An automatic fusion method and system based on knowledge distillation and graph neural networks

Abstract

This invention belongs to the field of automatic knowledge fusion, specifically involving an automatic fusion method and system based on knowledge distillation and graph neural networks. The method includes: collecting multi-source heterogeneous urban construction project data; extracting Logit vectors from teacher-student models and generating softened probability distributions; calculating teacher knowledge correction loss and non-target category relationship extraction loss based on Logit variance; completing distillation training by combining cross-entropy loss to obtain a student model for semantic feature extraction; constructing paired connected graphs based on source and target ontologies; fusing multi-dimensional similarity features; and outputting cross-ontology entity matching results after processing by a convolutional neural network, thus achieving standardized and unified fusion of multi-source urban construction knowledge. This invention solves the problems of heterogeneous and chaotic urban construction knowledge, low efficiency of manual fusion, poor matching accuracy, and difficulty in automated iterative updates.

Description

Technical Field

[0001] This invention belongs to the field of automatic knowledge fusion, specifically relating to an automatic fusion method and system based on knowledge distillation and graph neural networks. Background Technology

[0002] In complex engineering systems, urban construction management, housing and construction supervision, and the operation of large-scale infrastructure projects, urban construction project management systems often involve multi-level, multi-entity, and multi-element information interaction and decision-making collaboration. With the development of informatization and digitalization, the data scale upon which multi-source urban construction project management databases rely has expanded rapidly, and the data types exhibit a highly heterogeneous trend, including structured project ledgers, bidding data, contract lists, progress reports, cost data, unstructured construction logs, supervision reports, design drawings, BIM model data, spatiotemporal geographic information, and expert experiential knowledge. These information types have complex relationships, making it difficult to effectively model and integrate them using traditional rule-based or manual construction methods. This results in high knowledge acquisition costs, slow updates, and difficulty in adapting to rapidly changing business needs.

[0003] In recent years, the development of artificial intelligence, especially deep learning, has provided new technical pathways for knowledge modeling. However, deep neural networks typically rely on large amounts of labeled data. For scenarios like urban construction project management, which are highly specialized and characterized by implicit and structured knowledge, directly using deep learning models often faces problems such as data scarcity, insufficient generalization ability, and inadequate interpretability. Meanwhile, knowledge graph technology provides an important tool for representing entities, attributes, and relationships in urban construction project management; however, traditional knowledge graph construction methods heavily rely on manual rules or templates, making it difficult to achieve automated and high-precision knowledge fusion.

[0004] Graph Neural Networks (GNNs) have become a core technology for processing graph-structured data in recent years. Their ability to perform feature propagation and relational reasoning on graph structures makes them suitable for representing highly interconnected knowledge structures in urban construction project management. However, GNN model training also requires high-quality data, and its application in the construction and integration of knowledge in urban construction project management is constrained by issues such as data scarcity, strong heterogeneity, and high noise levels.

[0005] Against this backdrop, Knowledge Distillation (KD) technology offers a novel solution. Its core idea is to leverage a high-performance, parameter-rich "teacher model" to transfer knowledge to a lightweight "student model," enabling the student model to maintain high performance while possessing better deployability and generalization capabilities. For knowledge construction in urban construction project management, knowledge distillation can be used to extract refined knowledge from complex or multi-source models, allowing graph neural networks to obtain more effective structured representations even with insufficient or noisy data. Simultaneously, knowledge distillation can achieve representation alignment and transfer between different types of data sources, improving the efficiency and quality of knowledge fusion.

[0006] However, directly combining knowledge distillation with graph neural networks for urban construction project management modeling is not easy. On the one hand, urban construction project management knowledge has strong structural dependencies and cross-domain characteristics, and how to maintain the structural information and logical relationships in the knowledge during the distillation process is a challenge for existing methods. On the other hand, knowledge from different sources often has inconsistencies or even conflicts, and direct fusion will lead to a decline in model performance. Summary of the Invention

[0007] The technical problem this invention aims to solve is to provide an automatic fusion method and system based on knowledge distillation and graph neural networks. This addresses issues in urban construction project management modeling, such as high heterogeneity of multi-source knowledge, inconsistent quality, complex structural relationships, and the reliance on manual intervention and difficulty in automatic updates in traditional knowledge fusion methods. This enables automatic cross-system data association and business collaboration, and can automatically verify project information during construction permit processing, improving approval efficiency. Furthermore, this method replaces the high-cost and difficult-to-maintain traditional system integration method with automated matching rules, significantly reducing long-term operating costs.

[0008] An automatic fusion method based on knowledge distillation and graph neural networks includes: Obtain multi-source heterogeneous urban construction project data, input the data into the teacher model and student model, and extract the teacher Logit vector and student Logit vector respectively; The teacher's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the teacher's softening probability distribution. The student's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the student's softening probability distribution. The Logit variance is calculated based on the teacher's Logit vector. The Logit variance is then used to correct the teacher's Logit vector to obtain a corrected vector. A softmax transformation is performed on the corrected vector to obtain the corrected teacher softening probability distribution. The KL divergence is calculated based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. Based on the teacher softening probability distribution, non-target categories with a probability lower than the correct category are filtered out. The student Logit vector and teacher Logit vector corresponding to the non-target category are corrected respectively to obtain the corrected student Logit vector and the corrected teacher Logit vector. Based on the corrected student Logit vector and the corrected teacher Logit vector, the non-target category relationship extraction loss is obtained. The total training loss is obtained based on the teacher knowledge correction loss, the non-target category relationship extraction loss, and the classification cross-entropy loss. The student model is then optimized based on the total training loss to obtain the optimized student model. Obtain the source city construction project management ontology and the target city construction project management ontology to be matched, and construct a pairwise connected graph of project ontology entities. The nodes of the connected graph represent project entity concept pairs between the source and target ontologies, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is used as the source ontology, and the target city construction project management ontology is used as the target ontology. A basic similarity matrix is ​​constructed based on the attributes of entities in the source ontology and the target ontology. The entities of the source ontology are input into the optimized student model to obtain deep semantic features. An enhanced similarity matrix is ​​obtained based on the basic similarity matrix and the deep semantic features. The enhanced similarity matrix is ​​input into a convolutional neural network to obtain low-dimensional dense features. The low-dimensional dense features are concatenated with entity label similarity features to obtain initial node features. The initial node features are input into a graph neural network to complete feature propagation and aggregation on a pairwise connected graph to obtain node depth representation. The entity label similarity is calculated through item entity concept pairs. Based on the node depth representation, perform binary classification matching prediction, output the entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

[0009] Optionally, the teacher's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the teacher's softening probability distribution, and the student's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the student's softening probability distribution, expressed as follows:

[0010] in, For temperature coefficient, Let C be the Logit vector, and C be the total number of classes. Let m represent the j-th Logit vector, where m = T or S. When m = T, Let the teacher's Logit vector be... For the teacher's softened probability distribution, when m=S, Let the student's Logit vector be... Soften the probability distribution for students.

[0011] Optionally, the Logit variance is calculated based on the teacher Logit vector, and the teacher Logit vector is corrected using the Logit variance to obtain a corrected vector, including: Based on the teacher Logit vector, the average value of the teacher Logit vector is obtained; Based on the average value of the teacher's Logit vector and the average value of the teacher's Logit vector, the Logit variance is obtained; The teacher's Logit vector is corrected based on the Logit variance to obtain the corrected vector.

[0012] Optionally, the step of filtering non-target categories with probabilities lower than the correct category based on the teacher softening probability distribution, and correcting the Logit vectors corresponding to the non-target categories to obtain corrected student Logit vectors and corrected teacher Logit vectors, and obtaining the non-target category relation extraction loss based on the corrected student Logit vectors and corrected teacher Logit vectors, includes: Based on the teacher softening probability distribution, non-target categories with probabilities lower than the correct category are filtered to obtain a non-target category mask; Based on the non-target class mask, the Logit variance, and the preset coefficients, the Logit vectors corresponding to the non-target classes are modified to obtain the modified student Logit vector and the modified teacher Logit vector. The modified student Logit vector and the modified teacher Logit vector are input into the softmax function to obtain the modified teacher probability distribution and the modified student probability distribution; Based on the modified teacher probability distribution and the modified student probability distribution, the non-target category relationship extraction loss is calculated.

[0013] Optionally, the step of constructing a basic similarity matrix based on the attributes of entities in the source ontology and the target ontology, inputting the entities of the source ontology into the optimized student model to obtain deep semantic features, and obtaining an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features includes: Based on the data attributes and entity attributes of the source ontology and the target ontology, a basic similarity matrix is ​​calculated; The entities of the source ontology are input into the optimized student model to obtain deep semantic features; The cosine similarity is calculated based on the deep semantic features of the source and target ontologies. An enhanced similarity matrix is ​​constructed based on the basic similarity matrix and the cosine similarity.

[0014] Optional methods for calculating entity tag similarity include: Based on the label names of the source and target ontologies, calculate character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance normalization; The character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance are normalized and concatenated to obtain the entity label similarity.

[0015] Optionally, the initial node features are input into a graph neural network, and feature propagation and aggregation are performed on paired connected graphs to obtain the node depth representation, as follows:

[0016] in, For node depth representation, For activation function, For attention, For shared matrices, These are the initial node features.

[0017] An automatic fusion system based on knowledge distillation and graph neural networks includes: The first acquisition module is used to acquire multi-source heterogeneous urban construction project data, input the multi-source heterogeneous urban construction project data into the teacher model and the student model, and extract the teacher Logit vector and the student Logit vector respectively. The probability distribution calculation module is used to perform a softmax transformation on the teacher's Logit vector using a temperature coefficient to obtain the teacher's softening probability distribution, and to perform a softmax transformation on the student's Logit vector using a temperature coefficient to obtain the student's softening probability distribution. The teacher knowledge correction loss calculation module is used to calculate the Logit variance based on the teacher Logit vector, correct the teacher Logit vector using the Logit variance to obtain a corrected vector, perform a softmax transformation on the corrected vector to obtain the corrected teacher softening probability distribution, and calculate the KL divergence based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. The non-target category relation extraction loss calculation module is used to filter non-target categories with probabilities lower than the correct category based on the teacher softening probability distribution, correct the student Logit vector and teacher Logit vector corresponding to the non-target category respectively to obtain corrected student Logit vector and corrected teacher Logit vector, and obtain the non-target category relation extraction loss based on the corrected student Logit vector and teacher Logit vector. The optimization module is used to obtain the total training loss based on the teacher knowledge correction loss, the non-target category relationship extraction loss, and the classification cross-entropy loss, and to optimize the student model based on the total training loss to obtain an optimized student model. The second acquisition module is used to acquire the source city construction project management ontology and the target city construction project management ontology to be matched, and to construct a pairwise connected graph of project ontology entities. The nodes of the connected graph represent project entity concept pairs between the source and target ontology, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is the source ontology, and the target city construction project management ontology is the target ontology. An enhanced similarity matrix construction module is used to construct a basic similarity matrix based on the attributes of entities in the source ontology and the target ontology, input the entities of the source ontology into the optimized student model to obtain deep semantic features, and obtain an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features. The feature calculation module is used to input the enhanced similarity matrix into a convolutional neural network to obtain low-dimensional dense features, concatenate the low-dimensional dense features with entity label similarity features to obtain initial node features, input the initial node features into a graph neural network, complete feature propagation and aggregation on a pairwise connected graph to obtain node depth representation, and the entity label similarity is calculated through item entity concept pairs; The fusion module is used to perform binary classification matching prediction based on the node depth representation, output an entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

[0018] A terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs an automatic fusion method based on knowledge distillation and graph neural networks.

[0019] A computer-readable storage medium storing a computer program, which, when loaded and executed by a processor, employs an automatic fusion method based on knowledge distillation and graph neural networks.

[0020] The beneficial effects of this invention are: 1. Based on the adaptive Logit reconstruction online distillation (ALRKD), and with the help of teacher error correction (CC) and non-target class relationship learning (OC) dual channels, the Logit vector is extracted from the input data first, and the temperature softmax soft label is constructed. Then, the error correction is based on Logit variance through CC and the relationship between categories is extracted through OC, which reduces noise transmission and improves the accuracy of teacher knowledge. Finally, a unified knowledge expression with high consistency and strong robustness is formed to serve the integration of knowledge in multi-source urban construction project management.

[0021] 2. Utilizing the trained student model as a strongly optimized student model, deep semantic representations are generated for urban construction project management ontology entities, thereby enhancing the semantic representation capability of project ontology entities to the similarity matrix in the second stage. Based on this, a paired connected graph (PCG) of project ontology entities is first constructed, with nodes representing project entity concept pairs and edges representing project relationship pairs. Then, a four-layer similarity matrix of project ontology entities incorporating distillation knowledge is introduced, and features are extracted using CNN and transformed into dense vectors. Finally, residual GNNs with edge-aware attention are used to propagate features, combined with multi-dimensional label similarity, to avoid gradient vanishing. This method transforms complex project ontology entity matching into binary classification, providing support for the construction of a unified urban construction project management ontology library. Its accuracy and semantic understanding depth are significantly improved due to the introduction of distillation knowledge. This addresses the problems of high heterogeneity of multi-source knowledge, inconsistent quality, complex structural relationships, and the reliance on manual processes and difficulty in automatic updates in traditional knowledge fusion methods during urban construction project management modeling. It improves the accuracy of knowledge modeling in urban construction project management, thereby achieving automatic cross-system data association and business collaboration. When processing construction permits, project information can be automatically verified, improving approval efficiency. Meanwhile, this method replaces the high-cost and difficult-to-maintain traditional system integration method with automated matching rules, significantly reducing long-term operating costs. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an automatic fusion method based on knowledge distillation and graph neural networks according to the present invention.

[0023] Figure 2 This is a flowchart of the automatic knowledge fusion technology based on teacher knowledge correction, which is the subject of this invention. Detailed Implementation

[0024] An automatic fusion method based on knowledge distillation and graph neural networks, such as Figure 1 As shown, the present invention includes: S1. Obtain multi-source heterogeneous urban construction project data, input the data into the teacher model and student model, and extract the teacher Logit vector and student Logit vector respectively; Specifically, multi-source heterogeneous urban construction project data refers to data generated by different management departments and business processes in urban construction and management, which differ in source, format, and semantics. It includes four main categories: 1. Structured business data, such as project investment plans, bidding records, contract ledgers, and permit information; 2. Unstructured document and model data, such as feasibility study reports, design drawings, construction logs, and BIM models; 3. Spatiotemporal and IoT data, such as project GIS boundary lines, video surveillance streams, and equipment sensor data; 4. Experience and rule-based knowledge, such as historical problem-solving cases, expert consultation records, and industry standards. These data collectively constitute the input objects and knowledge sources for the methodology.

[0025] Specifically, urban construction project management data will be input into the teacher model. and student model Logit vectors are extracted from the penultimate layer of each of the two models. ,in Let C represent the Logit output for class m, and C represent the total number of classes. This is the last Logit vector. Logit serves as the foundation for structured probabilistic knowledge.

[0026] S2. Apply a softmax transformation to the teacher's Logit vector using a temperature coefficient to obtain the teacher's softening probability distribution. Apply a softmax transformation to the student's Logit vector using a temperature coefficient to obtain the student's softening probability distribution. Applying a softmax transform to the teacher's Logit vector using a temperature coefficient yields the teacher's softening probability distribution. Similarly, applying a softmax transform to the student's Logit vector using a temperature coefficient yields the student's softening probability distribution, expressed as follows:

[0027] in, For temperature coefficient, Let C be the Logit vector, and C be the total number of classes. Let m represent the j-th Logit vector, where m = T or S. When m = T, Let the teacher's Logit vector be... For the teacher's softened probability distribution, when m=S, Let the student's Logit vector be... Soften the probability distribution for students.

[0028] When T>1, the softmax function softens the labels. This softening makes the probability distribution more uniform, preventing the model from overconfidently predicting certain categories. By introducing a temperature hyperparameter for softening, the model pays more attention to subtle differences between categories, reflecting their underlying relationships. This provides more valuable information for the student model, helping it learn better. To facilitate the student model's learning from the teacher model, this invention uses the distance between their probability distributions to measure the KL divergence. A smaller KL divergence means that, under the guidance of the teacher model, the student model's output is closer to the teacher's.

[0029] S3. Calculate the Logit variance based on the teacher Logit vector, use the Logit variance to correct the teacher Logit vector to obtain the corrected vector, perform a softmax transformation on the corrected vector to obtain the corrected teacher softening probability distribution, calculate the KL divergence based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. The Logit variance is calculated based on the teacher's Logit vector, and the teacher's Logit vector is corrected using the Logit variance to obtain the corrected vector, which includes: Based on the teacher's Logit vector, obtain the average value of the teacher's Logit vector; The Logit variance is obtained from the mean of the teacher's Logit vector and the mean of the teacher's Logit vector. The teacher's Logit vector is corrected based on the Logit variance to obtain the correction vector.

[0030] Specifically, the teacher model may have biases in urban construction project management data. To ensure the quality of integrated knowledge, this invention proposes an adaptive error correction strategy based on Logit variance.

[0031] This invention uses the Adaptive Logit Reconstruction Knowledge Distillation (ALRKD) method for knowledge fusion. The overall process of the method is as follows: Figure 2As shown, given input urban construction project management data, Logit can be extracted from the penultimate layer of the teacher and student networks. Adaptive Logit reconstruction divides the knowledge transfer link into two parts: the first part is the correction of teacher errors (CorrectClass, CC), and the second part is the learning of relationships between non-target classes (OtherClasses, OC). For the first part, the correction of teacher errors, the variance of the output Logit is used as the correction gain. Variance, as a variable measuring the degree of fluctuation in a distribution, can help the teacher model improve its confidence in the correct class, thereby promoting error correction. For the second part, the learning of relationships between non-target classes, a selection strategy is adopted, not learning relationships between all non-target classes. Only classes with probabilities lower than the target class are learned, further extracting the correct inter-class relationships. The first stage of adaptive Logit reconstruction aims to correct the teacher network's erroneous predictions in the classification process. This invention proposes an adaptive correction strategy based on the sample's own characteristics for such erroneous samples. This method uses the standard deviation of the sample Logit as an adaptive gain to enhance the Logit corresponding to the correct class, thereby improving classification accuracy. Furthermore, standardizing the Logit helps minimize the difference between the correct class and other non-target classes, aligning the model's confidence with its accuracy and enhancing overall model calibration. The standard deviation can be represented as follows:

[0032] in, For Logit variance, The average value of the teacher's Logit vector. This represents the c-th element in all teacher Logit vectors, where c is the class index. This represents the total number of categories.

[0033] When the standard deviation of the sample is calculated, the teacher Logit vector can be corrected by adding the standard deviation to the Logit of the correct class, as follows:

[0034] in, For the correction vector, The mask for the correct category. Let be the teacher's Logi vector. This is used to assign a value of 1 to the position corresponding to the correct category, while assigning a value of 0 to all other positions corresponding to non-target categories.

[0035] Subsequently, a new probability distribution is obtained by applying the softmax function to the adjusted teacher Logit vector. The final loss function It can be represented as follows:

[0036]

[0037] in, For the first A corrected teacher Logit vector Let j be the corrected teacher Logit vector. This is the corrected probability distribution of teacher softening.

[0038] Teacher knowledge correction loss is represented as:

[0039] in, To correct knowledge loss for teachers express divergence, To correct the teacher softening probability distribution, Student softening probability distribution.

[0040] S4. Based on the teacher softening probability distribution, filter out non-target categories with probabilities lower than the correct category. Correct the student Logit vector and teacher Logit vector corresponding to the non-target category to obtain the corrected student Logit vector and corrected teacher Logit vector. Based on the corrected student Logit vector and teacher Logit vector, obtain the non-target category relationship extraction loss. Based on the teacher softening probability distribution, non-target categories with probabilities lower than the correct category are selected. The student and teacher Logit vectors corresponding to the non-target categories are corrected to obtain corrected student and teacher Logit vectors. Based on the corrected student and teacher Logit vectors, the non-target category relation extraction loss is obtained, including: Based on the teacher softening probability distribution, non-target categories with probabilities lower than the correct category are filtered to obtain the non-target category mask; Based on the non-target class mask, Logit variance, and preset coefficients, the Logit vectors corresponding to the non-target classes are modified to obtain the modified student Logit vector and the modified teacher Logit vector. By inputting the corrected student Logit vector and the corrected teacher Logit vector into the softmax function, the corrected teacher probability distribution and the corrected student probability distribution are obtained. The loss for extracting non-target category relationships is calculated based on the modified teacher probability distribution and the modified student probability distribution.

[0041] Specifically, in addition to target knowledge, urban construction project management data also contains a large amount of "inter-category relationship knowledge," such as the similarity between project constraints and the mutual influence between conflicting categories. This primarily targets relationships between non-target categories. These non-target categories are filtered, selecting those with classification probabilities lower than the target category as knowledge to be transferred. These categories with lower probability values ​​than the target category can be considered in the knowledge distillation process as not affecting classification accuracy but preserving the model's generalization ability. Therefore, strengthening the learning of these relationships between categories can improve the student model's generalization ability. To better emphasize the interrelationships between non-target categories, the probability values ​​of the remaining categories with probability values ​​greater than or equal to the target category are set to zero, thus preserving and emphasizing the relationships between the remaining non-target categories.

[0042] The corrected student Logit vector and the corrected teacher Logit vector are calculated as follows:

[0043] in, This represents the correction of the student's Logit vector and the correction of the teacher's Logit vector, when... When =T, To correct the teacher's Logit vector, when When =S, To correct the student's Logit vector, The Logit vectors consist of student Logit vectors and teacher Logit vectors. The student Logit vector is used for correction to obtain the corrected student Logit vector, and the teacher Logit vector is used for correction to obtain the corrected teacher Logit vector. For preset coefficients, This is a mask for classes whose probability values ​​are greater than or equal to the correct class. k is used to reduce the Logit of the outputs of those classes whose probability values ​​are greater than or equal to the correct class, thus ensuring that their probabilities can approach zero after undergoing the softmax transformation.

[0044] Unlike error correction, this operation is also applied to the output of the student model. The modified Logit is then fed into the softmax function to obtain a new probability distribution:

[0045]

[0046] in, To correct the teacher's Logit vector, To correct the student's Logit vector, Let j be the modified teacher's Logit vector. Let be the modified student's Logit vector. To correct the teacher probability distribution, To correct the student probability distribution.

[0047] This new probability distribution only includes the remaining classes whose classification probability is less than the target class probability, thus maintaining correctness while maximizing generalization to the student network. Finally, KL divergence is used to measure the difference between the two distributions. The loss function is as follows:

[0048] in, Extract loss for non-target category relationships.

[0049] S5. The total training loss is obtained based on the teacher knowledge correction loss, non-target category relationship extraction loss, and classification cross-entropy loss. The student model is then optimized based on the total training loss to obtain the optimized student model. Specifically, by combining the correct class error correction loss (LCC) and the non-target class relationship learning loss (LOC), we obtain the adaptive Logit reconstruction knowledge distillation loss:

[0050] in, and It's a hyperparameter.

[0051] Combine it with cross-entropy loss Combined, the total training loss is obtained. :

[0052]

[0053] in, For hyperparameters, The total number of samples, For the student model to the first The probability of a sample belonging to class c. For the first The true label of a sample, when the sample When the true category is c, It is 1 if it is true, otherwise it is 0.

[0054] In online knowledge distillation methods based on teacher-corrected knowledge, the final output of knowledge fusion is the final probability distribution p generated by the student model on the input data after training. S This fusion result is achieved through dual-path knowledge transfer (CC+OC) and dynamic loss optimization (…). )accomplish.

[0055] The output possesses three key characteristics. First, it exhibits high accuracy by adaptively correcting the classification bias of the teacher model through teacher error correction (CC), significantly improving the confidence of the correct category and ensuring the reliability of the classification results. Second, it demonstrates strong generalization ability by preserving the topological relationships between categories (such as similarity and conflict in project constraints) through non-target category relation learning (OC), enhancing the model's understanding of complex semantic associations. Finally, it contains structured knowledge; the output distribution not only includes category discrimination information but also deeply encodes the semantic associations between categories, thereby supporting fine-grained urban construction project decisions.

[0056] S6. Obtain the source city construction project management ontology and the target city construction project management ontology to be matched, and construct a pairwise connected graph of project ontology entities. In the connected graph, the nodes represent the project entity concept pairs between the source and target ontology, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is the source ontology, and the target city construction project management ontology is the target ontology. Specifically, a pairwise connectedivity graph (PCG) combines two directed graphs, establishing edges between two nodes that reflect their interactions. The construction of a pairwise connectedivity graph of project ontology entities allows graph neural network layers to effectively capture these interactions and learn efficient embeddings for each node pair to achieve binary classification (i.e., matching or not matching).

[0057] For two given city construction project management ontology O=(C,R,I) and O'=(C',R',I'), then PCG(O,O') is a construct of the form < The triples of R and Γ> are formed, where, R and Γ represent the set of nodes, the set of edges, and the set of edge types, respectively. Each element in R corresponds to a pair of item ontology entities between O and O′, and each element in R corresponds to a pair of item relations.

[0058] The method for constructing each edge is as follows:

[0059] Where x and x' represent the parent entities in the source project management ontology and the target project management ontology, respectively, y and y' represent the child entities, p represents the relationship between x and y, and p' represents the relationship between x' and y'.

[0060] S7. Construct a basic similarity matrix based on the attributes of entities in the source ontology and the target ontology, and input the entities of the source ontology into the student model to obtain deep semantic features. Obtain an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features. A basic similarity matrix is ​​constructed based on the attributes of entities in the source and target ontologies. Entities from the source ontology are then input into the student model to obtain deep semantic features. An enhanced similarity matrix is ​​then derived based on the basic similarity matrix and the deep semantic features, including: Based on the data attributes and entity attributes of the source and target ontologies, a basic similarity matrix is ​​calculated. The entity input from the source ontology is used to optimize the student model to obtain deep semantic features; The cosine similarity is calculated based on the deep semantic features of the source and target ontologies. An enhanced similarity matrix is ​​constructed based on the basic similarity matrix and the cosine similarity.

[0061] Specifically, a similarity matrix for project ontology entity pairs is introduced as the feature vector for project ontology entity pairs. Given a project ontology entity pair < , >, among which ∈O, ∈O', the rows and columns of the matrix include and The data properties include project code, project name, construction address, total cost, total construction period, start date, completion date, structure type, and building area; and object properties include construction unit, construction company, supervision unit, design unit, surveying unit, project manager, chief supervising engineer, contract section, building, work process, equipment, and documents. It also includes features such as project tags, annotations, disjoint classes, equivalence classes, subclasses, and parent classes. (o) and (o') represent the i-th feature values ​​of o and o', respectively. To capture the similarity between two item ontology entities o and o', an enhanced similarity matrix M for a single-layer item ontology entity pair is calculated. m×n Elements in M express (o) and The similarity calculation result of (o'). The feature values ​​of project ontology entities may be of various types, such as data, time, floating-point, integer, and string. To maintain simplicity and effectiveness, this invention treats all feature values ​​as text strings. To more comprehensively extract the feature correlations between two project ontology entities, this invention designs a four-layer basic similarity matrix and adds a deep semantic similarity layer based on the first-stage student model. The specific similarity algorithm selection is shown in the table below:

[0062] The newly added deep semantic similarity layer involves inputting the entity elements of the project ontology to be compared into the student model trained in the first part. In the process, the output vector (Logits) of the penultimate layer is extracted and denoted as follows: = ( )and = ( Subsequently, the cosine similarity between the two feature vectors is calculated as a new semantic similarity layer. This operation can be represented as:

[0063] This layer's similarity, together with the original edit distance, Jaccard similarity, and other layers, forms a richer multi-level similarity matrix for subsequent CNN feature extraction.

[0064] S8. Input the enhanced similarity matrix into the convolutional neural network to obtain low-dimensional dense features. Concatenate the low-dimensional dense features with the entity label similarity features to obtain the initial node features. Input the initial node features into the graph neural network to complete feature propagation and aggregation on the paired connected graph to obtain the node depth representation. The entity label similarity is calculated through the item entity concept pairs. Specifically, typically, an item ontology entity is described by only a small number of attributes and instances in the knowledge graph; that is, most attribute values ​​are empty for an item ontology entity. Therefore, the feature similarity matrix between two item ontology entities is usually a sparse matrix, with a large portion being 0. Furthermore, the similarity between some attributes may be useless or meaningless for detecting permutations and combinations. To automatically find useful similarity patterns in attribute values, this invention proposes an attribute feature extraction method based on convolutional neural networks. This method not only avoids the tedious work of manually extracting features and automatically obtains useful similarity features of item ontology entity pairs, but also encodes the sparse similarity matrix into a short and dense vector, obtaining the ideal node input vector for item ontology entity pairs. The invention uses the enhanced similarity matrix M of the item ontology entity pairs as input to the CNN network, ultimately outputting a low-dimensional dense feature vector X. This vector is concatenated with the label similarity feature vector of the item ontology entity name, and together they serve as the initial node feature input for the graph neural network (GNN).

[0065] In an Ontology Entity Pair Connected Graph (PCG), directed edges connecting nodes represent both the relationships between nodes and structural information. By propagating the attribute features of project ontology entity pairs along these edges, feature representations of project ontology entity pairs with neighbor node information are obtained. Therefore, this invention uses a Graph Neural Network (GNN) to propagate the attribute features of project ontology entity pairs on the PCG. By recursively aggregating the feature vectors of their neighbors to learn node representations in the graph, the GNN can combine node features and structural information from the graph.

[0066] Entity tag similarity calculation methods include: Based on the label names of the source and target ontologies, calculate character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance normalization; The entity label similarity is obtained by concatenating the normalized values ​​of character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance.

[0067] Specifically, the label or name of an entity in an ontology is often considered an important clue for predicting whether two entities match. Therefore, this invention treats the label attribute as a special attribute, calculating a separate name similarity vector, which is concatenated with the similarity vector X generated by the CNN model and used as input to the graph neural network model. To capture the similarity features of entity names from different perspectives, this invention uses several string-based similarity metrics, which are widely used in traditional similarity-based alignment methods. This invention uses the following similarity measures: ① String equality: Used to evaluate whether two strings are equal, by calculating characteristic values. .

[0068] ② Edit Distance (Levenshtein Distance): Also known as Levenshtein distance, it's a string metric that calculates the degree of difference between two strings. Levenshtein distance can be considered the minimum number of edits (such as modification, insertion, or deletion) required to move from one string to another. It's calculated by obtaining eigenvalues. .

[0069] ③ Edit Similarity: Represents the edit similarity between two strings. It is calculated using feature values. .

[0070] ④ Jaccard similarity: Also known as the Jaccard similarity coefficient, it is used to compare the similarity and differences between a finite set of samples. It is obtained by calculating eigenvalues. .

[0071] ⑤ Substring Similarity: This is calculated by finding the longest common substring between two strings. The characteristic value is obtained through calculation. .

[0072] ⑥ Normalized EditDistance: In some project ontology entities, the local names of elements are represented in the form of IDs, such as project codes, building numbers, process IDs, etc., which are meaningless. Therefore, this invention first simply obtains the mapping result by comparing the tag sets of element pairs, and uses the normalized edit distance similarity index to calculate the language similarity between tag sets. The formula is as follows. The feature value SIMname is obtained through calculation.

[0073] The entity-to-tag similarity feature vector of the project ontology is represented as a concatenated vector calculated from the above features:

[0074] It will be concatenated with the similarity vector X generated by the CNN to form the initial feature vector H of the project ontology entity pair, which will then serve as the input to the graph neural network model.

[0075] The feature propagation-based graph neural network model is a residual graph neural network model with edge-aware attention. Its attention mechanism is constructed by improving the GAT model. The model has two layers, each taking the features of the initial nodes. As input, where ∈R F N is the number of nodes in the PCG graph, and F is the dimension of the input features. Each layer generates a new feature vector as the input to the next layer, calculated as follows:

[0076] in, For node depth representation, For activation function, For attention, For shared matrices, These are the initial node features.

[0077] In the classic GAT model, attention The attention is calculated based on the features of nodes i and j. However, in the task of entity matching on the project ontology, the edge type between two nodes is important and should not be ignored. Therefore, this invention uses an edge-aware attention mechanism to calculate the attention. First, an attention coefficient needs to be calculated using a shared attention mechanism. The formula is as follows:

[0078] in ∈R F′ ( That is, with the edge<i,j> Type corresponding (Vector) represents the edge type from node i to node j. 'a' is a vector representing an edge of a single-layer feedforward neural network used to compute attention coefficients, and || denotes concatenation of vectors. Edge Type The vector is the element-wise absolute difference between the average vectors of the source and target nodes connected by the edge. Its calculation formula is as follows:

[0079] in, and These are the sets of node indices for outgoing and incoming edges of type k in the input paired connected graph. This indicates the number of elements in the set.

[0080] The attention coefficients are obtained according to the above formula, and then the attention is calculated by normalizing all the coefficients of its neighboring nodes using the softmax function. The specific formula is as follows.

[0081]

[0082] S9. Perform binary classification matching prediction based on node depth representation, output the entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

[0083] After feature propagation through the Graph Neural Network (GNN) layers, the model outputs the final vector representation of each node in the Pairwise Connected Graph (PCG) of item ontology entities. Thus, the final task of knowledge fusion is formalized as a binary classification problem based on node representations: for each node in the graph, the model needs to determine whether its corresponding pair of item ontology entities are semantically equivalent, i.e., outputting a label of 1 (match) or 0 (mismatch). The transformation process from the node vectors output by the GNN to the final binary classification result follows three clear steps: (1) Generation of final node representations After multiple rounds of iterative feature propagation and aggregation through a GNN layer based on edge-aware attention and residual connections, the model learns a deep representation vector for each node i in the graph that deeply integrates multiple types of information, denoted as . (i.e., the output of the last layer) ).

[0084] It is the final feature vector of the i-th node, which integrates automatic features from CNN, manually calculated label similarity features, and neighbor information aggregated from the PCG graph structure.

[0085] (2) Calculation of confidence score The final representation vector of the node Given a prediction layer (typically a combination of a fully connected layer and a non-linear activation function), calculate the confidence score that the candidate match belongs to the "match" category. This process is defined by the following formula:

[0086] in is the matching confidence score of the candidate matching pair corresponding to node i. It is a scalar with a value range between (0, 1). The closer the score is to 1, the higher the confidence of the model in determining that the pair is a match. It is the trainable weight matrix of the prediction layer. It is the trainable bias term of the prediction layer, and it is a scalar. ( ) is the Sigmoid activation function, which maps the result of a linear transformation to a probability value between 0 and 1, which serves as a confidence score.

[0087] (3) Binary classification decision Obtain the confidence score for each node. Then, a decision is made based on a preset classification threshold τ, and the final binary classification label is output:

[0088] in It is the final predicted label of node i, with a value of 0 (no match) or 1 (match). This is the classification decision threshold, a hyperparameter, typically set to 0.5 by default. All nodes with confidence scores that reach or exceed this threshold are considered "matches," otherwise they are considered "mismatches."

[0089] (4) Matching the mapping set After the model finishes running, all those predicted as "matches" (i.e. The nodes with a value of 1 together constitute the final result of this automatic knowledge fusion task, namely the matching mapping set. ={ | =< , relation, confidence , )>}.

[0090] in It is the k-th matching mapping in the set. , A pair of entities (or attributes) that originate from the source project management ontology and the target project management ontology and are determined to be semantically equivalent. The relation specifies... and The semantic relationship between them is usually an equivalence relationship in project ontology entity matching. That is, the confidence score calculated in the aforementioned steps. This serves as a reliable metric for this matching mapping.

[0091] Should The set directly provides core equivalence relation assertions for building a high-quality, unified urban construction project management ontology, and is the key output for completing the automatic conversion from multi-source heterogeneous knowledge to structured and integrated knowledge.

[0092] An automatic fusion system based on knowledge distillation and graph neural networks includes: The first acquisition module is used to acquire multi-source heterogeneous urban construction project data, input the multi-source heterogeneous urban construction project data into the teacher model and the student model, and extract the teacher Logit vector and the student Logit vector respectively. The probability distribution calculation module is used to perform a softmax transformation on the teacher's Logit vector using a temperature coefficient to obtain the teacher's softening probability distribution, and to perform a softmax transformation on the student's Logit vector using a temperature coefficient to obtain the student's softening probability distribution. The teacher knowledge correction loss calculation module is used to calculate the Logit variance based on the teacher Logit vector, correct the teacher Logit vector using the Logit variance to obtain a corrected vector, perform a softmax transformation on the corrected vector to obtain the corrected teacher softening probability distribution, and calculate the KL divergence based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. The non-target category relation extraction loss calculation module is used to filter non-target categories with probabilities lower than the correct category based on the teacher softening probability distribution, correct the student Logit vector and teacher Logit vector corresponding to the non-target category respectively to obtain corrected student Logit vector and corrected teacher Logit vector, and obtain the non-target category relation extraction loss based on the corrected student Logit vector and teacher Logit vector. The optimization module is used to obtain the total training loss based on the teacher knowledge correction loss, the non-target category relationship extraction loss, and the classification cross-entropy loss, and to optimize the student model based on the total training loss to obtain an optimized student model. The second acquisition module is used to acquire the source city construction project management ontology and the target city construction project management ontology to be matched, and to construct a pairwise connected graph of project ontology entities. The nodes of the connected graph represent project entity concept pairs between the source and target ontology, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is the source ontology, and the target city construction project management ontology is the target ontology. An enhanced similarity matrix construction module is used to construct a basic similarity matrix based on the attributes of entities in the source ontology and the target ontology, input the entities of the source ontology into the optimized student model to obtain deep semantic features, and obtain an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features. The feature calculation module is used to input the enhanced similarity matrix into a convolutional neural network to obtain low-dimensional dense features, concatenate the low-dimensional dense features with entity label similarity features to obtain initial node features, input the initial node features into a graph neural network, complete feature propagation and aggregation on a pairwise connected graph to obtain node depth representation, and the entity label similarity is calculated through item entity concept pairs; The fusion module is used to perform binary classification matching prediction based on the node depth representation, output an entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

[0093] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs an automatic fusion method based on knowledge distillation and graph neural networks.

[0094] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0095] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0096] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0097] In this terminal device, the automatic fusion method based on knowledge distillation and graph neural network in the above embodiments is stored in the memory of the terminal device and loaded and executed on the processor of the terminal device for convenient use.

[0098] This application also discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it employs the automatic fusion method based on knowledge distillation and graph neural networks described in the above embodiments.

[0099] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0100] The automatic fusion method based on knowledge distillation and graph neural networks in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0101] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0102] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. An automatic fusion method based on knowledge distillation and graph neural networks, characterized in that, include: Obtain multi-source heterogeneous urban construction project data, input the multi-source heterogeneous urban construction project data into the teacher model and student model, and extract the teacher Logit vector and student Logit vector respectively; The teacher's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the teacher's softening probability distribution. The student's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the student's softening probability distribution. The Logit variance is calculated based on the teacher's Logit vector. The Logit variance is then used to correct the teacher's Logit vector to obtain a corrected vector. A softmax transformation is performed on the corrected vector to obtain the corrected teacher softening probability distribution. The KL divergence is calculated based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. Based on the teacher softening probability distribution, non-target categories with a probability lower than the correct category are filtered out. The student Logit vector and teacher Logit vector corresponding to the non-target category are corrected respectively to obtain the corrected student Logit vector and the corrected teacher Logit vector. Based on the corrected student Logit vector and the corrected teacher Logit vector, the non-target category relationship extraction loss is obtained. The total training loss is obtained based on the teacher knowledge correction loss, the non-target category relationship extraction loss, and the classification cross-entropy loss. The student model is then optimized based on the total training loss to obtain the optimized student model. Obtain the source city construction project management ontology and the target city construction project management ontology to be matched, and construct a pairwise connected graph of project ontology entities. The nodes of the connected graph represent project entity concept pairs between the source and target ontologies, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is used as the source ontology, and the target city construction project management ontology is used as the target ontology. A basic similarity matrix is ​​constructed based on the attributes of entities in the source ontology and the target ontology. The entities of the source ontology are input into the optimized student model to obtain deep semantic features. An enhanced similarity matrix is ​​obtained based on the basic similarity matrix and the deep semantic features. The enhanced similarity matrix is ​​input into a convolutional neural network to obtain low-dimensional dense features. The low-dimensional dense features are concatenated with entity label similarity features to obtain initial node features. The initial node features are input into a graph neural network to complete feature propagation and aggregation on a pairwise connected graph to obtain node depth representation. The entity label similarity is calculated through item entity concept pairs. Based on the node depth representation, perform binary classification matching prediction, output the entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

2. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, The teacher's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the teacher's softening probability distribution. Similarly, the student's Logit vector is subjected to a softmax transformation using a temperature coefficient to obtain the student's softening probability distribution, expressed as follows: where, is the temperature coefficient, is the Logit vector, C is the total number of classes, denotes the j-th Logit vector, m = T or S, when m = T, is the teacher Logit vector, is the teacher softened probability distribution, when m = S, is the student Logit vector, is the student softened probability distribution.

3. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, The Logit variance is calculated based on the teacher's Logit vector, and the teacher's Logit vector is corrected using the Logit variance to obtain a corrected vector, including: Based on the teacher Logit vector, the average value of the teacher Logit vector is obtained; Based on the average value of the teacher's Logit vector and the average value of the teacher's Logit vector, the Logit variance is obtained; The teacher's Logit vector is corrected based on the Logit variance to obtain the corrected vector.

4. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, The process involves filtering non-target categories with probabilities lower than the correct category based on the teacher softening probability distribution, and then correcting the student and teacher Logit vectors corresponding to the non-target categories to obtain corrected student and teacher Logit vectors. Based on these corrected student and teacher Logit vectors, the non-target category relationship extraction loss is obtained, including: Based on the teacher softening probability distribution, non-target categories with probabilities lower than the correct category are filtered to obtain a non-target category mask; Based on the non-target class mask, the Logit variance, and the preset coefficients, the Logit vectors corresponding to the non-target classes are modified to obtain the modified student Logit vector and the modified teacher Logit vector. The modified student Logit vector and the modified teacher Logit vector are input into the softmax function to obtain the modified teacher probability distribution and the modified student probability distribution; Based on the modified teacher probability distribution and the modified student probability distribution, the non-target category relationship extraction loss is calculated.

5. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, The process involves constructing a basic similarity matrix based on the attributes of entities in the source and target ontologies, inputting entities from the source ontology into the optimized student model to obtain deep semantic features, and then obtaining an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features, including: Based on the data attributes and entity attributes of the source ontology and the target ontology, a basic similarity matrix is ​​calculated; The entities of the source ontology are input into the optimized student model to obtain deep semantic features; The cosine similarity is calculated based on the deep semantic features of the source and target ontologies. An enhanced similarity matrix is ​​constructed based on the basic similarity matrix and the cosine similarity.

6. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, Entity tag similarity calculation methods include: Based on the label names of the source and target ontologies, calculate character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance normalization; The character equivalence, edit distance, edit similarity, Jaccard similarity, substring similarity, and edit distance are normalized and concatenated to obtain the entity label similarity.

7. The automatic fusion method based on knowledge distillation and graph neural networks according to claim 1, characterized in that, The initial node features are input into the graph neural network, and feature propagation and aggregation are performed on the paired connected graph to obtain the node depth representation, which is represented as: in, For node depth representation, For activation function, For attention, For shared matrices, These are the initial node features.

8. An automatic fusion system based on knowledge distillation and graph neural networks, characterized in that, include: The first acquisition module is used to acquire multi-source heterogeneous urban construction project data, input the multi-source heterogeneous urban construction project data into the teacher model and the student model, and extract the teacher Logit vector and the student Logit vector respectively. The probability distribution calculation module is used to perform a softmax transformation on the teacher's Logit vector using a temperature coefficient to obtain the teacher's softening probability distribution, and to perform a softmax transformation on the student's Logit vector using a temperature coefficient to obtain the student's softening probability distribution. The teacher knowledge correction loss calculation module is used to calculate the Logit variance based on the teacher Logit vector, correct the teacher Logit vector using the Logit variance to obtain a corrected vector, perform a softmax transformation on the corrected vector to obtain the corrected teacher softening probability distribution, and calculate the KL divergence based on the corrected teacher softening probability distribution and the student softening probability distribution to obtain the teacher knowledge correction loss. The non-target category relation extraction loss calculation module is used to filter non-target categories with probabilities lower than the correct category based on the teacher softening probability distribution, correct the student Logit vector and teacher Logit vector corresponding to the non-target category respectively to obtain corrected student Logit vector and corrected teacher Logit vector, and obtain the non-target category relation extraction loss based on the corrected student Logit vector and teacher Logit vector. The optimization module is used to obtain the total training loss based on the teacher knowledge correction loss, the non-target category relationship extraction loss, and the classification cross-entropy loss, and to optimize the student model based on the total training loss to obtain an optimized student model. The second acquisition module is used to acquire the source city construction project management ontology and the target city construction project management ontology to be matched, and to construct a pairwise connected graph of project ontology entities. The nodes of the connected graph represent project entity concept pairs between the source and target ontology, and the edges represent project relationship pairs. The source city construction project management ontology to be matched is the source ontology, and the target city construction project management ontology is the target ontology. An enhanced similarity matrix construction module is used to construct a basic similarity matrix based on the attributes of entities in the source ontology and the target ontology, input the entities of the source ontology into the optimized student model to obtain deep semantic features, and obtain an enhanced similarity matrix based on the basic similarity matrix and the deep semantic features. The feature calculation module is used to input the enhanced similarity matrix into a convolutional neural network to obtain low-dimensional dense features, concatenate the low-dimensional dense features with entity label similarity features to obtain initial node features, input the initial node features into a graph neural network, complete feature propagation and aggregation on a pairwise connected graph to obtain node depth representation, and the entity label similarity is calculated through item entity concept pairs; The fusion module is used to perform binary classification matching prediction based on the node depth representation, output an entity matching mapping set between the source ontology and the target ontology, and fuse the source ontology and the target ontology according to the entity matching mapping set.

9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.

Images