutually reinforcing gradient coupling during training through shared parameters. Experiments show that the comprehensive diagnostic performance surpasses the simple superposition of the individual technologies, producing a superlinear synergistic effect. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of the improved large language model (ChatGLM4OBE) of this invention.

[0025] Figure 2 This is a schematic diagram of the cognitive hierarchy sensitive weighted attention module (CLSWA) of the present invention.

[0026] Figure 3 This is a schematic diagram of the two-dimensional semantic alignment subnetwork (VKDA) of the present invention.

[0027] Figure 4 This is a schematic diagram of the Bloom Partial Order Constrained Gradient Propagation Layer (BPOG) of the present invention. Detailed Implementation

[0028] Example 1

[0029] This embodiment provides an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints, including the following steps: Step S1: Obtain the course syllabus text, course objective description, and student assessment response data of the course to be diagnosed; perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; construct a diagnostic input word sequence based on the course syllabus text, the annotated course objective description, and the student assessment response data.

[0030] Step S1.1: Obtain the course syllabus text, course objective description, and student assessment response data of the course to be diagnosed; perform standardized cleaning and unified lexicalization on all texts of the course syllabus text, course objective description, and student assessment response data to convert natural language into a sequence of lexical IDs that the model can recognize.

[0031] Step S1.2: Based on Bloom's six levels of cognitive hierarchy (L1 memory, L2 comprehension, L3 application, L4 analysis, L5 evaluation, L6 creation), Bloom's cognitive hierarchy is used to annotate the action verbs and knowledge domains in the course objective descriptions, deriving the cognitive level of each course objective; the annotation results are mapped to corresponding word units to generate a Bloom's cognitive hierarchy label sequence corresponding to the original word unit sequence. : ; In the formula, express The Bloom's hierarchy of cognitive labels for the last word in the text; Indicates the total number of lexical elements; express The Middle Bloom's hierarchy of cognitive labels for each word element This indicates the position number of a word in the word sequence. 0 is the neutral label for all words in the general word sequence and student assessment response data. 1~6 correspond to L1 memory, L2 comprehension, L3 application, L4 analysis, L5 evaluation, and L6 creation, respectively.

[0032] Step S1.3: Following the logic of overall course requirements → specific teaching objectives → student learning performance, concatenate the word sequence of the course syllabus text, the annotated course objective descriptions, and student assessment response data, while simultaneously standardizing the sequence length to generate a unified format of original diagnostic input word sequence. : ; In the formula, This represents the sequence of terms in the course syllabus text. This indicates the total number of lexical units in the course syllabus. This indicates the last word in the sequence of words in the course outline text; This represents the sequence of lexical terms describing all course objectives. This indicates the total number of terms in the description of the course objectives. This indicates the last lexical element in the course objective lexical subsequence; This represents the sequence of words used by students to evaluate their responses. This indicates the total number of lexical units in the student's assessment response. This indicates the last word in the sequence of words used to evaluate a student's response.

[0033] Step S2: Construct the improved large language model (ChatGLM4OBE), such as Figure 1 As shown, the improved large language model consists of an embedding layer, N modified Transformer layers, a Bloom partial order constraint gradient propagation layer (BPOG), and an output layer connected in sequence. The process of obtaining and modifying the Transformer layer is as follows: A cognitive-level sensitive weighted attention module (CLSWA) is constructed, which replaces the original multi-head self-attention module in the original Transformer layer with the cognitive-level sensitive weighted attention module. Construct a two-dimensional semantic alignment subnetwork (VKDA) and connect it in parallel with the feedforward network path within the Transformer layer as a bypass.

[0034] Step S3: Input the diagnostic input word sequence into the embedding layer of the improved large language model for processing to obtain the diagnostic input embedding matrix.

[0035] Step S3.1: Define a learnable Bloom's six-level cognitive hierarchy embedding matrix Each row of the matrix corresponds to a learnable vector representation of Bloom's cognitive hierarchy; the original diagnostic input word sequence The Middle each word element Calculate its standard lexical embedding , This represents the embedding of basic lexical elements into the query matrix; based on lexical elements... Input word sequence in the original diagnosis Calculate the corresponding Rotational Position Code (RoPE) based on the position in the matrix. According to the word elements Location corresponding to Bloom's hierarchy of knowledge labels Bloom's hierarchical embeddings are calculated by using a lookup table index within the learnable Bloom's six-level cognitive embedding matrix. ; Step S3.2: Add the standard word embedding, rotation position encoding, and Bloom hierarchy embedding element-wise for each word to obtain the final input embedding vector for each word: ; In the formula, Indicates the first The final input embedding vector of each word.

[0036] Step S3.3: Concatenate the final input embedding vectors of all words to generate the diagnostic input embedding matrix. : ; In the formula, This represents the final input embedding vector of the last word; Represents the set of all real numbers; This represents a unified embedding feature dimension.

[0037] Step S4: The diagnostic input embedding matrix is ​​sequentially processed through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the cognitive-level perceptual attention output. After residual and layer normalization, an intermediate feature representation is obtained. The intermediate feature representation is then fed into the feedforward network and the two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final attention weights for target sensitivity in that layer. The alignment weights of the cognitive-level sensitive weighted attention module of the first Transformer layer are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers.

[0038] like Figure 2 As shown in the figure, the left side of the figure shows the diagnostic input word sequence and its corresponding Bloom level label input. The middle part shows the interaction between Bloom embedding and query projection and key projection to generate cognitive level perceptual query vector and cognitive level perceptual key vector. The logarithm of the transition matrix is ​​embedded in the attention score calculation path in the form of a differentiable bias term. The output comparison on the right side of the figure shows the difference in attention distribution between standard attention (ChatGLM-small baseline) and cognitive level sensitive attention (the model in this paper) after Bloom perceptual reweighting, intuitively demonstrating the modulation effect of the cognitive level bias term on attention allocation. The processing of the cognitive level sensitive weighted attention module in the modified Transformer layer includes: 1. Define the transition matrix of Bloom's cognitive levels. (Used to quantify the degree of preference for pairwise transitions between Bloom's six cognitive levels), the formula for calculating the elements of Bloom's cognitive level transition matrix is ​​as follows: ; In the formula, express The Middle line, number The elements of the column represent the elements starting from the first... The cognitive level shifts to the first... Preference weights at each cognitive level; Bloom's hierarchy of cognition labels representing two lexical units respectively; Represents an exponential function; Represents a learnable scalar parameter; Indicates from the first From the first cognitive level to the second Learnable bias terms at each cognitive level; the transition matrix of Bloom's cognitive levels satisfies hour, ; express The Middle line, number The elements of the column represent the elements starting from the first... The cognitive level shifts to the first... Preference weights at each cognitive level; 2. Project the inputs of the cognitive level-sensitive weighted attention module into query vectors. Key vector Sum value vector The input to the cognitive-level sensitive weighted attention module contains cognitive-level labels. Lexical units (cognitive level labels) These are the Bloom cognitive hierarchy labels pre-annotated for each lexical unit during the preprocessing stage. These are inherent properties of the lexical units and do not change with the feature encoding of the Transformer layer. The cognitive hierarchy-aware query vector is calculated. and cognitive level perception key vector : ; ; In the formula, Indicates cognitive level label The corresponding embedding vector (composed of the label) Embedding Matrix from Bloom's Six Levels of Cognition (obtained by index) This represents the learnable projection matrix corresponding to the query vector; This represents the learnable projection matrix corresponding to the key vector; This represents element-wise multiplication; The input of the cognitive level sensitive weighted attention module is the input of the modified Transformer layer; the input feature of the first modified Transformer layer is the diagnostic input embedding matrix, and the input of subsequent modified Transformer layers is the output of the previous modified Transformer layer. 3. Calculate the cognitive-level perceived attention score matrix based on the cognitive-level perceived query vector and the cognitive-level perceived key vector: ; In the formula, The cognitive hierarchy perceptual attention score matrix represents the first... line, number Column elements; Indicates the first The cognitive hierarchy of query vectors for each word element; Indicates the first The cognitive level perception key vector of each word element; Indicates transpose; This represents the feature dimension of the key vector; Indicates from the cognitive level Towards Jump transition weights Indicates the first Each word corresponds to a Bloom cognitive tag. Indicates the first Bloom's cognitive tags corresponding to each word element; express The logarithm of the value is embedded in the core computation path of the attention score as a differentiable bias term. This bias term, along with the inner product of the query vector and the key vector, lies on the same addition path, ensuring... The gradient is correctly propagated to during backpropagation. and It should be noted that this bias term is fundamentally different from fixed-position bias: fixed-position bias encodes the positional distance between tokens using a preset function, does not include learnable parameters, and does not carry domain semantics; the bias term in this invention originates from a parameter that includes learnable parameters. and The cognitive level transition matrix, the bias value varies with the cognitive level. The changes encode the transition relationships between Bloom's six cognitive levels and are continuously optimized through the training process. Furthermore, this invention simultaneously performs element-wise multiplicative modulation of the query vector and key vector through cognitive level embedding, forming a composite cognitive level perceptual structure together with the bias term—a composite structure unprecedented in existing technologies. In existing technologies, cognitive level information is only used as scalar weights or external feature tensors in the data preprocessing stage and does not participate in attention calculation; this invention directly embeds the logarithm of the hierarchical transition matrix into the attention score matrix. The additive path makes cognitive-level information a component of the attentional computational structure, rather than an external auxiliary signal; 4. Introduce Bloom's hierarchy of knowledge corresponding to the course objectives (CLO). Calculate the target-sensitive final attention weights : ; In the formula, Indicates from the cognitive level Bloom's hierarchy of knowledge corresponding to the course objectives (CLO) The transition weight of the jump; Indicates from the cognitive level Bloom's hierarchy of knowledge corresponding to the course objectives (CLO) Jump transition weights Indicates the first Bloom's cognitive tags corresponding to each word element; The cognitive hierarchy perceptual attention score matrix represents the first... line, number The elements of the column.

[0039] Course Learning Outcome is a core concept in Outcome-Based Education (OBE). It refers to the measurable and specific learning outcomes that students must achieve after the course, clearly defining the level of knowledge, skills, and cognitive abilities that students should master. The numerator and denominator structure in the above formula corresponds to the weighted Softmax normalization operation, outputting the normalized goal-sensitive final attention weights. 5. Use target-sensitive final attention weights on the value vector We perform weighted summation to obtain the output of the cognitive level sensitive weighted attention module.

[0040] like Figure 3 As shown, the left side of the figure represents the intermediate feature representation and the Bloom cognitive hierarchy label input corresponding to the word units; the upper middle part shows the Gumbel-Softmax hierarchical distribution calculation path for the verb dimension; and the lower middle part shows the graph attention encoding path for the knowledge domain dimension. The right side of the figure shows a comparison of the transformation effect from fixed weights to learnable weights. Here, "fixed weights" correspond to the manually preset fixed fusion constants in existing technologies, "learnable weights" correspond to the learnable dual-dimensional fusion coefficients of this invention, and "verb-knowledge domain aligned according to Bloom hierarchy" corresponds to the comprehensive alignment score output after fusing the verb dimension hierarchical distribution vector and the knowledge domain feature vector using the learnable fusion coefficients. The processing of the dual-dimensional semantic alignment sub-network includes: 1. Construct a learnable embedding dictionary and hierarchical prototype vectors for Bloom's action verbs: for each predefined Bloom action verb... Assign learnable embedding vectors , Represents the unified feature dimension of verbs and hierarchical prototype vectors; defines six learnable hierarchical prototype vectors. As differentiable centers of verb-level mapping, we model the distribution of Bloom's six cognitive levels in semantic space; The predefined Bloom action verbs refer to a set of signature verbs pre-compiled and organized according to six cognitive levels based on the revised Bloom Taxonomy of Educational Objectives. These include six subsets of verbs: memory, comprehension, application, analysis, evaluation, and creation (e.g., the memory level includes "recognize, remember, enumerate", the application level includes "use, calculate, solve", and the creation level includes "design, construct, create" etc.). The subsets are mutually exclusive. 2. Calculate the continuously differentiable Bloom hierarchy distribution of the action verbs in the course objectives: Calculate the action verbs in the description of the course objectives (corresponding to predefined Bloom action verbs). Euclidean distance between the vectors and the prototype vectors at each level: ; In the formula, Action verbs in the target description With the Euclidean distance between layer prototype vectors; Then, by combining the Gumbel-Softmax relaxation method with a temperature annealing strategy, the hierarchical distribution vector of behavioral verbs describing standardized curriculum objectives is obtained: ; In the formula, Indicates the first The Bloom's six-level cognitive probability distribution vector corresponding to each verb, that is, the hierarchical distribution vector of behavioral verbs describing standardized curriculum objectives; For temperature parameters; This represents the Gumbel-Softmax relaxation method.

[0041] The application of the Gumbel-Softmax relaxation method in this invention is not a conventional application to general classification tasks, but rather, together with learnable hierarchical prototype vectors, Euclidean distance calculation, and temperature annealing strategy, it constitutes a specific mapping structure: using six learnable prototype vectors... As the differentiable center point of Bloom's Level 6 in the semantic space, the continuous probability of hierarchical affiliation is determined by the Euclidean distance between the action verb embedding and each prototype vector, and then a differentiable hierarchical distribution is obtained through Gumbel-Softmax relaxation. This combined structure establishes a specific gradient path: the error gradient of the achievement prediction head can flow backward through the verb hierarchical recognition step, continuously optimizing the prototype vector. The position of the verb in the semantic space allows the verb-level mapping to be automatically refined as training data accumulates. In existing OBE techniques, the verb-level mapping is a discrete lookup table operation, which does not allow backpropagation of gradients and does not have such a learnable prototype vector structure.

[0042] 3. Graph attention encoding of knowledge domain feature vectors: Based on intermediate feature representation, feature vectors corresponding to all knowledge point words in the course objective description, student assessment and answer data and course outline text are extracted through word position indexing, and used as the initial feature vectors of the knowledge domain. A knowledge domain graph is constructed, with knowledge points as nodes and semantic associations and prior dependencies between knowledge points as edges. The initial feature vector is encoded using a multi-layer graph attention network, which adaptively aggregates the feature information of neighboring nodes to update the feature representation of each knowledge point node. The encoding formula is as follows: ; In the formula, Represents a node After the first The first layer obtained after neighborhood aggregation Layer updates feature vectors; Represents a node The set of all adjacent neighbor nodes in the knowledge domain graph; Indicates the first Layer nodes with neighboring nodes The learnable graph attention weight coefficients between them; Indicates the first Learnable linear transformation matrix used for layer feature mapping; Representing neighboring nodes In the The original feature vector of the layer; Represents a nonlinear activation function; After multi-layer graph attention encoding, the feature representations of all knowledge point nodes corresponding to each course objective and each student's assessment answer are aggregated to obtain the knowledge domain feature vector corresponding to the course objective. The knowledge domain feature vector corresponding to student assessment answers .

[0043] 4. Calculate the Verb Dimension Alignment Score (VAAC): Combine the hierarchical distribution vector of action verbs describing standardized curriculum objectives. And the Bloom's hierarchy of cognition labels corresponding to the lexical units, calculate the alignment score for the verb dimension: ; In the formula, Indicates the first The course objectives and the first Alignment scores between verb dimensions of assessment responses; A learnable weight matrix representing verb dimension alignment; The bias term indicating verb-dimensional alignment; This represents a vector concatenation operation; Indicates the first The cognitive level label vector corresponding to each assessment response is taken from the Bloom cognitive level label corresponding to the word.

[0044] 5. Calculate the Knowledge Domain Dimension Alignment Score (KDOC): based on the knowledge domain feature vectors corresponding to the course objectives. The knowledge domain feature vector corresponding to student assessment answers Calculate the similarity score along the knowledge domain dimension: ; In the formula, This represents the function for calculating cosine similarity. Indicates the first The course objectives and the first The assessment is based on the similarity alignment score of the responses in the semantic dimension of knowledge.

[0045] 6. Calculate the two-dimensional comprehensive alignment score SA: By using learnable two-dimensional fusion coefficients, the alignment scores of the verb dimension and the knowledge domain dimension are fused to obtain the comprehensive alignment score: ; In the formula, Indicates the first The course objectives and the first The final comprehensive alignment score, which integrates the verb cognition dimension and the knowledge domain semantic dimension, is used to evaluate the responses. This provides learnable two-dimensional fusion coefficients, replacing the fixed constants in existing technologies. This learnable two-dimensional fusion coefficient is another key feature that distinguishes this invention from existing VKDA technologies: in existing technologies... The fixed constants preset by humans cannot adaptively adjust the relative importance of the two dimensions; the present invention It automatically optimizes during training, adapting to the varying weighting requirements of verb hierarchy compatibility and knowledge domain coverage in different course scenarios, such as... Figure 3 As shown on the right, the training effect of the above-mentioned learnable dual-dimensional fusion coefficients is as follows: the fixed weights preset by humans in the existing technology are transformed into data-driven learnable weights, so that the verb dimension and the knowledge domain dimension can be adaptively aligned according to the Bloom hierarchy.

[0046] Finally, output the integrated alignment tensor. (in For the number of course objectives, (To assess the number of responses).

[0047] 7. Alignment tensor feedback and two-dimensional semantic alignment loss calculation: Alignment Tensor Feedback: Integrating the alignment tensor After being normalized by the Softmax activation function, the weight mapping matrix, which is used to map the course objective to the assessment response, is fed back to the next layer, the cognitive level-sensitive weighted attention module of the Transformer layer, in the final attention weight calculation step, corresponding to the Bloom cognitive level of the course objective (CLO). Jointly participate in weight normalization calculations to establish a fully end-to-end differentiable calculation path from the course outline text to the course objective achievement value; Alignment loss calculation: Taking the intermediate feature representation as input, the Bloom hierarchy probability distribution is first obtained through projection mapping. Then, based on the KL divergence between the Bloom hierarchy distributions of adjacent words, a two-dimensional semantic alignment loss is calculated. : ; In the formula, Indicates the length of the input text sequence; intermediate feature representation Projection matrix with two-dimensional semantic alignment The mapped Bloom hierarchy probability distribution is used to supervise the model's learning of semantic representations that conform to the Bloom hierarchy progression, and to optimize model parameters. express function; The function representing the calculation of KL divergence (Kullback-Leibler divergence); Indicates the first The probability distribution vector of Bloom's six cognitive levels corresponding to each word element; Indicates the first The probability distribution vector of Bloom's six cognitive levels corresponding to each word.

[0048] Step S5: Input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

[0049] like Figure 4 As shown, the processing steps of the Bloom partial order constraint gradient propagation layer include: 1. Based on Bloom's six levels of cognitive development (L1 memory, L2 comprehension, L3 application, L4 analysis, L5 evaluation, L6 creation), following the progressive rule that "the frequency of achievement at lower cognitive levels is no less than that at higher cognitive levels," a partial order relationship is constructed between any two different cognitive levels. Therefore, pairwise combinations from Bloom's six levels of cognitive development yield the following results. Each pair of valid partial order constraints (i.e., L1≥L2, L1≥L3, ..., L5≥L6) is encoded into a fixed constraint matrix. Fixed constraint matrix The element is defined as follows: ; In the formula, Represents a fixed constraint matrix The Middle Line number Column elements, Indicates a low level of cognition The frequency of achievement should be no less than that of higher cognitive levels. ; The symbol represents the number of combinations, used to calculate the number of combinations of any two different levels selected from the six Bloom cognitive levels.

[0050] 2. Using a lightweight prediction head (a single fully connected layer), the final hidden state is mapped to a probability distribution of Bloom's six levels of cognition. : ; In the formula, Indicates the final hidden state; The learnable weight matrix represents the prediction head; This represents the learnable bias term of the prediction head; , Indicates the first The probability of achieving a higher level of cognition. .

[0051] 3. For each pair that satisfies The cognitive level affects the computation of differentiable constraint violations: ; In the formula, Indicates the first The and the first The differentiable constraint violation quantity of each cognitive level; This indicates a modified linear unit activation function; , They represent the first , The probability of achieving a higher level of cognition; when hour, This indicates a low level of cognition. The probability of achieving this is lower than at higher cognitive levels. This violates the Bloom partial order constraint; when the constraint is satisfied, , No punishment.

[0052] 4. Combining linear and superlinear penalties, construct a composite partial order constraint penalty loss function. : ; In the formula, The adjustable weight hyperparameters represent the superlinear penalty term; Indicates batch; Indicates an indicator function.

[0053] 5. Backpropagate the gradient of the composite partial order constraint penalty loss function to the model's fully learnable parameters using the chain rule: ; In the formula, This represents all learnable parameters of the model, including the prediction head parameters, all modified Transformer layer weights, learnable parameters of the cognitively sensitive weighted attention module, and the hierarchical prototype vector of the two-dimensional semantic alignment subnetwork (VKDA). With learnable two-dimensional fusion coefficient This gradient propagation path achieves end-to-end optimization of partial order constraints, unlike the existing BPOG technology which only uses partial order constraints as external verification rules in the inference stage, allowing the constraint satisfaction capability to continuously improve with the accumulation of training data.

[0054] 6. Define a constraint-aware diagnostic generation gating mechanism to control the on-demand activation of the diagnostic process in the multi-task output layer, avoiding unnecessary computational overhead. Specifically: Based on differentiable constraint violation quantity Statistics show that cognitive levels affect The group violation rate (the group violation rate represents the percentage of violations of "low cognitive level" in the current small batch of data) Achievement rate no lower than high cognitive level (Constrained sample proportion); Based on the probability distribution of Bloom's six levels of cognitive hierarchy, the first... Probability of achieving the next level of cognition and the Probability of achieving the next level of cognition and cognitive levels The group violation rate is used to calculate the constraint-aware diagnostic generation gate value. : ; In the formula, Indicating cognitive levels on Group violation rate; This represents the learnable weight matrix for diagnostic generation gating; This represents the learnable bias term for diagnostic generation gating; When the gate value When the preset activation threshold is exceeded, the corresponding cognitive level is triggered. The fine-grained diagnostic generation process (this diagnostic generation process is designed to address violations of cognitive levels) Triggered specific diagnostic output pathways, including: forward reuse of Bloom's six levels of cognitive probability distribution and differentiable constraint violation quantities. And the composite partial order constraint penalty loss function; then connect it to the output layer to generate a diagnostic report fragment for the violation level pair, which includes a refined Bloom distribution, structured error description and improvement suggestions; When the gate value When the threshold is below the preset activation threshold, the fine-grained diagnostic generation process will not perform calculations.

[0055] Among them, the joint loss function of the improved large language model for: ; In the formula, The loss of the basic large language model is usually the cross-entropy loss; The weight hyperparameters representing the two-dimensional semantic alignment loss; The weight hyperparameters representing the penalty loss for the composite partial order constraint; An improved large language model is trained end-to-end using a joint loss function.

[0056] The experimental environment configuration for this embodiment is as follows: Apple Silicon M4 Max chip, 128GB RAM. Software environment: Python 3.10, PyTorch 2.1 deep learning framework, integrating the transformers library (Hugging Face open-source library), the peft library (efficient parameter fine-tuning library), and the accelerate core library (distributed training acceleration library). Attention calculation introduces the FlashAttention efficient attention mechanism to reduce GPU memory usage and improve training throughput. Training dataset: 5200 OBE course diagnostic samples, covering 24 core courses in engineering majors such as Computer Science, Electronic Information, and Mechanical Engineering. Bloom's hierarchy of knowledge is evenly distributed, and the training set (4680 samples) and validation set (520 samples) are divided in a 9:1 ratio. Training hyperparameters: AdamW optimizer, initial learning rate... The weight decay is 0.05, the random seed is 42, and the batch size is 16. A two-stage training strategy is adopted: the initial 5 rounds (learning rate...) + 10 rounds of continued training (learning rate) Early stopping was achieved based on validation loss. Statistical analysis used effect size (Cohen's d) and Bootstrap resampling (1000 times) to construct 95% confidence intervals, and all experimental results were reported based on the mean of 5-fold cross-validation.

[0057] This embodiment sets up comparative experiments: ChatGLM-small (the original large language model) and ChatGLM4OBE (the improved large language model of this invention). The comprehensive performance evaluation results are shown in Table 1. In Table 1, Val.Loss represents the total loss of the validation set, PPL represents the perplexity of the language model, ROUGE-L represents the longest common subsequence similarity, and Cohen'sd represents the Cohen effect size.

[0058] Table 1. Overall performance evaluation results (400 test samples, mean of 5-fold cross-validation)

[0059] As shown in Table 1, the satisfaction rate of Bloom's partial order constraint increased from 45.2% to 100.0%, the accuracy of course objective diagnosis increased by +7.70pp, and the format completeness rate increased by +18.5pp. The improvement of the above core indicators all have a large effect magnitude.

[0060] This embodiment also includes a cross-sectional comparison experiment with six benchmark models: GPT-2-medium (generative pre-trained Transformer-2 medium-sized model), LLaMA-7B (large language model-7B 4-bit quantized zero-shot model), BERT-base-Chinese (bidirectional encoder), ChatGLM-small+P-tuningv2 (ChatGLM-small + hint tuning v2), and ChatGLM-small+LoRA (ChatGLM-small + low-rank adaptive fine-tuning). The results of the multi-model cross-sectional comparison are shown in Table 2.

[0061] Table 2. Results of cross-sectional comparison of multiple models

[0062] As shown in Table 2, the improved large language model (ChatGLM4OBE) of this invention, with the same number of parameters, has a course objective diagnosis accuracy 2.1 times that of the suboptimal model, and a partial order constraint satisfaction rate of 100.0%, far exceeding the suboptimal model's 65.2%. LLaMA-7B has 20 times the number of parameters as ChatGLM4OBE, yet its zero-sample course objective accuracy is lower, verifying that parameter size cannot replace architectural-level structural transformation for outcome-oriented education.

[0063] The ablation experiment results of this embodiment are shown in Table 3.

[0064] Table 3 Comparison of ablation test results

[0065] In Table 3, the baseline model is ChatGLM-small (the original large language model), +CLSWA is the baseline model with a cognitive-level sensitive weighted attention module, +VKDA is the baseline model with a two-dimensional semantic alignment sub-network, +BPOG is the baseline model with a Bloom partial order constraint gradient propagation layer, and Full is the baseline model with a cognitive-level sensitive weighted attention module, a two-dimensional semantic alignment sub-network, and a Bloom partial order constraint gradient propagation layer, which is the improved large language model of this invention. As can be seen from Table 3, the full model (Full) achieves the optimal value in all five evaluation metrics.

[0066] This embodiment also included a cross-disciplinary generalization experiment. 100 samples were drawn from each of the Electronic Information Engineering major's "Analog Circuits" and the Mechanical Engineering major's "Fundamentals of Mechanical Manufacturing" courses, which were not included in the training. The cross-disciplinary generalization test results are shown in Table 4.

[0067] Table 4. Cross-disciplinary generalization test results

[0068] As shown in Table 4, the accuracy of course objective diagnosis for foreign majors decreased by 12.0% and 17.3% relative to the training domain, which is within a reasonable range of generalization decay. The satisfaction rate of Bloom's partial order constraint remained at 100.0% across all three majors, proving that the internalization of Bloom's partial order constraint does not depend on the specific professional vocabulary distribution and has cross-professional universality.

[0069] Example 2

[0070] This embodiment provides an outcome-oriented educational diagnostic system based on Bloom's hierarchy of cognitive constraints, used to execute an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints as described in Embodiment 1, including: The diagnostic input word sequence construction module is used to obtain the course outline text, course objective description, and student assessment response data of the course to be diagnosed, and to perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; based on the course outline text, the annotated course objective description, and the student assessment response data, the diagnostic input word sequence is constructed. The model building module is used to build an improved large language model. The improved large language model consists of an embedding layer, an N-layer modified Transformer layer, a Bloom partial order constraint gradient propagation layer, and an output layer connected in sequence. The modified Transformer layer includes replacing the original multi-head self-attention module in the original Transformer layer with a cognitive level sensitive weighted attention module, and connecting the two-dimensional semantic alignment sub-network in parallel with the feedforward network path in the Transformer layer in the form of a bypass. The diagnostic input lexical sequence processing module is used to process the diagnostic input lexical sequence into the embedding layer of the improved large language model to obtain the diagnostic input embedding matrix. The diagnostic input embedding matrix processing module processes the diagnostic input embedding matrix through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the cognitive-level perceptual attention output. After residual and layer normalization, an intermediate feature representation is obtained. This intermediate feature representation is then fed in parallel into a feedforward network and a two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final target-sensitive attention weights for that layer. The alignment weights of the first cognitive-level sensitive weighted attention module are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers. The output module is used to input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

[0071] Example 3

[0072] This embodiment provides an electronic device, including a processor, a memory, and a bus. The processor and the memory are connected via the bus. The memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to Embodiment 1.

[0073] Example 4

[0074] A non-volatile computer storage medium storing computer-executable instructions that execute an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints, as described in Embodiment 1.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints, characterized in that, Includes the following steps: Step S1: Obtain the course outline text, course objective description, and student assessment response data of the course to be diagnosed; perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; construct a diagnostic input word sequence based on the course outline text, the annotated course objective description, and the student assessment response data. Step S2: Construct an improved large language model, which consists of an embedding layer, an N-layer modified Transformer layer, a Bloom partial order constraint gradient propagation layer, and an output layer connected in sequence; wherein, the modified Transformer layer includes: replacing the original multi-head self-attention module in the original Transformer layer with a cognitive level sensitive weighted attention module, and connecting the two-dimensional semantic alignment sub-network in parallel with the feedforward network path in the Transformer layer in the form of a bypass; Step S3: Input the diagnostic input word sequence into the embedding layer of the improved large language model for processing to obtain the diagnostic input embedding matrix; Step S4: The diagnostic input embedding matrix is ​​sequentially processed through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the attention output of the cognitive level perception. After residual and layer normalization, an intermediate feature representation is obtained. The intermediate feature representation is then fed into the feedforward network and the two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final attention weights of the target sensitivity of the corresponding layer. The alignment weights of the cognitive-level sensitive weighted attention module of the first Transformer layer are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers. Step S5: Input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

2. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints as described in claim 1, characterized in that: The construction of the diagnostic input embedding matrix in step S3 includes: calculating the standard word embedding, rotation position encoding, and Bloom hierarchical embedding obtained from the learnable Bloom six-level cognitive hierarchy embedding matrix for each word in the diagnostic input word sequence, and adding the three elements one by one to obtain the final input embedding vector.

3. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to claim 2, characterized in that: The processing steps of the cognitive-level sensitive weighted attention module include: Define the transition matrix of Bloom's cognitive levels; The inputs of the cognitive-level sensitive weighted attention module are projected into query vectors, key vectors, and value vectors, respectively. Based on the Bloom cognitive hierarchy labels of the lexical units, the query vector and key vector are subjected to element-wise multiplication modulation of cognitive hierarchy embedding to obtain cognitive hierarchy perceived query vector and cognitive hierarchy perceived key vector. Attention scores are calculated based on the cognitive level-aware query vector and cognitive level-aware key vector, and the logarithmic bias term obtained from the cognitive level transition matrix is ​​embedded into the addition path of the attention scores; the Bloom cognitive level corresponding to the course objectives is introduced to calculate the final attention weights that are sensitive to the objectives; the final attention weights are used to weight and sum the value vectors to obtain the output.

4. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to claim 3, characterized in that: The logarithmic bias term is obtained by taking the logarithm of the transition weights in the cognitive level transition matrix. The transition matrix contains learnable scalar parameters and learnable bias terms, and satisfies asymmetric constraints.

5. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to claim 4, characterized in that: The processing steps of the two-dimensional semantic alignment sub-network include: Construct a learnable embedding dictionary and hierarchical prototype vectors for Bloom's action verbs; Calculate the Euclidean distance between the course objective action verbs and the prototype vectors at each level, and obtain the hierarchical distribution vectors of the standardized course objective description action verbs through a Gumbel-Softmax relaxation combined with a temperature annealing strategy. , represented as: ； In the formula, Indicates the first The Bloom's six-level cognitive probability distribution vector corresponding to each verb, that is, the hierarchical distribution vector of behavioral verbs describing standardized curriculum objectives; For temperature parameters; Action verbs in the target description With the Euclidean distance between layer prototype vectors; This represents the Gumbel-Softmax relaxation method; Based on intermediate feature representations, an initial feature vector of the knowledge domain is extracted. Then, the knowledge domain feature vectors corresponding to the course objectives and the knowledge domain feature vectors corresponding to student assessment responses are obtained through graph attention network encoding. The hierarchical distribution vector of action verbs in combination with standardized curriculum objectives And the Bloom's hierarchy of cognition labels corresponding to the lexical units, calculate the alignment score for the verb dimension; Based on the knowledge domain feature vector corresponding to the course objectives The knowledge domain feature vector corresponding to student assessment answers Calculate the similarity score along the knowledge domain dimension; By using a learnable two-dimensional fusion coefficient, the alignment scores of the verb dimension and the knowledge domain dimension are fused to obtain a comprehensive alignment score; The comprehensive alignment score is passed to the final attention weight calculation step of the cognitive level sensitive weighted attention module of the next layer of the Transformer layer, and participates in the weight normalization operation together with the Bloom cognitive level corresponding to the course objectives.

6. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to claim 5, characterized in that: The processing steps of the Bloom partial order constraint gradient propagation layer include: Construct a fixed constraint matrix based on the partial order relation of Bloom's six-level cognitive hierarchy; The final hidden state is mapped to a probability distribution of Bloom's six cognitive levels by using a prediction head; For each pair satisfied The cognitive level affects the computation of differentiable constraint violations: ； In the formula, Indicates the first The and the first The differentiable constraint violation quantity of each cognitive level; This represents the modified linear unit activation function; , They represent the first , The probability of achieving a higher level of cognition; Represents a fixed constraint matrix The Middle Line 1 Column elements; when hour, This indicates a low level of cognition. The probability of achieving this is lower than at higher cognitive levels. This violates the Bloom partial order constraint; when the constraint is satisfied, , No punishment; Constructing a composite partial order constraint penalty loss function by combining linear and superlinear penalties : ； In the formula, The adjustable weight hyperparameters represent the superlinear penalty term; Indicates batch; Indicates an indicator function; The gradient of the composite partial order constraint penalty loss function is backpropagated to the model's full-link learnable parameters; the model's full-link learnable parameters include the prediction head parameters, the weights of all modified Transformer layers, the learnable parameters of the cognitive level sensitive weighted attention module, the hierarchical prototype vector of the two-dimensional semantic alignment subnetwork, and the learnable two-dimensional fusion coefficients.

7. The outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints according to claim 6, characterized in that: Attention scores are calculated based on the cognitive hierarchy-aware query vector and the cognitive hierarchy-aware key vector, and are expressed as follows: ； In the formula, The cognitive hierarchy perceptual attention score matrix represents the first... line, number Column elements; Indicates the first The cognitive hierarchy of query vectors for each word element; Indicates the first The cognitive level perception key vector of each word element; Indicates transpose; This represents the feature dimension of the key vector; Indicates from the cognitive level Towards Jump transition weights Indicates the first Each word corresponds to a Bloom cognitive tag. Indicates the first Bloom's cognitive tags corresponding to each word element; express The logarithm of .

8. An outcome-oriented educational diagnostic system based on Bloom's hierarchy of cognitive constraints, characterized in that, include: The diagnostic input word sequence construction module is used to obtain the course outline text, course objective description, and student assessment response data of the course to be diagnosed, and to perform Bloom's hierarchy of knowledge annotation on the action verbs and knowledge domains in the course objective description; based on the course outline text, the annotated course objective description, and the student assessment response data, the diagnostic input word sequence is constructed. The model building module is used to build an improved large language model. The improved large language model consists of an embedding layer, an N-layer modified Transformer layer, a Bloom partial order constraint gradient propagation layer, and an output layer connected in sequence. The modified Transformer layer includes replacing the original multi-head self-attention module in the original Transformer layer with a cognitive level sensitive weighted attention module, and connecting the two-dimensional semantic alignment sub-network in parallel with the feedforward network path in the Transformer layer in the form of a bypass. The diagnostic input lexical sequence processing module is used to process the diagnostic input lexical sequence into the embedding layer of the improved large language model to obtain the diagnostic input embedding matrix. The diagnostic input embedding matrix processing module processes the diagnostic input embedding matrix through N Transformer layers. In each Transformer layer, the cognitive-level sensitive weighted attention module first calculates the cognitive-level perceptual attention output. After residual and layer normalization, an intermediate feature representation is obtained. This intermediate feature representation is then fed in parallel into a feedforward network and a two-dimensional semantic alignment sub-network for processing. The alignment score output by the two-dimensional semantic alignment sub-network is passed to the cognitive-level sensitive weighted attention module of the next Transformer layer to guide the calculation of the final attention weights for the target sensitivity of the corresponding layer. The alignment weights of the cognitive-level sensitive weighted attention module in the first Transformer layer are uniformly initialized. The output of the feedforward network is passed to the next Transformer layer after residual and layer normalization. The final hidden state is obtained after processing by N Transformer layers. The output module is used to input the final hidden state into the Bloom partial order constraint gradient propagation layer, and map the final hidden state to the probability distribution of the Bloom six-level cognitive hierarchy through the prediction head; the output layer decodes and generates the course goal achievement value and structured improvement suggestions based on the final hidden state and the probability distribution of the Bloom six-level cognitive hierarchy, and finally outputs a diagnostic report.

9. An electronic device comprising a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein... The memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory, characterized in that it executes an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints as described in any one of claims 1-7.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform an outcome-oriented educational diagnostic method based on Bloom's hierarchy of cognitive constraints as described in any one of claims 1-7.

Images