An AI-based business object modeling method in a face-to-object, medium and system

Abstract

The application provides an AI-based object-oriented business object modeling method, medium and system, belongs to the technical field of AI-based object-oriented business object modeling methods, media and systems, and solves the technical problem of low business object recognition accuracy caused by low-frequency long-tail distribution of professional terms in natural language text of industrial business requirements by receiving industrial business requirement text, using natural language processing technology and an industrial vertical field continuous pre-training corpus to complete standardized mapping, inputting standardized text into a hierarchical gated knowledge injection double-flow encoding model to identify core business objects, attributes and classification labels, and then automatically generating a business object model of a unified modeling language specification through an artificial intelligence model combined with a Monte Carlo tree search heuristic sampling strategy and a VF2++ graph matching algorithm, and finally calculating a field bias comprehensive score according to verification feedback of business personnel and adaptively fine-tuning model parameters.

Description

Technical Field

[0001] This invention belongs to the technical field of AI-based object-oriented business object modeling methods, media and systems. Specifically, it relates to an AI-based object-oriented business object modeling method, media and system. Background Technology

[0002] In the field of industrial software engineering, object-oriented modeling is a core step in transforming industrial business requirements into executable system designs. Traditional business object modeling methods rely on business analysts manually reading requirement texts, identifying objects, attributes, and relationships one by one using a unified modeling language specification, and selecting modeling templates based on industry experience to complete model construction. With the deepening of the Industrial Internet and intelligent manufacturing, the scale of business requirement texts in fields such as machinery manufacturing, chemical engineering, and metallurgy continues to expand. The professional terms involved in these texts are highly vertical and have a low-frequency, long-tail distribution, making traditional manual modeling methods inefficient. The industry has begun to introduce natural language processing and pre-trained language model technologies to automate the extraction of business objects and is attempting to use knowledge graphs to assist entity recognition to compensate for the shortcomings of general language models in understanding the semantics of industrial terms.

[0003] However, in existing technologies, general pre-trained language models cannot cover the real semantic space for vector representations of industrial terms. When knowledge graph features and text features are fused, there is a semantic space misalignment problem, which makes it impossible for the knowledge injection strength to adaptively adjust with the text context. Low-quality knowledge interferes with the normal encoding process, ultimately causing a serious decline in the accuracy of business object recognition for low-frequency long-tail industrial terms.

[0004] In other words, existing technologies suffer from the technical problem of low accuracy in identifying business objects due to the low-frequency, long-tail distribution of technical terms in natural language texts used for industrial business needs. Summary of the Invention

[0005] In view of this, the present invention provides an AI-based method, medium, and system for modeling business objects in object-oriented programming, which can solve the technical problem in the prior art where the low-frequency long-tail distribution of professional terms in natural language texts for industrial business needs leads to low accuracy in identifying business objects.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides an AI-based method for modeling business objects in object-oriented programming, comprising the following steps:

[0007] The system receives industrial business requirement texts in natural language, performs cleaning, word segmentation, part-of-speech tagging and named entity recognition through natural language processing technology, and standardizes the mapping of professional terms by combining a continuously pre-trained corpus of the industrial vertical field to obtain standardized requirement text data.

[0008] Standardized requirement text data is input into a hierarchical gating knowledge injection dual-stream coding model, which outputs the object confidence score, business object classification label, and object attribute label for each business object. The weight of the graph flow knowledge injection is adjusted according to the confidence interval rule based on the object confidence score to complete the identification of core business objects, object attributes, and business object classification labels.

[0009] A weighted adjacency matrix is ​​constructed based on the identified core business objects and their classification labels. The weighted adjacency matrix is ​​then subjected to Laplacian matrix spectral decomposition. A low-rank sparse approximation graph is extracted based on the spectral energy retention threshold. The semantic similarity between objects is calculated on the low-rank sparse approximation graph using a random walk algorithm. The semantic similarity between objects is then concatenated with the low-frequency feature vectors of the corresponding nodes and input into a multilayer perceptron to complete the relationship type discrimination and output a precise business object relationship network.

[0010] The industrial scenario type is determined based on the business object classification tags. The corresponding modeling template library is called, and the accurate business object relationship network and modeling template graph are constructed into an attribute graph. The VF2++ graph matching algorithm combined with the pruning strategy is used to reduce the search space. The Monte Carlo tree search heuristic sampling strategy is introduced to prioritize the exploration of high-confidence template paths and automatically generate a business object model that conforms to the unified modeling language specification.

[0011] The business object model is submitted to business personnel for verification. Based on the number of missing objects, attribute errors, and relationship misjudgments in the verification feedback, a comprehensive domain deviation score is calculated. When the comprehensive domain deviation score falls into different intervals, different learning rate strategies are used to fine-tune the hierarchical gating knowledge injection dual-stream coding model. The model parameters are updated until the business object model meets the business requirements.

[0012] The industrial vertical domain continuous pre-training corpus consists of three types of data: industrial patent documents, industry standard documents, and equipment operation manuals. A domain-adaptive continuous pre-training strategy is used to perform secondary training on the basic pre-training model, and a domain-specific terminology vocabulary is dynamically injected into the word embedding layer through a vocabulary enhancement method.

[0013] The vocabulary enhancement method refers to injecting newly added domain terms into the model word embedding matrix in a vector initialization manner. The initialization vector is obtained by statistical calculation of the co-occurrence of the term in the context of the continuously pre-trained corpus of the industrial vertical domain.

[0014] The hierarchical gating knowledge injection dual-stream coding model adopts a dual-stream parallel coding architecture. The text stream processes standardized requirement text data based on an improved RoBERTa encoder, while the graph stream performs hierarchical aggregation of multi-hop neighbors of the industrial knowledge graph based on a graph attention network. The first hop extracts attribute knowledge, the second hop extracts relational knowledge, and the third hop extracts scene knowledge.

[0015] The text stream uses the Longformer architecture with sparse attention mechanism instead of the standard Transformer. It captures long-distance dependencies across sentences through a hybrid strategy of sliding window attention and global attention, supporting the processing of ultra-long texts with more than 4096 tags.

[0016] The text stream and the graph stream are injected with cross-stream information through a dynamic gating fusion mechanism. The fusion gate weight of the dynamic gating fusion mechanism is dynamically calculated by the confidence score of the currently labeled object. An adaptive skip connection is introduced inside the network. When the dependence of a certain layer of the text stream on the graph features is lower than the dependence threshold, the graph injection step is skipped and the text stream features are directly transmitted.

[0017] The hierarchical gating knowledge injection dual-stream coding model fusion contrastive learning technique applies contrastive loss to the text stream representation and graph stream representation of the same entity for pre-alignment, narrowing the cosine distance between the two and ensuring the consistency of the dual-stream semantic space; the output layer fuses conditional random field decoding to complete entity sequence labeling and output object confidence scores.

[0018] The specific rule for adjusting the graph flow knowledge injection weights based on the object confidence score according to the confidence interval rule is as follows: Let the object confidence score be... ,when The knowledge injection weight of the time-mapping graph is set to 0.2. The knowledge injection weight of the time-based graph flow is set to 0.5. The weight for knowledge injection into the time-map flow is set to 0.8.

[0019] The weighted adjacency matrix is ​​constructed by using core business objects as nodes, candidate pairs of associations determined by the classification labels of business objects as edges, and the semantic similarity between objects of the candidate pairs as edge weights; the Laplacian matrix spectral decomposition extracts low-frequency feature vectors corresponding to smaller eigenvalues ​​as the global structural representation of the graph.

[0020] Wherein, the spectral energy retention threshold refers to the threshold before retention. Each eigenvalue corresponds to a threshold representing the proportion of energy to the total energy of all eigenvalues. The low-rank sparse approximation graph retains only the first... The graph structure corresponding to each low-frequency feature vector is obtained by removing low-weight redundant relation edges, and the reference range for the spectral energy retention threshold is 70% to 99%.

[0021] The random walk algorithm starts with each core business object node on a low-rank sparse approximation graph and iterates through several steps according to the edge weight probability. The co-occurrence frequency between nodes is used as the semantic similarity between objects. Core business object pairs with higher semantic similarity between objects are given higher priority as association candidates in relation type discrimination.

[0022] The VF2++ graph matching algorithm introduces a candidate node sorting strategy and a backtracking pruning mechanism. During the recursive search process, it prioritizes expanding the node with the highest degree and the strongest constraints. During the matching process, it also verifies the consistency between node attributes and edge attributes and skips subtrees that do not meet the constraints by combining the pruning strategy.

[0023] The Monte Carlo tree search heuristic sampling strategy constructs a search tree iteratively through four stages: selection, expansion, simulation, and backpropagation. It uses the average confidence score of existing matching results as a heuristic signal to prioritize exploring paths with historically high object confidence scores.

[0024] The formula for calculating the comprehensive score of the domain deviation is as follows: ,in The overall score is based on the domain deviation. For the number of missing objects, For the total number of objects, For the number of attribute errors, For the total number of attributes, For the number of misjudged relationships, For the total number of relations, , , The weighted coefficients are and satisfy the following conditions: .

[0025] Specifically, the method of using different learning rate strategies to fine-tune the hierarchical gated knowledge injection dual-stream coding model when the comprehensive score of domain deviation falls into different intervals is as follows: when Learning rate Perform full parameter fine-tuning, when Learning rate Fine-tune the top-level parameters when Learning rate Only the output layer parameters are fine-tuned.

[0026] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described AI-based object-oriented business object modeling method.

[0027] A third aspect of the present invention provides an AI-based object-oriented business object modeling system, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0028] This invention proposes an artificial intelligence-based object-oriented modeling method for business objects. It constructs a continuously pre-trained corpus for industrial vertical domains and combines lexical enhancement methods to expand the vector representation of low-frequency long-tail terms. At the same time, it adopts a hierarchical gated knowledge injection dual-stream coding model to pre-align the text stream and graph stream through comparative loss. Then, through a dynamic gated fusion mechanism and adaptive skip connections, it realizes context-adaptive adjustment of knowledge injection intensity and low-quality knowledge filtering. This fundamentally solves the two intrinsic factors that restrict accuracy: misalignment of the two-stream semantic space and interference from low-quality knowledge.

[0029] This invention employs a domain-adaptive continuous pre-training strategy to enable the model to adaptively shift to the distribution of industrial language. A vocabulary augmentation method injects low-frequency term vectors into the word embedding matrix, ensuring that long-tail terms have semantically distinguishable vector representations. Contrastive loss pre-alignment eliminates the semantic gap between the text stream and the graph stream. A dynamic gating fusion mechanism adaptively adjusts the graph knowledge injection weights based on the object confidence score. Adaptive skip connections directly bypass the knowledge injection path when the graph dependency is low, thereby significantly improving the accuracy and robustness of business object recognition in low-frequency long-tail industrial term scenarios.

[0030] In summary, this invention solves the technical problem mentioned in the background art of low accuracy in identifying business objects due to the low-frequency, long-tail distribution of technical terms in natural language text for industrial business needs. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention.

[0032] Figure 2 A bar chart showing the distribution of the number of five types of relationships in the precise business object relationship network.

[0033] Figure 3 A weight distribution diagram of different confidence paths during the Monte Carlo tree search template matching process. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0035] like Figure 1 The diagram shown is a flowchart of a business object modeling method based on AI in object-oriented programming, provided by the first aspect of this invention. This method includes the following steps:

[0036] S01. Receive natural language industrial business requirement text, and complete cleaning, word segmentation, part-of-speech tagging and named entity recognition through natural language processing technology. Combine the continuous pre-trained corpus of the industrial vertical field to standardize the mapping of professional terms and obtain standardized requirement text data.

[0037] S02. Input standardized requirement text data into a hierarchical gating knowledge injection dual-stream coding model, output the object confidence score, business object classification label and object attribute label for each business object, adjust the graph flow knowledge injection weight according to the confidence interval rule based on the object confidence score, and complete the identification of core business objects, object attributes and business object classification labels;

[0038] S03. Based on the identified core business objects and business object classification labels, construct a weighted adjacency matrix, perform Laplace matrix spectral decomposition on the weighted adjacency matrix, extract a low-rank sparse approximation graph according to the spectral energy retention threshold, calculate the semantic similarity between objects on the low-rank sparse approximation graph through a random walk algorithm, concatenate the semantic similarity between objects with the low-frequency feature vector of the corresponding node, and input the concatenation into a multilayer perceptron to complete the relationship type discrimination, and output an accurate business object relationship network.

[0039] S04. Determine the industrial scenario type based on the business object classification tags, call the corresponding modeling template library, construct the accurate business object relationship network and modeling template graph into an attribute graph, use the VF2++ graph matching algorithm combined with pruning strategy to reduce the search space, introduce the Monte Carlo tree search heuristic sampling strategy to prioritize the exploration of high confidence template paths, and automatically generate a business object model that conforms to the unified modeling language specification.

[0040] S05. Submit the business object model to the business personnel for verification. Calculate the domain deviation comprehensive score based on the number of missing objects, attribute errors, and relationship misjudgments in the verification feedback. When the domain deviation comprehensive score falls into different intervals, use different learning rate strategies to fine-tune the hierarchical gating knowledge injection dual-stream coding model and update the model parameters until the business object model meets the business requirements.

[0041] The industrial vertical domain continuous pre-training corpus consists of three types of data: industrial patent documents, industry standard documents, and equipment operation manuals. A domain-adaptive continuous pre-training strategy is used to perform secondary training on the basic pre-trained model. A domain-specific terminology lexicon is dynamically injected into the word embedding layer using a lexical enhancement method, enabling the model's vector representation of low-frequency, long-tail terminology to cover the real semantic space. The domain-adaptive continuous pre-training strategy refers to continuously updating the model parameters using the industrial vertical domain continuous pre-training corpus on top of the general pre-trained model, causing the model to adaptively shift to the language distribution in the industrial domain, thereby improving its semantic encoding ability for industrial terminology. The lexical enhancement method involves injecting newly added domain terms into the model's word embedding matrix using vector initialization. The initialization vectors are obtained by statistically calculating the co-occurrence of the terms in the context of the industrial vertical domain continuous pre-training corpus, allowing the model's lexicon to expand its representation ability for terminology without destroying the original semantic structure.

[0042] The hierarchical gated knowledge injection dual-stream coding model employs a dual-stream parallel coding architecture: the text stream processes standardized requirement text data based on an improved RoBERTa encoder, using a Longformer architecture with sparse attention mechanisms instead of the standard Transformer, supporting ultra-long text processing with over 4096 tags, and capturing long-distance dependencies across sentences through a hybrid strategy of sliding window attention and global attention; the graph stream uses a graph attention network to hierarchically aggregate multi-hop neighbors of the industrial knowledge graph, extracting attribute knowledge in the first hop, relational knowledge in the second hop, and scene knowledge in the third hop; cross-stream information injection between the text stream and the graph stream is achieved through a dynamic gated fusion mechanism, where the fusion gate weights are dynamically calculated based on the current tag's object confidence score, replacing simple concatenation; adaptive skip connections are introduced within the network, where when a certain layer of the text stream... When the dependence on graph features is lower than the dependence threshold, the graph injection step is skipped, and text stream features are directly transmitted to avoid interference from low-quality knowledge. The output layer fuses conditional random field decoding to complete entity sequence labeling and output object confidence scores. The network fuses contrastive learning technology to apply contrastive loss to the text stream representation and graph stream representation of the same entity for pre-alignment, narrowing the cosine distance between the two and ensuring consistency of the two-stream semantic space. The hierarchical gated knowledge injection two-stream coding model realizes dynamic fusion of symbol space and vector space through parallel two-stream coding. Contrast loss pre-alignment eliminates the semantic gap between the two feature spaces. The dynamic gated fusion mechanism enables the knowledge injection intensity to adaptively adjust with the text context. Adaptive skip connections filter low-quality knowledge interference, thereby synergistically improving the accuracy and robustness of business object recognition in three dimensions: industrial long-tail terminology recognition, cross-sentence relation extraction, and knowledge-assisted entity classification.

[0043] The steps for establishing the training dataset for the hierarchical gated knowledge injection dual-stream coding model specifically include: collecting business modeling case texts from the fields of mechanical manufacturing, chemical industry, and metallurgical industry. Each sample contains the original business requirement text, labeled business object sequence labels, object attribute labels, and object relationship triples. At the same time, a corresponding industrial knowledge graph subgraph is constructed as the graph stream input. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1. Low-frequency term samples are expanded to a number close to that of high-frequency samples using synonym substitution and context enhancement.

[0044] The training steps of the hierarchical gated knowledge injection dual-stream coding model specifically include: first, pre-training the text stream and graph stream separately using a single stream; then, using a contrastive loss function to pre-align the dual-stream representations of the same entity, so that the cosine distance between the graph stream representation and the text stream representation converges to within the alignment threshold; finally, joint fine-tuning is performed, with the batch size set to 32 and the initial learning rate set to... A linear learning rate decay strategy is adopted, with a maximum of 50 training epochs. The harmonic mean of accuracy and recall is used as the evaluation metric. Training stops when the harmonic mean of the validation set no longer improves for 5 consecutive epochs. The dependency threshold is determined by iterating through candidate values ​​from 0.1 to 0.9 on the validation set and selecting the candidate value corresponding to the highest harmonic mean of the validation set. The alignment threshold is determined by experimentally comparing candidate cosine distance values ​​in the range of 0.05 to 0.5 on the validation set and selecting the candidate value corresponding to the highest harmonic mean of the validation set.

[0045] Wherein, the object confidence score refers to the maximum predicted probability of the category to which the hierarchical gated knowledge injection dual-stream coding model belongs when performing entity recognition on a certain business object; the specific rule for adjusting the graph flow knowledge injection weights according to the confidence interval rule based on the object confidence score is as follows: Let the object confidence score be... ,when At that time, the fusion gate weights are biased towards the text flow, and the graph flow knowledge injection weight is set to 0.2; when At that time, the fusion gate weights are balanced, and the graph flow knowledge injection weight is set to 0.5; when At that time, the fusion gate weight is biased towards the graph flow, and the graph flow knowledge injection weight is set to 0.8; the above confidence interval boundary values ​​are determined by performing grid search experiments on candidate boundary values ​​on the validation set, and the boundary value combination with the highest harmonic mean of the validation set is used as the final setting.

[0046] The formula for calculating the comprehensive score of the domain deviation is as follows: ;in The overall score is based on the domain deviation. To verify the number of missing objects in the feedback, For the total number of objects, For the number of attribute errors, For the total number of attributes, For the number of misjudged relationships, For the total number of relations, , , The weighted coefficients are and satisfy the following conditions: The weighting coefficients are determined by experimentally comparing candidate weight combinations on the validation set, with the weight combination that yields the highest harmonic mean on the validation set after fine-tuning being used as the final setting; when At that time, the learning rate was used. Full parameter fine-tuning is performed on the hierarchical gating knowledge injection two-stream coding model; when At that time, the learning rate was used. Fine-tuning of the top-level parameters of the hierarchical gating knowledge injection two-stream coding model; when At that time, the learning rate was used. Only the output layer parameters of the hierarchical gating knowledge injection dual-stream coding model are fine-tuned.

[0047] The weighted adjacency matrix is ​​constructed using core business objects as nodes, candidate association pairs determined by the classification labels of business objects as edges, and semantic similarity between objects in the candidate association pairs as edge weights. The Laplacian matrix spectral decomposition refers to transforming the weighted adjacency matrix into a Laplacian matrix, performing eigenvalue decomposition on the Laplacian matrix, and extracting low-frequency eigenvectors corresponding to smaller eigenvalues ​​as the global structural representation of the graph. The spectral energy retention threshold refers to the threshold value before spectral energy retention. Each eigenvalue corresponds to a threshold value representing the proportion of energy to the total energy of all eigenvalues. The low-rank sparse approximation graph retains only the first... The graph structure corresponding to each low-frequency feature vector is obtained by removing low-weight redundant relation edges; the spectral energy retention threshold is set by testing the retention ratio of candidates from 70% to 99% sequentially on the validation set, and the ratio with the highest relation type discrimination accuracy is used as the final setting; the low-frequency feature vector captures the macro-clustering and grouping information of business objects in the relation network, and the high-frequency feature vector corresponds to local noise perturbation. By retaining only the low-frequency components, the relation network is denoised and sparsified, providing high-quality, low-dimensional, dense feature input for multilayer perceptron relation type discrimination, reducing the computational load of the graph and suppressing the interference of noise edges on the relation classification results.

[0048] The random walk algorithm starts with each core business object node on a low-rank sparse approximation graph and iterates through several steps according to the edge weight probability. The co-occurrence frequency between nodes is used as the semantic similarity between objects. Core business object pairs with high semantic similarity between objects are given higher priority as association candidates in relation type discrimination. The semantic similarity between objects is concatenated with the low-frequency feature vector of the corresponding node and used as the input of the multilayer perceptron. The multilayer perceptron outputs the relation type discrimination result. The relation type discrimination results of all core business object pairs and the semantic similarity between objects together constitute a precise business object relation network.

[0049] The VF2++ graph matching algorithm is an improved graph isomorphic subgraph matching algorithm based on the VF2 algorithm. By introducing a candidate node sorting strategy and a backtracking pruning mechanism, it prioritizes expanding the node with the highest degree and strongest constraints during the recursive search process, thereby compressing the number of effective branches in the search tree. In S04, the precise business object relationship network and the modeling template graph are represented as attribute graphs, respectively. The node attributes are the business object classification labels, and the edge attributes are the relationship type discrimination results. The VF2++ graph matching algorithm simultaneously checks the consistency of node attributes and edge attributes during the matching process. Combined with the pruning strategy, it skips subtrees that do not meet the constraints, reducing the template matching search space to an acceptable range when there are many business objects and many relationship types. The VF2++ graph matching algorithm uses the candidate node sorting strategy to make the search tree expand along the branch with the strongest constraints first. The backtracking pruning mechanism immediately terminates the search of the current branch when a local inconsistency is found, reducing the overall search complexity from exponential to polynomial. Thus, even when there are many business objects and many relationship types, template matching can still be completed in a limited time.

[0050] The Monte Carlo tree search heuristic sampling strategy refers to iteratively constructing a search tree on the template matching search tree through four stages: selection, expansion, simulation, and backpropagation. Using the average object confidence score of existing matching results as a heuristic signal, it prioritizes exploring paths with historically high object confidence scores, concentrating search resources on the search branches most likely to produce high-quality template matching results. The Monte Carlo tree search heuristic sampling strategy works synergistically with the VF2++ graph matching algorithm, further avoiding invalid traversals of low-quality paths through probabilistic sampling based on search space pruning, thus compressing the template matching computation time to the second level.

[0051] The specific implementation of step S01 is as follows: After receiving the original natural language industrial business requirement text, the text is first cleaned to remove invalid symbols, noise characters, and redundant formatting content, ensuring the structural consistency of the text. Then, a word segmentation tool based on an industrial domain dictionary is used to segment the cleaned text, and a part-of-speech tagging model is used to assign a part-of-speech tag to each word, providing grammatical feature support for subsequent named entity recognition. In the named entity recognition stage, a conditional random field sequence labeling method is used to identify candidate professional terminology entities in the text. After recognition, the candidate terms are mapped and compared with a domain-specific terminology vocabulary built in a continuously pre-trained corpus of the industrial vertical domain. Non-standard expressions are standardized and replaced, so that synonymous and variant terms in industrial texts from different sources are uniformly mapped to standard terminology entries, ultimately outputting standardized requirement text data. The industrial vertical domain continuous pre-training corpus consists of three types of data: industrial patent documents, industry standard documents, and equipment operation manuals. A domain-adaptive continuous pre-training strategy is used to perform secondary training on the basic pre-training model. By using a vocabulary enhancement method, low-frequency long-tail terms are injected into the word embedding matrix with vectors initialized by context co-occurrence statistics, so that low-frequency terms can obtain semantically distinguishable dense representations.

[0052] The specific implementation of step S02 is as follows: Standardized requirement text data is input into a hierarchical gated knowledge injection dual-stream coding model. The model adopts a dual-stream parallel coding architecture. The text stream uses a Longformer architecture instead of the standard Transformer, and processes ultra-long texts with more than 4096 tags through a hybrid strategy of sliding window attention and global attention to capture long-distance dependencies across sentences. The graph stream uses a graph attention network to perform multi-hop neighbor hierarchical aggregation on the industrial knowledge graph, extracting attribute knowledge in the first hop, relation knowledge in the second hop, and scene knowledge in the third hop. During the training phase, the text stream and graph stream are first pre-trained separately in single stream, and then a contrastive loss is applied to make the cosine distance of the dual-stream representations of the same entity converge to within the alignment threshold (reference value is 0.1 to 0.2), eliminating the semantic gap between the two feature spaces. During joint inference, the dynamic gating fusion mechanism dynamically calculates the fusion gate weights based on the current labeled object confidence scores, achieving adaptive injection of cross-stream information. When the dependency of a certain layer in the text stream on graph features is lower than the dependency threshold (reference value is 0.3), the adaptive skip connection directly bypasses the graph injection step to pass text stream features, avoiding interference from low-quality knowledge. The output layer uses conditional random field decoding to complete entity sequence labeling and outputs the object confidence score, business object classification label, and object attribute label for each business object. The graph stream knowledge injection weights are adjusted according to the confidence interval rules based on the object confidence score: when... The time weight is set to 0.2, when The time weight is set to 0.5, when The time weight was set to 0.8, and the confidence interval boundary values ​​were determined by conducting a grid search experiment on the validation set.

[0053] The specific implementation of step S03 is as follows: Using the identified core business objects as nodes, the associated candidate pairs determined by the business object classification labels as edges, and the semantic similarity between the objects in the associated candidate pairs as edge weights, a weighted adjacency matrix is ​​constructed. Subsequently, the weighted adjacency matrix is ​​transformed into a Laplacian matrix, and eigenvalue decomposition is performed on the Laplacian matrix. Low-frequency eigenvectors corresponding to smaller eigenvalues ​​are extracted as the global structure representation of the graph. Based on the spectral energy retention threshold (reference value is 85% to 95%), the remaining eigenvectors are retained... The low-frequency feature vectors corresponding to each feature value are used to remove low-weight redundant relation edges, resulting in a low-rank sparse approximation graph, thus achieving noise reduction and sparsification of the relation network. On this low-rank sparse approximation graph, starting from each core business object node, a random walk is iteratively performed according to the edge weight probability, and the co-occurrence frequency between nodes is counted as the semantic similarity between objects. The semantic similarity between objects is concatenated with the low-frequency feature vectors of the corresponding nodes to form a joint feature vector, which is then input into a multilayer perceptron to determine the relation type. The output is the relation type of all core business object pairs, which, together with the semantic similarity between objects, constitutes a precise business object relation network.

[0054] The specific implementation of step S04 is as follows: Based on the business object classification labels, the industrial scenario type is determined. The corresponding modeling template graph from the modeling template library is then called. The precise business object relationship network and the modeling template graph are represented as attribute graphs, with node attributes representing business object classification labels and edge attributes representing relationship type discrimination results. The VF2++ graph matching algorithm is used to perform subgraph isomorphic matching between attribute graphs. The VF2++ algorithm prioritizes expanding the node with the highest degree and strongest constraints through a candidate node sorting strategy. Combined with a backtracking pruning mechanism, the current branch is immediately terminated upon discovering local inconsistencies, reducing the search complexity from exponential to polynomial level. The matching process simultaneously verifies the consistency of node and edge attributes to ensure the semantic correctness of template matching. Based on this, a Monte Carlo tree search heuristic sampling strategy is introduced. A search tree is constructed through four iterative stages: selection, expansion, simulation, and backpropagation. The average object confidence score of existing matching results is used as a heuristic signal to prioritize exploring high-confidence template paths, further compressing the template matching computation time to the second level. Finally, a business object model conforming to the unified modeling language specification is automatically generated based on the matching results.

[0055] The specific implementation of step S05 is as follows: Submit the automatically generated business object model to business personnel for manual verification, and collect the number of missing objects in the verification feedback. Attribute error count Number of misjudgments in relation Combine the total number of objects Total number of attributes With all relations According to the formula Comprehensive score of deviation in computing domain Weighting coefficients , , satisfy Furthermore, the candidate weight combinations were determined through experimental comparison on the validation set. The interval to which it belongs adopts a differentiated fine-tuning strategy: when Learning rate Full parameter fine-tuning is performed on the hierarchical gating knowledge injection two-stream coding model; when Learning rate Fine-tune the top-level parameters; when Learning rate Only the output layer parameters are fine-tuned, and the model parameters are iteratively updated until the business object model meets the business requirements.

[0056] It should be noted that the key technologies of this invention include: the contrastive loss pre-alignment mechanism in the hierarchical gated knowledge injection dual-stream coding model, by narrowing the cosine distance between the representations of the same entity in the text stream and the graph stream, ensures that the two heterogeneous feature spaces are in a unified semantic metric space before fusion, fundamentally guaranteeing that the fusion weights calculated by the dynamic gated fusion mechanism are semantically comparable, and eliminating the semantic distortion problem caused by the simple splicing of heterogeneous features in traditional knowledge injection methods; the dynamic gated fusion mechanism works synergistically with adaptive skip connections, the former continuously adjusting the knowledge injection intensity based on the object confidence score, and the latter blocking the transmission of low-quality knowledge in a discrete bypass manner when the dependency is below a threshold. The two complementary approaches form a continuous and discrete dual-layer filtering structure, enabling knowledge injection to effectively supplement insufficient textual evidence while avoiding interference from graph noise when textual evidence is sufficient. The combined application of Laplacian matrix spectral decomposition and random walk algorithms filters out high-frequency local noise disturbances in the relation network by retaining low-frequency feature vectors. Random walks are performed on the dimensionality-reduced sparse graph to obtain macroscopic semantic similarity, providing high-quality low-dimensional dense features for multilayer perceptron relation type discrimination. The three key technologies, from semantic alignment and knowledge filtering to relation network construction, work in sequence and synergistically to ensure that the overall business object modeling process balances accuracy and robustness in low-frequency long-tail industrial terminology scenarios.

[0057] It should be noted that this invention also solves the following technical problem: In industrial business requirement modeling scenarios, when the number of business objects and relationship types is large, the template matching process faces the problem of combinatorial explosion due to the exponential growth of the search space. The computation time of traditional subgraph isomorphic matching algorithms on large-scale attribute graphs is difficult to meet the needs of actual engineering. This invention introduces a candidate node sorting strategy and a backtracking pruning mechanism through the VF2++ graph matching algorithm. It prioritizes expanding the search tree along the branch with the strongest constraints and immediately terminates the current branch when a local inconsistency is found, reducing the overall search complexity from exponential to polynomial. On this basis, a Monte Carlo tree search heuristic sampling strategy is superimposed, using the average object confidence score as a heuristic signal to guide search resources to focus on high-quality matching paths and avoid ineffective traversal of low-confidence paths. The two work together to compress the template matching computation time to the second level, thereby solving the technical problem that the template matching search space is too large and the computation time is unacceptable when the number of business objects and relationship types is large.

[0058] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described AI-based object-oriented business object modeling method.

[0059] A third aspect of the present invention provides an AI-based object-oriented business object modeling system, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0060] Specifically, the principle of this invention is as follows: The core reason why this invention can solve the above-mentioned technical problems lies in the fact that the following three levels of technical logic form a complete closed loop. First, the industrial vertical domain continuous pre-training corpus uses a domain-adaptive continuous pre-training strategy to shift the model parameters towards the distribution of industrial language. The vocabulary enhancement method initializes low-frequency term vectors with context co-occurrence statistics and injects them into the word embedding matrix, enabling long-tail terms that were originally indistinguishable by general models to obtain semantically distinguishable dense representations, thus eliminating the fundamental defect of missing low-frequency term representations at the vocabulary level. Second, the contrastive loss pre-alignment narrows the cosine distance between the text flow representation and the graph flow representation of the same entity, ensuring that the two types of features are in the same semantic space before fusion. This guarantees that the fusion weight calculation of the dynamic gating fusion mechanism has mathematical meaning and avoids knowledge injection distortion caused by semantic space misalignment. Third, the dynamic gating fusion mechanism dynamically adjusts the graph knowledge injection weights based on the object confidence score. When text evidence is sufficient, it reduces graph dependence; when text evidence is insufficient, it enhances graph supplementation. The adaptive skip connection further bypasses the graph injection path directly when the dependence is below a threshold, preventing low-quality knowledge from interfering with text flow features. The three aspects, from representation construction and spatial alignment to adaptive fusion, progress sequentially and logically constitute a complete causal chain. Therefore, the technical solution of this invention conforms to the inherent logic of solving the above-mentioned core technical problems.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0062] The specific implementation of step S01 is as follows: After receiving the original natural language industrial business requirement text, a cleaning operation is first performed to remove invalid symbols and redundant formatting content. Then, the cleaned text is segmented, part-of-speech tagging is performed, and named entity recognition is conducted to obtain a candidate set of professional terms. The candidate terms are then standardized and mapped to the domain-specific terminology vocabulary in the continuously pre-trained corpus of the industrial vertical domain, replacing non-standard expressions with standard terminology entries, and outputting standardized requirement text data. The vocabulary enhancement method injects newly added domain terms into the model's word embedding matrix. The initialization vector is obtained by statistically calculating the context co-occurrence of the terms in the continuously pre-trained corpus of the industrial vertical domain. Let the term... The initial word vectors are The calculation formula is expressed as follows:

[0063] ;

[0064] In the formula, For the term The set of all context words in the corpus For set Number of elements For context words The corresponding vector in the word embedding matrix of the basic pre-trained model Word embedding dimension and All have the same dimensions. The formula uses real-valued vectors, ensuring that both sides of the formula have the same dimensions. By averaging the context word vectors, the formula ensures that the initial vectors of low-frequency terms maintain a consistent semantic direction with their semantic neighbors, thereby enabling low-frequency long-tail terms to obtain semantically distinguishable dense representations.

[0065] The specific implementation of step S02 is as follows: Assume the output layer addresses the business object. The predicted probability vector of the category is Then the object confidence score The calculation formula is expressed as follows:

[0066] ;

[0067] In the formula, Total number of categories for business objects For the output layer to business objects Belongs to the The predicted probability of the class, with a value range of 1. For business objects The object confidence score ranges from 100 to 100. and All values ​​are dimensionless probability values, with unified dimensions on both sides of the formula. Standardized requirement text data is input into a hierarchical gated knowledge injection dual-stream coding model. The model employs a dual-stream parallel coding architecture. The text stream uses a Longformer architecture, employing a hybrid strategy of sliding window attention and global attention to process ultra-long texts with over 4096 tags, capturing long-distance dependencies across sentences. The graph stream uses a graph attention network to perform multi-hop neighbor hierarchical aggregation on the industrial knowledge graph. Let the... Skip graph attention network for nodes The aggregated output is Its formula is expressed as follows:

[0068] ;

[0069] In the formula, For the layer index of the graph attention network, Extract attribute knowledge corresponding to one jump. This corresponds to the knowledge of two-hop extraction relationships. Knowledge extraction from three-hop scenarios For nodes The set of neighboring nodes For the first Learnable weight matrix of layer For nodes In the eigenvectors of the layer To determine the hidden layer feature dimensions of a graph attention network The activation function is typically ReLU. For the first Layer nodes For neighboring nodes The attention coefficient, a dimensionless scalar, is calculated using the following formula:

[0070] ;

[0071] In the formula, For the first Layer-learnable attention vector This represents a vector concatenation operation, resulting in a concatenated vector with dimension 1. For nodes In the The eigenvectors of the layer; LeakyReLU is a linear rectified function with a negative slope, the empirical value of which is 0.2. This is an inner product operation, and the result is a dimensionless scalar with the numerator and denominator having the same dimensions. These are dimensionless normalization coefficients, ensuring dimensionless uniformity on both sides of the formula. and All A 3D real vector, with unified dimensions on both sides of the formula. Contrastive loss pre-alignment enables the text stream representation of the same entity. With spectral flow characterization The cosine distance converges to the alignment threshold. within, Given the dimension of the feature vector after two-stream fusion, the contrast loss formula is expressed as follows:

[0072] ;

[0073] In the formula, The inner product of two vectors and They are respectively and of Norm; the ratio of the two is the cosine similarity, a dimensionless scalar. The alignment threshold, a dimensionless scalar, was determined through experimental comparison of candidate cosine distance values ​​in the range of 0.05 to 0.5 on the validation set. The loss scalar is dimensionless, and the dimensions on both sides of the formula are unified. The dynamic gating fusion mechanism dynamically calculates the fusion gate weights based on the confidence scores of the currently labeled objects. Its formula is expressed as follows:

[0074] ;

[0075] In the formula, The text stream feature vector of the current label The feature vectors of the currently labeled graph flow are both of dimension 1. Gating weight matrix For bias vector Concatenated vector of features from text stream and graph stream The sigmoid activation function, after element-wise application, has an output range of... For dimensionless gated vectors, the dimensions are unified on both sides of the formula. The fused feature vector. for:

[0076] ;

[0077] In the formula, For element-wise multiplication and All 3D real vector; fusion gate weights Spectral Flow Components in Chinese The mean of each element is adjusted according to the confidence interval rule: when When the graph flow knowledge injection weight is set to 0.2; when Set to 0.5 when; The time is set to 0.8 and , The units are the same, and the units on both sides of the formula are consistent. (Text stream, number...) Dependence of layers on map features Defined as the mean of the graph flow components in the fusion gate weights, the formula is as follows:

[0078] ;

[0079] In the formula, For the first Total number of tags processed by layer For the first Layer The fusion gate weights corresponding to each marker The mean of all elements is a dimensionless scalar. Weights of the corresponding spectral flow components It is a dimensionless scalar, and its range of values ​​is 1. The dimensions on both sides of the formula are consistent; when When, skip the first The layer's map injection step directly... Pass to the next layer The dependency threshold, a dimensionless scalar, is determined by iterating through candidate values ​​from 0.1 to 0.9 on the validation set, and selecting the candidate value corresponding to the highest harmonic mean of the validation set. A reference value of 0.3 is used. The output layer fuses conditional random field decoding to complete entity sequence labeling and outputs the object confidence score for each business object. Business object category tags and object attribute tags.

[0080] The specific implementation of step S03 is as follows: using the identified Each core business object constructs a weighted adjacency matrix for nodes. Matrix elements For business objects With business objects Semantic similarity between objects, with a value range of , is a dimensionless quantity, diagonal element .Will Transform into a normalized Laplace matrix Its formula is expressed as follows:

[0081] ;

[0082] In the formula, identity matrix For a degree matrix, its diagonal elements Off-diagonal elements are 0 for a diagonal matrix with a power of -1 / 2 , , All For a real number matrix, all terms are dimensionless, and the dimensions are consistent on both sides of the formula. The eigenvalue decomposition is expressed by the following formula:

[0083] ;

[0084] In the formula, This is an eigenvalue diagonal matrix, with the eigenvalues ​​arranged in ascending order. All are dimensionless real numbers. For the corresponding eigenvector matrix, For the first Unit eigenvectors corresponding to each eigenvalue , All For real number matrices, the dimensions are consistent on both sides of the formula. Based on the spectral energy retention threshold. Before keeping A low-frequency feature vector, The following conditions must be met:

[0085] ;

[0086] In the formula, The number of candidate eigenvalues ​​to retain, which is a positive integer. The spectral energy retention threshold is a dimensionless ratio. It is determined by sequentially testing the retention rates of candidates from 70% to 99% on the validation set, and the ratio that yields the highest accuracy in relation type discrimination is used as the final setting. The reference value is 85% to 95%. and Both are the sum of eigenvalues, and the ratio between them is dimensionless, ensuring that the dimensions are consistent on both sides of the formula. (Before extraction) Low-frequency eigenvector matrix Low-weight redundant edges are removed to obtain a low-rank sparse approximation graph. A random walk is then performed on this low-rank sparse approximation graph, starting from each core business object node. The semantic similarity between objects is obtained by normalizing the co-occurrence frequency among nodes. , is a dimensionless scalar, and its range is . .Will With nodes , The low-frequency feature vectors are concatenated to form a joint feature vector. Its formula is expressed as follows:

[0087] ;

[0088] In the formula, for The Row, i.e., node Corresponding low-frequency feature vector for The Row, i.e., node Corresponding low-frequency feature vector It is a dimensionless scalar. and Each component is a dimensionless real number. Since it is a dimensionless vector, the dimensions are unified on both sides of the formula. The input is a multilayer perceptron, and the output is a relation type determination result. The relation type determination results of all core business object pairs are compared with the semantic similarity between the objects. Together they form a precise network of business object relationships.

[0089] The specific implementation of step S04 is as follows: Based on the business object classification labels, determine the industrial scene type; call the corresponding modeling template library; and represent the precise business object relationship network and the modeling template graph as attribute graphs, respectively. Node attributes are business object classification labels, and edge attributes are relationship type discrimination results. The VF2++ graph matching algorithm prioritizes expanding the node with the highest degree and strongest constraints through a candidate node sorting strategy. Combined with a backtracking pruning mechanism, it immediately terminates the current branch when a local inconsistency is found, reducing the overall search complexity from exponential to polynomial. The Monte Carlo tree search heuristic sampling strategy iteratively constructs the search tree on the template matching search tree through four stages: selection, expansion, simulation, and backpropagation. The average confidence score of the existing matching results is used as a heuristic signal. Node selection adopts the confidence upper bound formula, specifically described as follows:

[0090] ;

[0091] In the formula, For search tree nodes The mean confidence score of historical matching results is a dimensionless scalar with a range of values ​​of [value missing]. For nodes The number of visits is a positive integer. For nodes The number of visits to the parent node is a positive integer. The coefficient is an exploration coefficient, a dimensionless scalar, with an empirical value of . It is a dimensionless scalar. and Both are dimensionless quantities, and the dimensions on both sides of the formula are consistent. The synergistic effect of these two factors reduces the template matching calculation time to the second level.

[0092] The specific implementation of step S05 is as follows: Submit the business object model to business personnel for verification, and collect the number of missing objects in the verification feedback. Attribute error count Number of misjudgments in relation Combine the total number of objects Total number of attributes With all relations The comprehensive score for domain deviation is calculated according to the following formula. :

[0093] ;

[0094] In the formula, The comprehensive score for domain deviation has a range of values. , is a dimensionless quantity To verify the number of missing objects in the feedback, a non-negative integer is used. The total number of objects is a positive integer. The number of attribute errors is a non-negative integer, expressed here as... Mark to avoid the base of the natural logarithm Confusion The total number of attributes is a positive integer. The number of misjudged relationships is a non-negative integer. Let be the total number of relations, and be a positive integer. , , These are object omission rate, attribute error rate, and relationship misjudgment rate, all of which are dimensionless ratios. , , These are weighting coefficients, all of which are dimensionless scalars, satisfying... By conducting experiments and comparisons on the validation set to test candidate weight combinations, the weight combination that yields the highest harmonic mean on the validation set after fine-tuning is taken as the final setting. This is a weighted sum of dimensionless ratios, with dimensions unified on both sides of the formula. Based on... The interval to which it belongs adopts a differentiated fine-tuning strategy: when Learning rate Full parameter fine-tuning is performed on the hierarchical gating knowledge injection two-stream coding model; when Learning rate Fine-tune the top-level parameters; when Learning rate Only the output layer parameters are fine-tuned, and the model parameters are iteratively updated until the business object model meets the business requirements.

[0095] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians set up a test environment and collected the business requirements text of CNC machining equipment from a mechanical manufacturing enterprise as input. Using the technical solution of the invention, they completed the full-process test of automated modeling from the original requirements text to the business object model of the unified modeling language specification.

[0096] Technical personnel selected a batch of business requirement texts describing CNC machining center equipment management, work order scheduling, and tool maintenance. The total number of texts was 320, with an average length of 2180 characters per text. The texts contained highly specialized, low-frequency, long-tail terms such as spindle speed compensation, tool wear coefficient, and G-code macro program calls. Low-frequency terms (terms appearing less than 0.1% of the total in the corpus) accounted for 38% of all terms in the entire sample, constituting the main challenge of this test.

[0097] In step S01, technicians input 320 raw texts into the preprocessing pipeline. After cleaning, word segmentation, part-of-speech tagging, and named entity recognition, the texts are standardized by combining them with a lexicon of CNC machining terminology built from a continuously pre-trained corpus in the industrial vertical field. A vocabulary enhancement method injects low-frequency terms such as spindle speed compensation and tool wear coefficient into the word embedding matrix using vectors initialized with contextual co-occurrence statistics. This ensures that low-frequency terms have semantically distinguishable vector representations in subsequent encoding stages, outputting 320 standardized requirement text data.

[0098] In step S02, the standardized requirement text data is input into a hierarchically gated knowledge injection dual-stream coding model. The text stream uses a Longformer architecture to process ultra-long texts, while the graph stream uses a graph attention network to perform three-hop hierarchical aggregation on the CNC machining domain knowledge graph, extracting attribute knowledge, relational knowledge, and scene knowledge. Contrastive loss pre-alignment ensures that the cosine distance between the text stream and graph stream representations converges to within 0.15. The dynamic gating fusion mechanism adaptively adjusts the graph stream knowledge injection weights based on the object confidence scores of each business object, and adaptive skip connections directly bypass the graph injection path when the dependency is below 0.3. The model identifies a total of 1284 core business objects, with business object classification labels covering four categories: equipment entities, work order entities, tool entities, and parameter entities, as shown in Table 1. The output object confidence score distribution is also shown.

[0099] Table 1. Statistical Table of Business Object Classification and Identification Results

[0100]

[0101] In step S03, a weighted adjacency matrix is ​​constructed using 1284 core business objects as nodes. Eigenvalue decomposition is performed on the Laplacian matrix, and the spectral energy retention threshold is set to 90%. [Preparation of the spectral energy is then performed.] A low-rank sparse approximation graph is obtained from low-frequency feature vectors. After removing low-weight redundant relation edges, the number of relation edges is compressed from the initial 18762 to 4231. A random walk is performed on the low-rank sparse approximation graph, and the co-occurrence frequency between nodes is used as the semantic similarity between objects. The semantic similarity is concatenated with the low-frequency feature vectors and then input into a multilayer perceptron to complete relation type discrimination. The output is a precise business object relation network containing 5 relation types. The distribution of the 5 relation types and their number of relations is as follows: Figure 2 As shown.

[0102] In step S04, based on the business object classification tags, the industrial scenario type is determined to be a CNC machining equipment management scenario. The corresponding modeling template library is then invoked to construct an attribute graph by combining the precise business object relationship network with the modeling template graph. The VF2++ graph matching algorithm, combined with a pruning strategy, compresses the search space from its initial size to an acceptable range. The Monte Carlo tree search heuristic sampling strategy prioritizes exploring high-quality template paths using the mean of the object confidence score as a heuristic signal. Figure 3 The image shows the sampling weight distribution of different confidence paths during the template matching process in Monte Carlo tree search. Template matching is completed in seconds, automatically generating a business object model that conforms to the Unified Modeling Language specification.

[0103] In step S05, business personnel verify the generated business object model, and the feedback results are shown in Table 2.

[0104] Table 2. Statistical Table of Verification Feedback from Business Personnel

[0105]

[0106] Based on the verification feedback, a comprehensive score of domain deviation is calculated, with weighting coefficients. , , Calculated ,belong Interval, using learning rate Only the output layer parameters are fine-tuned, the model parameters are updated, and the verification is submitted again to ensure that the business object model meets the business requirements.

[0107] Compared to traditional manual modeling methods, this invention achieves several fundamental advancements at the technical principle level. Traditional methods rely on manual text reading to identify objects one by one. The recognition of low-frequency, long-tail industrial terms depends entirely on the accumulated experience of business personnel, and cannot formally and quantitatively represent the semantics of the terms. This leads to low-frequency terms being easily overlooked or misjudged in complex requirement texts. This invention constructs semantically distinguishable dense vector representations for low-frequency terms through a lexical enhancement method, enabling low-frequency terms to have the same semantic accessibility as high-frequency terms during the encoding stage, thus eliminating the root cause of the low accuracy of low-frequency term recognition at the representation level. Traditional knowledge graph-assisted methods typically use a hard-joining approach to fuse textual and graph features. The measurements of the two types of features in different semantic spaces are incomparable, and the feature quality after fusion is severely affected by semantic distortion. This invention uses contrastive loss pre-alignment to ensure that the two types of features are in a unified semantic measurement space before fusion. The dynamic gating fusion mechanism and adaptive skip connections further ensure the adaptive controllability of the strength and quality of knowledge injection, achieving a fundamental improvement at the fusion mechanism level. In the template matching stage, traditional subgraph isomorphism methods face combinatorial explosion on large-scale attribute graphs. The synergistic effect of the VF2++ algorithm and Monte Carlo tree search reduces the search complexity from exponential to polynomial, making template matching possible in large-scale industrial modeling scenarios within engineering time constraints.

[0108] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0109] Table 3. Variable Explanation Table (Part 1)

[0110]

[0111] Table 4. Variable Explanation Table (Part Two)

[0112]

[0113] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for modeling business objects in an object-oriented programming paradigm based on artificial intelligence, characterized in that, Includes the following steps: The system receives industrial business requirement texts in natural language, performs cleaning, word segmentation, part-of-speech tagging and named entity recognition through natural language processing technology, and standardizes the mapping of professional terms by combining a continuously pre-trained corpus of the industrial vertical field to obtain standardized requirement text data. Standardized requirement text data is input into a hierarchical gating knowledge injection dual-stream coding model, which outputs the object confidence score, business object classification label, and object attribute label for each business object. The weight of the graph flow knowledge injection is adjusted according to the confidence interval rule based on the object confidence score to complete the identification of core business objects, object attributes, and business object classification labels. A weighted adjacency matrix is ​​constructed based on the identified core business objects and their classification labels. The weighted adjacency matrix is ​​then subjected to Laplacian matrix spectral decomposition. A low-rank sparse approximation graph is extracted based on the spectral energy retention threshold. The semantic similarity between objects is calculated on the low-rank sparse approximation graph using a random walk algorithm. The semantic similarity between objects is then concatenated with the low-frequency feature vectors of the corresponding nodes and input into a multilayer perceptron to complete the relationship type discrimination and output a precise business object relationship network. The industrial scenario type is determined based on the business object classification tags. The corresponding modeling template library is called, and the accurate business object relationship network and modeling template graph are constructed into an attribute graph. The VF2++ graph matching algorithm combined with the pruning strategy is used to reduce the search space. The Monte Carlo tree search heuristic sampling strategy is introduced to prioritize the exploration of high-confidence template paths and automatically generate a business object model that conforms to the unified modeling language specification. The business object model is submitted to business personnel for verification. Based on the number of missing objects, attribute errors, and relationship misjudgments in the verification feedback, a comprehensive domain deviation score is calculated. When the comprehensive domain deviation score falls into different intervals, different learning rate strategies are used to fine-tune the hierarchical gating knowledge injection dual-stream coding model. The model parameters are updated until the business object model meets the business requirements.

2. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 1, characterized in that, The industrial vertical domain continuous pre-training corpus consists of three types of data: industrial patent documents, industry standard documents, and equipment operation manuals. A domain-adaptive continuous pre-training strategy is used to perform secondary training on the basic pre-training model, and a domain-specific terminology vocabulary is dynamically injected into the word embedding layer through a vocabulary enhancement method.

3. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 2, characterized in that, The vocabulary enhancement method refers to injecting new domain terms into the model word embedding matrix in a vector initialization manner. The initialization vector is obtained by statistical calculation of the co-occurrence of the term in the context of the continuously pre-trained corpus of the industrial vertical domain.

4. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 3, characterized in that, The hierarchical gated knowledge injection dual-stream coding model adopts a dual-stream parallel coding architecture. The text stream processes standardized requirement text data based on an improved RoBERTa encoder, while the graph stream performs hierarchical aggregation of multi-hop neighbors of the industrial knowledge graph based on a graph attention network. The first hop extracts attribute knowledge, the second hop extracts relational knowledge, and the third hop extracts scene knowledge.

5. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 4, characterized in that, The text stream uses the Longformer architecture with sparse attention mechanism instead of the standard Transformer. It captures long-distance dependencies across sentences through a hybrid strategy of sliding window attention and global attention, supporting ultra-long text processing with more than 4096 tags.

6. The method for modeling business objects in an object-oriented programming system based on artificial intelligence according to claim 5, characterized in that, The text stream and the graph stream are injected with cross-stream information through a dynamic gating fusion mechanism. The fusion gate weight of the dynamic gating fusion mechanism is dynamically calculated by the confidence score of the currently labeled object. An adaptive skip connection is introduced inside the network. When the dependence of a certain layer of the text stream on the graph features is lower than the dependence threshold, the graph injection step is skipped and the text stream features are directly transmitted.

7. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 6, characterized in that, The hierarchical gated knowledge injection dual-stream coding model fusion contrastive learning technique applies contrastive loss to the text stream representation and graph stream representation of the same entity for pre-alignment, narrowing the cosine distance between the two and ensuring the consistency of the dual-stream semantic space. The output layer fuses conditional random field decoding to complete entity sequence labeling and output object confidence scores.

8. The method for modeling business objects in an object-oriented programming based on artificial intelligence according to claim 7, characterized in that, The weighted adjacency matrix is ​​constructed by using core business objects as nodes, candidate pairs of associations determined by the classification labels of business objects as edges, and the semantic similarity between objects in the candidate pairs as edge weights; the Laplacian matrix spectral decomposition extracts low-frequency feature vectors corresponding to smaller eigenvalues ​​as the global structural representation of the graph.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the AI-based object-oriented business object modeling method according to any one of claims 1-8.

10. A business object modeling system based on AI in object-oriented programming, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images