Document processing method and system fusing industry knowledge graph and dynamic process weaving network

Abstract

The application belongs to the field of official document intelligent processing, and discloses an official document processing method and system fusing industry knowledge graph and dynamic process weaving. The application firstly extracts and determines exclusive semantic analysis triplets, constructs an industry fusion knowledge graph, and generates a professional pseudo-official document word vector in linkage with a platform, forms an examination and approval flow path parameter in combination with an official document identity vector and an urgency vector relying on a work flow engine, extracts a key semantic analysis feature vector of a triplet cluster formed in the knowledge graph with an input official document based on a bidirectional relative entropy, calculates and obtains an official document text network correlation feature and a dynamic process weaving urgency related feature to construct an official document intelligent processing model, and realizes output of an official document intelligent processing real-time method for examination and approval decision assistance during official document processing based on the model. The application deeply fuses an industry knowledge base, business data and an intelligent work flow, and realizes automatic processing of official document semantic analysis, opinion generation and process flow transfer in all links.

Description

Technical Field

[0001] This invention belongs to the field of intelligent document processing technology, and in particular relates to a document processing method and system that integrates industry knowledge graphs and dynamic process networks. Background Technology

[0002] Government and state-owned enterprise documents face challenges such as complex business categories, numerous professional terms, scattered business data, diverse document types, and flexible and ever-changing approval processes. Traditional document processing methods rely on manual review of each document, manual drafting of opinions, and fixed workflows, resulting in low processing efficiency, insufficient information mining, homogenized opinions, and rigid approval processes, making them unsuitable for the complex and ever-changing needs of document handling. Current mainstream document processing technologies mostly focus on basic functions such as general text extraction and meeting content monitoring, only achieving superficial information capture. They cannot combine industry-specific business logic for deep semantic understanding, nor can they connect business databases to achieve data linkage. Furthermore, they generally employ fixed approval processes, failing to adaptively adjust workflows based on the actual situation of the document, resulting in insufficient overall intelligence.

[0003] Therefore, existing technologies face the challenge of how to deeply integrate industry knowledge bases, underlying business data, and intelligent workflow engines, relying on large language models combined with proprietary semantic parsing models to achieve deep semantic parsing of official documents, linking business data to generate data-driven draft opinions, and driving dynamic adaptation of the approval process based on the parsing results to complete the intelligent handling of official documents throughout the entire process. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a document processing method and system that integrates industry knowledge graphs and dynamic process networks.

[0005] In a first aspect of the present invention, a document processing method integrating industry knowledge graphs and dynamic process networks is provided, the method comprising:

[0006] Based on a large language model combined with a dedicated industry lexicon and business database, we extract and determine dedicated semantic parsing triples and construct an industry-integrated knowledge graph.

[0007] By linking with the underlying digital foundation data platform, specialized word vectors for drafting are generated, and by relying on the workflow engine and combining document identity vectors and urgency vectors, dynamic process network urgency features are formed.

[0008] Based on entropy-based similarity extraction technology, key semantic parsing feature vectors of triplet clusters formed with input documents in the knowledge graph are obtained, and document text network relational features are calculated and obtained. Based on the key semantic parsing feature vectors, document text network relational features, specialized drafting word vectors, dynamic process network urgency features, and corresponding document intelligent processing methods, a document intelligent processing model is constructed.

[0009] Based on the aforementioned document intelligent processing model, the document intelligent dynamic processing method is output when the document to be processed is input, and document intelligent processing for approval decision assistance is performed according to the aforementioned document intelligent dynamic processing method.

[0010] Furthermore, the specialized proposed word vector is composed of the latent intent features of official documents and the number of proposed approval opinions in historical official documents retrieved from the underlying digital base data platform based on the exclusive semantic parsing triplet.

[0011] Furthermore, the approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

[0012] Furthermore, the entropy-based similarity extraction technique employs bidirectional divergence relative entropy to extract the key semantic parsing feature vector. The bidirectional divergence relative entropy is obtained through knowledge graph and triple cluster calculation. The triple cluster is obtained by processing the input document. All triples are extracted from the input document to obtain the triple cluster. The triple with the highest bidirectional divergence relative entropy with the knowledge graph in the triple cluster is used as the key semantic parsing feature vector of the intelligent document processing model.

[0013] Furthermore, the network relational features of the official document text are obtained by taking the average value of the feature vector centrality of the input document triples and the industry-integrated knowledge graph.

[0014] Furthermore, before constructing the intelligent document processing model based on the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature, and the corresponding intelligent document processing method, the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature are normalized and then spatial feature vectors are concatenated to obtain the intelligent document processing feature. The intelligent document processing feature is then used as the input feature for constructing the intelligent document processing model.

[0015] Furthermore, the document intelligent processing model adopts an ensemble neural network model improved based on overlapping training set data. This ensemble neural network model is obtained by integrating K sub-neural network models. In the K sub-neural network models trained on each sub-training set, the proportion of overlapping data in the training set of the Kth sub-neural network model is [missing information]. ,in Less than or equal to 0.3 and Greater than 0.

[0016] It also provides a document processing system that integrates industry knowledge graphs and dynamic workflow networks. This system implements a document processing method that integrates industry knowledge graphs and dynamic workflow networks. The system includes an industry-integrated knowledge graph module, an underlying digital foundation data platform module, a workflow engine module, a document intelligent processing model construction module, and a document intelligent dynamic processing module.

[0017] The industry-integrated knowledge graph module is used to extract and determine exclusive semantic parsing triples based on a large language model combined with a dedicated industry lexicon and business database, and to construct an industry-integrated knowledge graph.

[0018] The underlying digital base data platform module is used to generate specialized word vectors in conjunction with the underlying digital base data platform.

[0019] The workflow engine module is used to form a dynamic process network with urgency characteristics by combining the document identity vector and the urgency vector with the workflow engine.

[0020] The document intelligent processing model construction module is used to obtain the key semantic parsing feature vectors of the triple clusters formed with the input documents in the knowledge graph based on entropy similarity extraction technology, and to calculate and obtain the document text network relationship features. Based on the key semantic parsing feature vectors, the document text network relationship features, the specialized drafting word vectors, the dynamic process network urgency features, and the corresponding document intelligent processing methods, a document intelligent processing model is constructed.

[0021] The document intelligent dynamic processing module is used to output the document intelligent dynamic processing method when the document to be processed is input based on the document intelligent processing model, and to perform document intelligent processing to assist approval decision-making based on the document intelligent dynamic processing method.

[0022] Furthermore, the specialized proposed word vector is composed of the latent intent features of official documents and the number of proposed approval opinions in historical official documents retrieved from the underlying digital base data platform based on the exclusive semantic parsing triplet.

[0023] Furthermore, the approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

[0024] Therefore, the beneficial effects of this invention are that it utilizes entropy characteristics to calculate word vector similarity, which can measure the semantic similarity of two word vectors from the dimension of information distribution consistency. This is more in line with the semantic representation logic of word vectors in large models and is highly suitable for the document keyword and business word vector matching scenarios of this patent. Compared with the construction of traditional non-overlapping batch neural network models, this invention utilizes the coupling theory of overlap ratio and loss function. By adjusting the overlap ratio, it adaptively changes the learning intensity of the loss for important documents, balancing training stability and long-tail sample fitting effect. Increasing the overlap ratio increases the frequency of important documents appearing in multiple batches, improving the expected contribution of such samples to the global loss, and shifting the focus of the loss function fitting towards the complex business document processing method. The training sample distribution of multiple sub-networks tends to be consistent, the average loss of each branch is balanced, and the stability of the prediction result is greatly improved after integration and fusion. At the same time, this overlap interval scheme is suitable for the private computing power of government intranets, reducing training hardware costs and making it more feasible for engineering implementation. This solution achieves deep coupling between industry knowledge base and semantic parsing, linkage between business data and proposed opinions, and adaptive workflow driven by semantic analysis results, breaking through the limitations of traditional technologies such as shallow semantic understanding, templated content, and fixed processes. This solution boasts high semantic recognition accuracy, opinion content closely aligned with actual business operations, and strong process adaptability, significantly improving the overall efficiency of official document processing. It can be widely applied to government and state-owned enterprise document processing scenarios such as market supervision and environmental protection.

[0025] Further embodiments and improvements of the present invention will be described in conjunction with the accompanying drawings and specific examples. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the integrated neural network of the present invention;

[0027] Figure 2 This is a schematic diagram of the document processing method of the present invention that integrates industry knowledge graphs and dynamic process networks;

[0028] Figure 3 This is a schematic diagram of a document processing system that integrates industry knowledge graphs and dynamic process networks in an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the neural network loss function in an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of an electronic device structure for implementing the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0031] The invention will now be further described in conjunction with the accompanying drawings and specific embodiments. The integrated neural network model improved based on overlapping training set data used in this invention is a pattern recognition multi-model loss function improvement and fusion construction after specific training of the neural network model in the scenario of this application and corresponding optimization processing of input data. In the scenario of intelligent document approval processing of the input model, the feature processing of data and the training integration of the model adopt different training methods from the original neural network classification model.

[0032] This invention integrates industry knowledge graphs and dynamic process networks to create an innovative intelligent document processing model. It is applied to the fields of data processing optimization and model training optimization of large models, belonging to the field of artificial intelligence optimization and the field of artificial intelligence technology.

[0033] In a first aspect of the present invention, a document processing method integrating industry knowledge graphs and dynamic process networks is provided, the method comprising:

[0034] Based on a large language model combined with a dedicated industry lexicon and business database, we extract and determine dedicated semantic parsing triples and construct an industry-integrated knowledge graph.

[0035] By linking with the underlying digital foundation data platform, specialized word vectors for drafting are generated, and by relying on the workflow engine and combining document identity vectors and urgency vectors, dynamic process network urgency features are formed.

[0036] Based on entropy-based similarity extraction technology, key semantic parsing feature vectors of triplet clusters formed with input documents in the knowledge graph are obtained, and document text network relational features are calculated and obtained. Based on the key semantic parsing feature vectors, document text network relational features, specialized drafting word vectors, dynamic process network urgency features, and corresponding document intelligent processing methods, a document intelligent processing model is constructed.

[0037] Based on the aforementioned document intelligent processing model, the document intelligent dynamic processing method is output when the document to be processed is input, and document intelligent processing for approval decision assistance is performed according to the aforementioned document intelligent dynamic processing method.

[0038] In this embodiment, relying on the powerful semantic understanding capabilities of the large language model, and combining the industry-specific thesaurus of government and public institutions with the full-domain business database, refined semantic adaptation and deep analysis are carried out. The system automatically completes the identification of document entities, extraction of business relationships, and mining of attribute parameters, accurately extracting and determining the document-specific entity-relationship-attribute semantic parsing triples. By aggregating, deduplicating, and iterating the associations of single document triples and clusters of triples from multiple documents of the same origin, the system automatically constructs an industry-integrated knowledge graph adapted to the office scenarios of government and public institutions, realizing the structured and graph-based accurate modeling of document business entities, business logic relationships, departmental authority and responsibility attributes, and document processing rules.

[0039] In one embodiment of this application, a common project funding application document from public institutions is selected as the processing sample. The original core text of the document is: "The Municipal Housing and Construction Bureau applies to the Municipal Finance Bureau for the allocation of special renovation funds for the 2026 urban old residential area renovation project. This project is led by the Municipal Housing and Construction Bureau and implemented in coordination with the street offices. The funding review and approval need to be completed within 7 working days. It is a key municipal-level supervised livelihood project." The system performs cleaning and preprocessing on the original document, removing redundant punctuation, blank lines, and invalid modifiers, while retaining the core business semantic text, providing standardized text input for triple extraction.

[0040] Based on a large language model, combined with semantic boundaries constrained by industry-specific thesaurus for government and public institutions, and linking with field mapping relationships in the business database, the preprocessed official document text undergoes three-dimensional mining of entities, relationships, and attributes. Multiple initial semantic triples are automatically extracted, and through dual verification by the industry thesaurus and the business database, erroneous and ambiguous triples are eliminated. The final standardized and compliant exclusive semantic parsing triples are as follows:

[0041] Tripartite Group 1 (Business Entity - Execution Relationship - Collaborating Units): <Municipal Housing and Construction Bureau, leading coordination, Sub-district Office>

[0042] Group 2 (Business Item - Application Relationship - Approval Authority): <Old Residential Area Renovation Project, Fund Disbursement Application, Municipal Finance Bureau>

[0043] Triad 3 (Document Items - Attribute Constraints - Processing Time Limit): <Fund Disbursement Approval, Completion Time Limit, 7 Working Days>;

[0044] Triad Group 4 (Project Type - Attribute Definition - Supervision Level): <Old Residential Area Renovation Project, Project Attribute: Key Municipal Supervision Project for People's Livelihood>

[0045] During the extraction process, industry-specific terms such as "leading and coordinating," "fund allocation," and "project supervision" are accurately identified using an industry-specific thesaurus. The authenticity of departmental responsibilities and project supervision attributes is verified using a business database to ensure that the extracted triplets fully conform to the business rules of public institutions, with no semantic bias or business logic errors. This forms a set of triplets specific to each official document. To facilitate subsequent processing by the intelligent document processing model, these triplets are represented in a vector form that the model can recognize. For example, triplet 1 can be represented as (5, 0.3, 3), and other triplets are represented similarly. In principle, when constructing an industry-integrated knowledge graph, the aforementioned triples 1, 2, 3, and 4 are connected by relation edges to form part of the industry-integrated knowledge graph. Since they are extracted from the same document, their relation edges are set to 1. If the triples extracted from other documents are not closely related, the relation edges for the two types of triples are set to 0. This does not mean that the industry-integrated knowledge graph is constructed from only two types of documents, but rather it is a knowledge graph network constructed from multiple categories of triples extracted from multiple types of documents. This will not be elaborated further here.

[0046] like Figure 2 The diagram shows a document processing method that integrates industry knowledge graphs and dynamic process networks.

[0047] Furthermore, the specialized proposed word vector is composed of the latent intent features of official documents and the number of proposed approval opinions in historical official documents retrieved from the underlying digital base data platform based on the exclusive semantic parsing triplet.

[0048] The data source for constructing specialized drafting word vectors consists of two parts: first, semantic features of latent intent in official documents extracted from the underlying digital platform, using dedicated semantic parsing triples as retrieval keys; and second, frequency features of various types of proposed approval opinions in historical official documents of similar business obtained from platform statistics. After normalizing and merging these two types of features, specialized drafting word vectors for intelligent document drafting tasks are formed. The latent intent features of official documents belong to deep semantic features, capturing implicit semantics such as document requests, processing purposes, and business constraints. The frequency of historical proposed approval opinions belongs to statistical prior features; the higher the frequency, the stronger the fit of the proposed opinion, serving as a vector weight term to improve the relevance of intelligent drafting and automatic drafting results.

[0049] In one embodiment of this application, if the exclusive semantic parsing triple is (5, 0.3, 3), it indicates that the keyword of the official document is "the Municipal Housing and Construction Bureau leading and coordinating with the sub-district offices." Based on this triple, the data platform clusters and recalls a total of 327 historical official documents of the same type in recent years, namely "the Municipal Housing and Construction Bureau leading and coordinating with the sub-district offices." The frequency of historical proposed handling and proposed approval opinions at all approval nodes is statistically analyzed to form frequency features, specifically:

[0050] (1) Frequently requested opinion 1: "After verification, the project is found to be true. It is recommended that the Finance Bureau review the project and submit it to the leadership for approval" (appeared 216 times).

[0051] (2) Frequently requested opinion 2: "The materials are complete and meet the funding application specifications. It is recommended to approve the project" (appeared 189 times).

[0052] (3) Low-frequency comments: "The budget needs to be supplemented with supporting materials before review" (appeared 23 times).

[0053] The corresponding specialized proposed word vector is composed of (proposed opinion 1, proposed opinion 2, proposed opinion 3, frequency 1, frequency 2, frequency 3), where the frequency is the number of times the corresponding proposed opinion appears. In the above embodiment, it is (2, 6, 9, 216, 189, 23). The corresponding proposed opinions are also transformed into feature vector values ​​to adapt to the construction of the subsequent intelligent document processing model.

[0054] The specialized word vectors selected in this application differ from general pre-trained word vectors. These vectors integrate prior statistical data on industry-specific documents, thus addressing the shortcomings of general vectors in adequately adapting to the intent and approval opinions of specialized documents in official documents.

[0055] Furthermore, the approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

[0056] In this embodiment, the document identification vector is obtained based on the unit with the highest level and the most frequent occurrences in the document, while the urgency vector is set accordingly based on the document level. The workflow engine extracts the units corresponding to the corresponding feature values ​​in the document, extracting the feature values ​​of the highest-level unit and the unit mentioned most frequently. The approval workflow path parameters are obtained by concatenating the aforementioned feature values ​​and the urgency vector.

[0057] Official documents often involve multiple departments and levels of stakeholders during processing. The administrative level of different units determines the approval level, and the frequency of unit mentions indicates the core responsible party for the matter. This invention abandons the single keyword matching method and integrates two dimensions of features: administrative level weight priority and text mention frequency priority, ensuring accurate identification of the document's sponsoring entity, jurisdictional level, and review scope. An urgency vector differentiates between ordinary, urgent, and special urgency levels, used to control node skipping, parallel weaving, and timeliness supervision weights in the process chain. The approval flow path parameters obtained by concatenating these three types of features serve as the core input parameters for the dynamic process weaving engine to automatically generate approval topology, routing branches, and process priorities.

[0058] Taking the official document "Request for 550,000 Yuan in Infrastructure Repair Funds for XX Street" as an example:

[0059] First, the entire document undergoes entity identification. The document mentions the following entities in sequence: Subdistrict Office, District Finance Bureau, District Government, and the Deputy District Head in charge. Based on a preset administrative level weighting system, the system determines the highest-level unit to be the District Government (district level) and extracts its administrative level feature vector. Simultaneously, the frequency of entity mentions in the entire document is counted. The Subdistrict Office is the most frequently mentioned business entity in this document, and its high-frequency entity feature vector is extracted. These two types of vectors are merged to generate a unique document identity vector. Further, the document's urgency label is identified. If the document is a routine matter, a corresponding low-priority urgency vector is generated according to the level mapping rules. If it is an urgent or super-urgent document, a high-weight urgency vector is generated to trigger the green channel and overdue expedited supervision mechanism. The workflow engine concatenates the feature values ​​of the highest-level unit, the most frequent unit, and the urgency vector to obtain complete approval flow path parameters.

[0060] Furthermore, the entropy-based similarity extraction technique employs bidirectional divergence relative entropy to extract the key semantic parsing feature vector. This bidirectional divergence relative entropy is obtained through knowledge graph and triple cluster calculation. The triple cluster is obtained by processing input documents. All triples are extracted from the input documents to form the triple cluster. The triple with the highest bidirectional divergence relative entropy in the triple cluster relative to the knowledge graph is used as the key semantic parsing feature vector for the intelligent document processing model. The formula for calculating the bidirectional divergence relative entropy between a triple in the knowledge graph and the triple obtained from document processing is as follows:

[0061]

[0062] In the formula, Let P and Q be the relative entropy of the two-way divergence of the triplet vectors. The normalized probability of the feature vector of the i-th triplet of the input document. Let n be the normalized probability of the i-th triplet feature vector in the knowledge graph, and n be the number of triplet feature vectors in the input document. The value of n should be less than or equal to the number of triplet clusters in the industry fusion knowledge graph.

[0063] In this application, the triplet features are used to perform probability distribution. The triplet features are then normalized using softmax to map the vectors to a distribution where each dimension is between 0 and 1. Finally, the relative entropy of the bidirectional divergence between the two triplet vectors is calculated. To characterize the degree of difference between the probability distributions of two word vectors, the smaller the relative entropy of the semantic divergence of word vectors, the more consistent the distribution, and the higher the similarity of the relative entropy of the semantic divergence of word vectors. The triple feature vectors extracted from the input document and the triple feature vectors with the largest bidirectional divergence relative entropy of the knowledge graph triple cluster are selected as the key semantic parsing feature vectors.

[0064] Traditional word vector similarity methods often use cosine similarity and Euclidean distance, which only measure spatial geometric distance and ignore the disorder of probability distribution and information uncertainty within the vectors, thus limiting their adaptability. Utilizing entropy to calculate word vector similarity allows for the measurement of semantic similarity between two word vectors from the perspective of information distribution consistency. This aligns better with the semantic representation logic of word vectors in large models and is highly suitable for the document keyword and business word vector matching scenarios described in this patent.

[0065] Furthermore, the network relational features of the official document text are obtained by taking the average of the feature vector centrality of the input document triples and the industry-integrated knowledge graph. The formula for calculating the feature vector centrality value of one of the input document triples and the industry-integrated knowledge graph is as follows:

[0066]

[0067] In the formula, The centrality of one of the triples in the input document with the feature vector of the industry-integrated knowledge graph. The maximum eigenvalue of the weight matrix, also known as the normalization coefficient, is determined according to the number of nodes in the industry fusion knowledge graph. In this embodiment, if the number of nodes in the industry fusion knowledge graph exceeds 100, the value is 50; otherwise, the value is 5. To traverse all other nodes h of the industry fusion knowledge graph, The service co-occurrence weights between node h and its neighbor node t. It is the centrality score of the neighbor node t itself.

[0068] Compared to existing keyword vectors and entropy similarity semantic features, which can only express the literal meaning of text, the network relational features of official document texts are obtained by averaging the feature vector centralities of the input document triples and the industry-integrated knowledge graph. This introduces the global graph topological influence, allowing the model to process and acquire more classification meanings simultaneously, such as what the document says and how important it is in the entire government business network. This eliminates the deficiency of pure text modeling in identifying major supervisions and cross-departmental core businesses. Feature vector centrality is a scoring indicator of the importance of nodes in a graph network. The importance of a node is not only determined by the number of nodes it connects to (degree centrality), but also by the importance of its connected neighbors. Feature vector centrality values ​​can be used to characterize the importance of network nodes. Ordinary degree centrality only counts the number of connections, while feature vector centrality can identify a "few key hub nodes." For example, a document that is simultaneously connected to high-level urgent instructions and leadership supervision documents, even if it has few associated documents, still has a very high centrality score and is judged as a core document. There is a transmission relationship in the circulation, drafting, and dissemination of official documents. Important documents will generate a large number of forwardings, assistance, and replies. Eigenvector centrality measures the propagation weight of a document within the entire document circulation network. This metric can be used as graph feature input to improve the accuracy of classification / hierarchy models. The centrality score of each document is used as a one-dimensional feature, concatenated with word vectors, identity vectors, and urgency vectors, and then fed into classification, recommendation, and intelligent drafting models to distinguish between ordinary circulated documents and those requiring supervision, urgent requests, or those requiring the signature of the top leader.

[0069] Furthermore, before constructing the intelligent document processing model based on the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature, and the corresponding intelligent document processing method, the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature are normalized and then spatial feature vectors are concatenated to obtain the intelligent document processing feature. The intelligent document processing feature is then used as the input feature for constructing the intelligent document processing model.

[0070] The general word vectors are trained on general news and web page corpora to distinguish semantic differences in government professional terms such as "draft for handling," "submit for review," "supervise," "co-sign," "main recipient," "copy to," "special mention," and "official telegram" in official documents. The specialized drafting word vectors are word embeddings specifically designed for official document drafting scenarios, retrained / fine-tuned using private government business data. The specialized drafting word vectors, generated based on underlying business data, integrate document identity and urgency vectors. This underlying business data includes the sources of acquired document data, including: existing document databases: original texts of received, issued, requested, approved, and correspondent documents from previous years; historical drafting work order data: historical manual drafting opinions, responsible departments, handlers, and distribution results for each document; business metadata: receiving unit, document type, processing time limit, historical cooperation records, and leadership instructions; and a government terminology dictionary: document-specific terminology, organization names, functional entries, and industry-specific terms.

[0071] The steps for generating specialized draft word vectors include cleaning the business corpus, removing meaningless characters, garbled characters in tables, and blank draft forms, and standardizing official document terminology. In this application, "top leader" is standardized as "principal leader," and "urgent / urgent document" is standardized with uniform tags.

[0072] Secondly, incremental training / fine-tuning of domain-specific corpora is performed. BERT / Word2Vec is pre-trained using general methods, and then fine-tuned a second time using a million-level official document business corpus. The training objective is not only semantic similarity, but also includes an auxiliary task for proposed classification: input text, predicting the historical responsible unit, urgency level, and type of proposed opinion.

[0073] By leveraging the underlying work order annotation data for supervision and constraints, keywords frequently processed by the same functional department are grouped closer together in the vector space, while the vectors of words corresponding to "urgent," "expedited," and "routine" are differentiated. Keywords related to leadership instructions are separated from those circulated in general by vector clustering. Each processed word, phrase, and entire document is mapped to a fixed-length dense vector, specifically adapted to intelligent assignment and recommendation models.

[0074] The document identity is integrated by combining conventional rule encoding with fully connected mapping. First, one-hot encoding / ordered numerical scoring is performed, such as assigning 5 points to the signature of the chief officer and 1 point to the ordinary employee. Then, it is compressed into a fixed-length identity vector through two layers of fully connected network to obtain the document identity vector.

[0075] This invention unifies and normalizes the following features: key semantic parsing feature vectors obtained through bidirectional divergence relative entropy matching; document text network relational features constructed based on the mean of feature vector centrality from a knowledge graph; specialized drafting word vectors generated from underlying business data; and dynamic process network urgency features obtained by fusing document identity and urgency vectors. This eliminates differences in the dimensions and numerical ranges of various features, and then concatenates them into a fusion feature set, which is input into the intelligent document processing model for training. Normalization addresses the issue of inconsistent scales among multi-source heterogeneous features, ensuring stable model convergence. Simultaneously, it integrates four complementary types of information: semantic distribution, business topology weights, real business data, and process timeliness. This allows the model to fully learn the comprehensive relational patterns of document text content, business level, data foundation, and processing time, overcoming the limitations of single-text feature modeling. Ultimately, this enables the accurate output of dynamic processing solutions tailored to each document, achieving intelligent approval decision support. After unifying and normalizing four types of multi-dimensional heterogeneous features, the model is fused to eliminate the interference of different feature numerical magnitudes, thereby improving the stability and convergence efficiency of model training. At the same time, it integrates multi-dimensional information such as semantics, graph topology, business data, and process timeliness to comprehensively characterize the comprehensive attributes of official documents, thereby significantly improving the accuracy of dynamic process adaptive adjustment and approval auxiliary decision-making.

[0076] Furthermore, the document intelligent processing model adopts an ensemble neural network model improved based on overlapping training set data. This ensemble neural network model is obtained by integrating K sub-neural network models. In the K sub-neural network models trained on each sub-training set, the proportion of overlapping data in the training set of the Kth sub-neural network model is [missing information]. ,in Less than or equal to 0.3 and If the value is greater than 0, then the loss function of the Kth sub-neural network model is:

[0077]

[0078] In the formula, Let $\frac{ ... To fix the overlap ratio of the training set data in the training set of the Kth sub-neural network model, The total number of dimensions of the output of the k-th sub-neural network. Let be the predicted value of the k-th sub-neural network for the s-th index vector of sample a. The k-th sub-neural network is used to manually label the s-th item of sample a with the true label value. For example... Figure 1 This is a schematic diagram of the integrated neural network of this application, where the hidden layers are corresponding parallel sub-neural networks, and the final output result is obtained by integrating the results of each sub-neural network.

[0079] In this embodiment, the integrated neural network model comprises K=5 parallel sub-neural networks. Each sub-network uses the following training hyperparameters: the AdamW optimizer is selected, the initial learning rate is set to 0.0001, the minimum learning rate lower bound is 0.0000001, the weight decay coefficient is 0.0005, and the gradient clipping threshold is 1.0; the training batch size is 128, the overlap ratio of the training sets of each sub-network is 0.2, and the number of overlapping samples in a single batch is 26; the total number of training rounds is set to 150, with an early stopping mechanism enabled, and training is terminated if the loss does not decrease after 15 consecutive rounds; each sub-network uses three fully connected hidden layers with dimensions of 448, 256, and 128 respectively, the Dropout probability of each layer is 0.2, the activation function is ReLU, and the model output dimension is 8; all input features are normalized to [0,1] using Min-Max, and the prediction results of multiple sub-networks are integrated and fused using an equal-weighted average. The document intelligent processing methods obtained from the prediction results of each sub-network include annotation at a certain location in the document, as well as the setting of the number of processes and the personnel involved in the approval process. The above document intelligent processing methods and / or document intelligent dynamic processing methods are all set according to the values ​​of the document intelligent processing model. This is a setting made by technical personnel based on the characteristics of the input and output data, and does not represent a unique setting; therefore, detailed settings will not be elaborated here. Figure 4 The diagram shows the principle of the neural network loss function.

[0080] Traditional fixed adjacent batches retain overlapping document samples, the sample distribution between batches will not be completely separated, there is a positive covariance in the batch gradient, and the gradient variance is significantly reduced; The interval constraints avoid excessive overlap, which can lead to sample duplication and redundancy, and repeated amplification of noise; at the same time, they retain the basic ability to smooth gradients.

[0081] The gradient fluctuation of the loss function in this application is reduced, the loss decrease curve of each backpropagation round is smooth, there is no need to frequently reduce the learning rate, and the loss converges to a steady-state low value faster under the same number of iterations.

[0082] Compared to traditional non-overlapping batches, where samples are completely mutually exclusive, leading to batches consisting entirely of ordinary documents and batches entirely of supervisory documents, resulting in drastic gradient jumps and difficulty in converging loss oscillations, this invention utilizes the coupling theory of overlap ratio and loss function. By adjusting the overlap ratio, it adaptively changes the learning intensity of the loss function for important documents, balancing training stability with the fitting effect of long-tail samples. Increasing the overlap ratio increases the frequency of important documents appearing in multiple batches, enhancing the expected contribution of such samples to the global loss and shifting the focus of the loss function fitting towards methods for handling complex business documents. The training sample distribution of multiple sub-networks tends to be consistent, the average loss of each branch is balanced, and the stability of the prediction results is significantly improved after integration and fusion. Furthermore, this overlap interval scheme is adaptable to the private computing power of government intranets, reducing training hardware costs and making it more practical for engineering implementation.

[0083] It also provides a document processing system that integrates industry knowledge graphs and dynamic workflow networks. This system implements a document processing method that integrates industry knowledge graphs and dynamic workflow networks. The system includes an industry-integrated knowledge graph module, an underlying digital foundation data platform module, a workflow engine module, a document intelligent processing model construction module, and a document intelligent dynamic processing module.

[0084] The industry-integrated knowledge graph module is used to extract and determine exclusive semantic parsing triples based on a large language model combined with a dedicated industry lexicon and business database, and to construct an industry-integrated knowledge graph.

[0085] The underlying digital base data platform module is used to generate specialized word vectors in conjunction with the underlying digital base data platform.

[0086] The workflow engine module is used to form a dynamic process network with urgency characteristics by combining the document identity vector and the urgency vector with the workflow engine.

[0087] The document intelligent processing model construction module is used to obtain the key semantic parsing feature vectors of the triple clusters formed with the input documents in the knowledge graph based on entropy similarity extraction technology, and to calculate and obtain the document text network relationship features. Based on the key semantic parsing feature vectors, the document text network relationship features, the specialized drafting word vectors, the dynamic process network urgency features, and the corresponding document intelligent processing methods, a document intelligent processing model is constructed.

[0088] The document intelligent dynamic processing module is used to output a document intelligent dynamic processing method based on the document intelligent processing model when a document is input, and to perform document intelligent processing to assist in approval decision-making based on the document intelligent dynamic processing method. Figure 3The diagram shows a document processing system that integrates industry knowledge graphs and dynamic process networks.

[0089] Furthermore, the specialized proposed word vector is composed of the latent intent features of official documents and the number of proposed approval opinions in historical official documents retrieved from the underlying digital base data platform based on the exclusive semantic parsing triplet.

[0090] The data source for constructing specialized drafting word vectors consists of two parts: first, semantic features of latent intent in official documents extracted from the underlying digital platform, using dedicated semantic parsing triples as retrieval keys; and second, frequency features of various types of proposed approval opinions in historical official documents of similar business obtained from platform statistics. After normalizing and merging these two types of features, specialized drafting word vectors for intelligent document drafting tasks are formed. The latent intent features of official documents belong to deep semantic features, capturing implicit semantics such as document requests, processing purposes, and business constraints. The frequency of historical proposed approval opinions belongs to statistical prior features; the higher the frequency, the stronger the fit of the proposed opinion, serving as a vector weight term to improve the relevance of intelligent drafting and automatic drafting results.

[0091] In one embodiment of this application, if the exclusive semantic parsing triple is (5, 0.3, 3), it indicates that the keyword of the official document is "the Municipal Housing and Construction Bureau leading and coordinating with the sub-district offices." Based on this triple, the data platform clusters and recalls a total of 327 historical official documents of the same type in recent years, namely "the Municipal Housing and Construction Bureau leading and coordinating with the sub-district offices." The frequency of historical proposed handling and proposed approval opinions at all approval nodes is statistically analyzed to form frequency features, specifically:

[0092] (1) Frequently requested opinion 1: "After verification, the project is found to be true. It is recommended that the Finance Bureau review the project and submit it to the leadership for approval" (appeared 216 times).

[0093] (2) Frequently requested opinion 2: "The materials are complete and meet the funding application specifications. It is recommended to approve the project" (appeared 189 times).

[0094] (3) Low-frequency comments: "The budget needs to be supplemented with supporting materials before review" (appeared 23 times).

[0095] The corresponding specialized proposed word vector is composed of (proposed opinion 1, proposed opinion 2, proposed opinion 3, frequency 1, frequency 2, frequency 3), where the frequency is the number of times the corresponding proposed opinion appears. In the above embodiment, it is (2, 6, 9, 216, 189, 23). The corresponding proposed opinions are also transformed into feature vector values ​​to adapt to the construction of the subsequent intelligent document processing model.

[0096] The specialized word vectors selected in this application differ from general pre-trained word vectors. These vectors integrate prior statistical data on industry-specific documents, thus addressing the shortcomings of general vectors in adequately adapting to the intent and approval opinions of specialized documents in official documents.

[0097] Furthermore, the approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

[0098] In this embodiment, the document identification vector is obtained based on the unit with the highest level and the most frequent occurrences in the document, while the urgency vector is set accordingly based on the document level. The workflow engine extracts the units corresponding to the corresponding feature values ​​in the document, extracting the feature values ​​of the highest-level unit and the unit mentioned most frequently. The approval workflow path parameters are obtained by concatenating the aforementioned feature values ​​and the urgency vector.

[0099] Official documents often involve multiple departments and levels of stakeholders during processing. The administrative level of different units determines the approval level, and the frequency of unit mentions indicates the core responsible party for the matter. This invention abandons the single keyword matching method and integrates two dimensions of features: administrative level weight priority and text mention frequency priority, ensuring accurate identification of the document's sponsoring entity, jurisdictional level, and review scope. An urgency vector differentiates between ordinary, urgent, and special urgency levels, used to control node skipping, parallel weaving, and timeliness supervision weights in the process chain. The approval flow path parameters obtained by concatenating these three types of features serve as the core input parameters for the dynamic process weaving engine to automatically generate approval topology, routing branches, and process priorities.

[0100] Taking the official document "Request for 550,000 Yuan in Infrastructure Repair Funds for XX Street" as an example:

[0101] First, the entire document undergoes entity identification. The document mentions the following entities in sequence: Subdistrict Office, District Finance Bureau, District Government, and the Deputy District Head in charge. Based on a preset administrative level weighting system, the system determines the highest-level unit to be the District Government (district level) and extracts its administrative level feature vector. Simultaneously, the frequency of entity mentions in the entire document is counted. The Subdistrict Office is the most frequently mentioned business entity in this document, and its high-frequency entity feature vector is extracted. These two types of vectors are merged to generate a unique document identity vector. Further, the document's urgency label is identified. If the document is a routine matter, a corresponding low-priority urgency vector is generated according to the level mapping rules. If it is an urgent or super-urgent document, a high-weight urgency vector is generated to trigger the green channel and overdue expedited supervision mechanism. The workflow engine concatenates the feature values ​​of the highest-level unit, the most frequent unit, and the urgency vector to obtain complete approval flow path parameters.

[0102] Therefore, the beneficial effects of this invention are that it utilizes entropy characteristics to calculate word vector similarity, which can measure the semantic similarity of two word vectors from the dimension of information distribution consistency. This is more in line with the semantic representation logic of word vectors in large models and is highly suitable for the document keyword and business word vector matching scenarios of this patent. Compared with the construction of traditional non-overlapping batch neural network models, this invention utilizes the coupling theory of overlap ratio and loss function. By adjusting the overlap ratio, it adaptively changes the learning intensity of the loss for important documents, balancing training stability and long-tail sample fitting effect. Increasing the overlap ratio increases the frequency of important documents appearing in multiple batches, improving the expected contribution of such samples to the global loss, and shifting the focus of the loss function fitting towards the complex business document processing method. The training sample distribution of multiple sub-networks tends to be consistent, the average loss of each branch is balanced, and the stability of the prediction result is greatly improved after integration and fusion. At the same time, this overlap interval scheme is suitable for the private computing power of government intranets, reducing training hardware costs and making it more feasible for engineering implementation. This solution achieves deep coupling between industry knowledge base and semantic parsing, linkage between business data and proposed opinions, and adaptive workflow driven by semantic analysis results, breaking through the limitations of traditional technologies such as shallow semantic understanding, templated content, and fixed processes. This solution boasts high semantic recognition accuracy, opinion content closely aligned with actual business operations, and strong process adaptability, significantly improving the overall efficiency of official document processing. It can be widely applied to government and state-owned enterprise document processing scenarios such as market supervision and environmental protection.

[0103] like Figure 5 As shown, the document processing method and system of the present invention, which integrates industry knowledge graphs and dynamic process networks, runs through conventional components of a computing device, including a processor, memory, communication interface, and bus.

[0104] Of course, it is understood that each embodiment of the present invention can achieve one of the effects individually, and the combination of multiple embodiments of the present invention can achieve all the above effects. However, it is not required that each embodiment of the present invention achieve all the above advantages and effects, because each embodiment of the present invention can constitute a separate technical solution and make one or more contributions to the prior art. Each embodiment of the present invention is not affected, and the solution can still be implemented even if one embodiment is deleted.

[0105] For any module structures not specifically defined in this invention, the existing technical descriptions shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered part of this invention and used to understand the meaning of certain technical features or parameters.

Claims

1. A document processing method that integrates industry knowledge graphs and dynamic process networks, characterized in that: The method includes: Based on a large language model combined with a dedicated industry lexicon and business database, we extract and determine dedicated semantic parsing triples and construct an industry-integrated knowledge graph. By linking with the underlying digital foundation data platform, specialized word vectors for drafting are generated, and by relying on the workflow engine and combining document identity vectors and urgency vectors, dynamic process network urgency features are formed. Based on entropy-based similarity extraction technology, key semantic parsing feature vectors of triplet clusters formed with input documents in the knowledge graph are obtained, and document text network relational features are calculated and obtained. Based on the key semantic parsing feature vectors, document text network relational features, specialized drafting word vectors, dynamic process network urgency features, and corresponding document intelligent processing methods, a document intelligent processing model is constructed. Based on the aforementioned document intelligent processing model, the document intelligent dynamic processing method is output when the document to be processed is input, and document intelligent processing for approval decision assistance is performed according to the aforementioned document intelligent dynamic processing method.

2. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 1, characterized in that: The specialized proposed word vector is composed of the potential intent features of official documents and the number of proposed approval opinions in historical official documents, retrieved from the underlying digital base data platform based on the exclusive semantic parsing triple.

3. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 1, characterized in that: The approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

4. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 1 or 3, characterized in that: The entropy-based similarity extraction technique uses bidirectional divergence relative entropy to extract the key semantic parsing feature vector. The bidirectional divergence relative entropy is obtained through knowledge graph and triple cluster calculation. The triple cluster is obtained by processing the input document. All triples are extracted from the input document to obtain the triple cluster. The triple with the highest bidirectional divergence relative entropy with the knowledge graph in the triple cluster is used as the key semantic parsing feature vector of the intelligent document processing model.

5. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 1, 2, or 3, characterized in that: The network relationship features of the official document text are obtained by taking the average value of the feature vector centrality of the input document triples and the industry fusion knowledge graph.

6. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 5, characterized in that: Before constructing the intelligent document processing model based on the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature, and the corresponding intelligent document processing method, the key semantic parsing feature vector, the document text network relational feature, the specialized drafting word vector, and the dynamic process network urgency feature are normalized and then spatial feature vectors are concatenated to obtain the intelligent document processing feature. The intelligent document processing feature is used as the input feature for constructing the intelligent document processing model.

7. The document processing method integrating industry knowledge graphs and dynamic process networks as described in claim 6, characterized in that: The document intelligent processing model adopts an ensemble neural network model improved based on overlapping training set data. This ensemble neural network model is obtained by integrating K sub-neural network models. In the K sub-neural network models trained on each sub-training set, the proportion of overlapping data in the training set of the Kth sub-neural network model is [missing information]. ,in Less than or equal to 0.3 and Greater than 0.

8. A document processing system integrating industry knowledge graphs and dynamic workflow networks, the system implementing the method described in claim 1, the system comprising an industry-integrated knowledge graph module, an underlying digital foundation data platform module, a workflow engine module, a document intelligent processing model construction module, and a document intelligent dynamic processing module, characterized in that: The industry-integrated knowledge graph module is used to extract and determine exclusive semantic parsing triples based on a large language model combined with a dedicated industry lexicon and business database, and to construct an industry-integrated knowledge graph. The underlying digital base data platform module is used to generate specialized word vectors in conjunction with the underlying digital base data platform. The workflow engine module is used to form a dynamic process network with urgency characteristics by combining the document identity vector and the urgency vector with the workflow engine. The document intelligent processing model construction module is used to obtain the key semantic parsing feature vectors of the triple clusters formed with the input documents in the knowledge graph based on entropy similarity extraction technology, and to calculate and obtain the document text network relationship features. Based on the key semantic parsing feature vectors, the document text network relationship features, the specialized drafting word vectors, the dynamic process network urgency features, and the corresponding document intelligent processing methods, a document intelligent processing model is constructed. The document intelligent dynamic processing module is used to output the document intelligent dynamic processing method when the document to be processed is input based on the document intelligent processing model, and to perform document intelligent processing to assist approval decision-making based on the document intelligent dynamic processing method.

9. The document processing system integrating industry knowledge graphs and dynamic process networks as described in claim 8, characterized in that: The specialized proposed word vector is composed of the potential intent features of official documents and the number of proposed approval opinions in historical official documents, retrieved from the underlying digital base data platform based on the exclusive semantic parsing triple.

10. The document processing system integrating industry knowledge graphs and dynamic process networks as described in claim 8, characterized in that: The approval workflow path parameters are formed by combining the document identity vector and the urgency vector with the workflow engine.

Images