CN-121811207-B - Multi-view image transducer cognitive evaluation system, method and storage medium fusing explicit characteristics

CN121811207BCN 121811207 BCN121811207 BCN 121811207BCN-121811207-B

Abstract

The invention discloses a multi-view graph converter cognitive evaluation system, a multi-view graph converter cognitive evaluation method and a storage medium which are integrated with explicit characteristics. The system comprises a feature embedding module, a view information aggregation module, a multi-view feature fusion module and a cognition evaluation module, wherein the feature embedding module is used for generating basic characterizations for three types of entities of learners, topics and knowledge concepts, respectively integrating difficulty coefficients of the topics and the knowledge concepts into the characterizations, the view information aggregation module is used for constructing three heterogeneous bipartite graphs of learners-topics, learners-knowledge concepts and topics-knowledge concepts and updating node characterizations based on a graph transducer, the multi-view feature fusion module is used for fusing characterizations of the entities in different views to generate specialized characterizations, and the cognition evaluation module is used for mapping the specialized characterizations of the learners into a vector space taking the number of the knowledge concepts as a dimension, and the obtained vector is a cognition evaluation result and is used for predicting the probability of answering any topic. According to the method, accuracy and interpretability of cognitive assessment are remarkably improved through explicit difficulty fusion, multi-view graph structure modeling and graph transform depth interaction.

Inventors

Zhao Xiangdie
Nie Ruotong
MA YINGHONG
LIU ZHIYUAN
YOU XUEMEI

Assignees

山东师范大学

Dates

Publication Date: 20260512
Application Date: 20260311

Claims (8)

1. The multi-view image transducer cognitive evaluation system integrating the explicit characteristics is characterized by comprising the following steps: the system comprises a feature embedding module, a feature embedding module and a feature embedding module, wherein the feature embedding module is used for defining basic representation of an entity in a vector space in a form of a trainable matrix, the entity comprises a learner, a question and a knowledge concept, and is used for defining the question difficulty coefficient and the knowledge concept difficulty coefficient and mapping the question difficulty coefficient and the knowledge concept difficulty coefficient to the basic representation of the question and the knowledge concept respectively, and in the feature embedding module, the question difficulty coefficient and the knowledge concept difficulty coefficient are mapped to the basic representation of the question and the knowledge concept respectively through a proprietary linear transformation layer, and the feature embedding module is specifically as follows: ; ; In the formula, 、 Respectively representing the question difficulty coefficient and the knowledge concept difficulty coefficient embedding; 、 Respectively representing the question difficulty coefficient and the knowledge concept difficulty coefficient; 、 respectively representing weight matrix parameters and bias parameters of the question difficulty coefficient embedded model; 、 respectively representing weight matrix parameters and bias parameters in the knowledge concept difficulty coefficient embedding model; Then: The initial characterization of the learner is the basic characterization of the learner, the initial characterization of the topic is the basic characterization of the learner and the embedded addition of the topic difficulty coefficient, and the initial characterization of the knowledge concept is the basic characterization of the learner and the embedded addition of the knowledge concept difficulty coefficient; The system comprises an in-view information aggregation module, a graph transformation module and a graph analysis module, wherein the in-view information aggregation module is used for establishing a heterogeneous bipartite graph and updating node characterization in the heterogeneous bipartite graph based on a graph transformation module, wherein the heterogeneous bipartite graph comprises a learner-topic graph, a learner-knowledge conceptual graph and a topic-knowledge conceptual graph; in the multi-view feature fusion module, two sub-tokens of a single entity in different heterogeneous bipartite graphs are stacked into a view matrix, the view matrix is input into a self-attention module, the self-attention module calculates similarity scores among sub-tokens, and allocates attention weights to each sub-token based on the similarity scores, and calculates weighted aggregate tokens of the entities based on the attention weights allocated in the sub-tables; Carrying out average pooling on the weighted aggregation characterization along the dimension of the view matrix, and obtaining the specialization characterization of the entity through layer normalization; the cognitive evaluation module is used for mapping the specialized characterization signs of the entities into a vector space taking the number of knowledge concepts as a dimension, taking the mapped characterization of a learner as a learner cognitive evaluation result, and taking the predicted correct probability of answering questions to the learner as an output result of the cognitive evaluation module.
2. The multi-view transducer cognitive assessment system incorporating explicit features of claim 1, Representing interactive record characterization among learners, questions and knowledge concepts by using a triplet set, and representing the knowledge concepts associated with the questions by using a Q matrix; the question difficulty coefficient is defined as the difference between 1 and the question response accuracy of all learners in the interaction record; the knowledge concept difficulty coefficient is defined as the difference between 1 and the accuracy of the solution of all learners to the knowledge concept related questions in the interaction record.
3. The multi-view transducer cognitive assessment system incorporating explicit features of claim 2, And in the cognitive evaluation module, mapping the specialization representation of the entity into a vector space with the number of knowledge concepts as dimensions through an independent linear layer, and enhancing the topic representation through a knowledge enhancement mechanism.
4. A multi-view transducer cognitive assessment system incorporating explicit features as claimed in claim 3, The knowledge enhancement mechanism is configured to: Averaging the mapped knowledge concept set associated with the topic to obtain a knowledge background vector of the topic; And splicing the mapped topic representation with the knowledge background vector to obtain the topic representation after knowledge enhancement.
5. The multi-view transducer cognitive assessment system incorporating explicit features of claim 4, And in the cognition evaluation module, calculating the vector inner product of the characteristic of the learner after mapping and the characteristic of the question after knowledge enhancement, and outputting the prediction probability of the answer of the learner to the question through function mapping.
6. The multi-view graph transform cognitive evaluation method based on the multi-view graph transform cognitive evaluation system with the integrated explicit characteristics is characterized by comprising the following steps of: s1, acquiring records of questions answered by learners, acquiring Q matrixes corresponding to the questions answered, and distributing an initial learning embedded vector to the learners; s2, constructing a learner-topic map, a learner-knowledge conceptual map and a topic-knowledge conceptual map; s3, running a multi-view graph transform cognitive evaluation system with integrated explicit characteristics, which is completed by training, on the three heterogeneous bipartite graphs, and updating all nodes; S4, fusing the characterization of the learner in the learner-question map and the learner-knowledge conceptual map, and outputting the specialized characterization of the learner; s5, mapping the specialized representation of the learner to the cognitive evaluation module, and outputting a cognitive evaluation result of the learner through the cognitive evaluation module.
7. The method for multi-view map transducer cognitive assessment with fusion of explicit features according to claim 6, The method is used for predicting the correct answering probability of the learner on the new questions which are not answered, and comprises the following steps of obtaining the question marks of the new questions which are not answered and the associated knowledge concept marks thereof, inputting the specialization marks of the learner, the question marks of the new questions which are not answered and the associated knowledge concept marks thereof into the cognitive evaluation module so as to output the predicting probability of the new questions which are not answered.
8. A computer readable storage medium, characterized in that a computer program is stored thereon, which program, when being executed by a processor, implements the method of any of claims 6-7.

Description

Multi-view image transducer cognitive evaluation system, method and storage medium fusing explicit characteristics Technical Field The application relates to the technical field of cognitive state evaluation and prediction, in particular to a multi-view graph transform cognitive evaluation system, a multi-view graph transform cognitive evaluation method and a storage medium which are fused with explicit features. Background The online learning platform is stepping into the key stage of quantity expansion and technical development, and the education field is continuously generating and accumulating unprecedented learner behavior data. The rich behavior data provides infinite possibility for cognitive evaluation and personalized learning guidance of the user. The learner, the knowledge concept and the questions can be used as three independent entities, and the process of solving the questions after the learner learns the knowledge concept can be regarded as comprehensive interaction. The comprehensive interaction is split, and the interactions of independent entities are exposed, namely learning knowledge concepts by a learner, answering questions by the learner and containing knowledge concepts by questions. Through split interaction, the cognitive state of a learner can be characterized and inferred, the inherent attribute and the associated characteristic of the knowledge concept are characterized, and the knowledge investigation requirement and the difficulty characteristic of the subject are characterized. Early cognitive diagnostic models were based on statistical methods, relying on mathematical methods of logistic functions and parameter estimation for reasoning, typical methods such as project reaction theory and DINA. Early approaches had good interpretability, but their limited modeling capabilities are difficult to describe real-world complex interactions, and diagnostic accuracy and generalization are constrained. With the intervention of deep learning, the cognitive diagnostic model is able to more deeply mine the intrinsic correlations between entities. Deep learning can describe the internal association between entities, but in the information aggregation stage, the GNN aggregator on which the current deep learning model depends has insufficient expression capability when facing the dependency relationship of high-frequency dynamic and complex interleaving in knowledge interaction. In addition, in the feature utilization dimension, existing depth models rely primarily on implicit tokens learned from learner question recordings, which cannot be incorporated into display tokens such as question difficulty. From an educational perspective, explicit features have a definite educational significance, limiting the token richness of the model when explicit features cannot be incorporated. Disclosure of Invention In a first aspect, the application provides a multi-view map transducer cognitive evaluation system incorporating explicit features. The technical scheme of the application is as follows, a multi-view image transducer cognitive evaluation system integrating explicit characteristics comprises: the feature embedding module is used for defining basic representation of an entity in a vector space in a form of a trainable matrix, wherein the entity comprises a learner, a question and a knowledge concept, and is used for defining a question difficulty coefficient and a knowledge concept difficulty coefficient and mapping the question difficulty coefficient and the knowledge concept difficulty coefficient to the basic representation of the question and the knowledge concept respectively; The system comprises an in-view information aggregation module, a graph transformation module and a graph analysis module, wherein the in-view information aggregation module is used for establishing a heterogeneous bipartite graph and updating node characterization in the heterogeneous bipartite graph based on a graph transformation module, wherein the heterogeneous bipartite graph comprises a learner-topic graph, a learner-knowledge conceptual graph and a topic-knowledge conceptual graph; the multi-view feature fusion module is used for fusing the representation of the entity in each heterogeneous bipartite graph and outputting the specialization representation of each entity; the cognitive evaluation module is used for mapping the specialized characterization signs of the entities into a vector space taking the number of knowledge concepts as a dimension, taking the mapped characterization of a learner as a learner cognitive evaluation result, and taking the predicted correct probability of answering questions to the learner as an output result of the cognitive evaluation module. Further, representing interactive record characterization among learners, questions and knowledge concepts by using a triplet set, and representing the knowledge concepts associated with the questions by using a Q matrix; the question di