Standard file auxiliary writing method and system based on artificial intelligence

Abstract

The invention discloses a standard file auxiliary writing method and system based on artificial intelligence, and the method comprises the steps: obtaining a standard file corpus, carrying out the analysis of the standard file corpus, extracting standard terms, and constructing the knowledge representation of the standard terms; constructing a standard clause association graph based on the standard clause knowledge representation, and representing the association relationship between the standard clauses; applying a graph mining algorithm to perform structured analysis on the association graph, and identifying key terms and term association paths; based on the association graph and the structured analysis result, constructing an intelligent auxiliary writing model, receiving user input, performing standard term recommendation, automatic content generation and consistency check, and assisting the user to complete standard file writing; by means of the knowledge graph and the deep learning technology, intelligent analysis and auxiliary compiling of the standard file are achieved, the compiling efficiency and quality of the standard file are improved, and the problems that in a traditional standard compiling method, efficiency is low, and the quality is difficult to guarantee are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of standard document writing, and in particular relates to an artificial intelligence-based auxiliary writing method and system for standard documents. Background Art

[0002] At present, the preparation of standard documents mainly relies on manual work, which poses many problems and challenges. First, the number of standard documents is huge and the structure is complex, so the compilers need to spend a lot of time to consult and refer to relevant standards. Second, there are complex references and correlations between different standards, which are difficult to fully grasp manually, and easily lead to inconsistencies between standards. Third, the standard clauses have high requirements for standardization and accuracy, and manual writing is prone to problems such as inaccurate expressions and inconsistent use of terminology. Fourth, with the accelerated frequency of standard updates in various industries, the workload of standard writing has increased sharply, and it is difficult to meet the needs of efficient writing by relying solely on manual methods.

[0003] Existing technologies to assist with standard writing primarily include template-based editing tools, keyword-based search systems, and simple consistency checking tools. These technologies suffer from limitations such as single functionality, low intelligence, and an inability to connect knowledge. Template-based editing tools offer only fixed formats and lack intelligent support for content; keyword-based search systems rely on precise search terms and fail to understand semantic similarities; and simple consistency checking tools can only perform superficial syntax and formatting checks, failing to verify deeper semantic and logical consistency.

[0004] These technical means cannot effectively solve the core problems in the preparation of standard documents, and there is an urgent need to develop intelligent standard document auxiliary writing technology. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an artificial intelligence-based standard document auxiliary writing method and system. Through knowledge graph technology and deep learning models, it realizes intelligent analysis, association mining and auxiliary writing of standard clauses, improves the efficiency and quality of standard document writing, reduces the workload of writers, and ensures the consistency and standardization of standard documents.

[0006] In a first aspect, the present invention provides an artificial intelligence-based standard document auxiliary writing method, the method comprising the following steps:

[0007] Step S1: Obtain a standard document corpus, analyze the standard document corpus, extract standard clauses and construct a standard clause knowledge representation.

[0008] Step S2: constructing a standard clause association graph based on the standard clause knowledge representation, wherein the association graph is used to represent the association relationship between standard clauses.

[0009] Step S3: Apply a graph mining algorithm to perform structured analysis on the association graph to identify key terms and term association paths.

[0010] Step S4, based on the association graph and structured analysis results, build an intelligent auxiliary writing model, receive user input, recommend standard terms, automatically generate content and perform consistency verification based on the user input, and assist the user in completing the standard document writing.

[0011] Furthermore, the analysis of the standard document corpus, extraction of standard clauses and construction of standard clause knowledge representation specifically includes:

[0012] Get the standard document corpus D = {d1, d2, ...d i ,…,d n}, where d i Represents the i-th standard file, preprocesses the standard file, identifies the clause boundaries and obtains the boundary identification set B of the i-th standard file i :

[0013] B i ={(s1,e1),(s2,e2),…,(s m ,e m )},s m and e m Respectively represent standard files d i The starting and ending positions of the mth clause identified in .

[0014] Furthermore, according to the identified clause boundary set B i , from the standard document d i Extract the corresponding clause text set C i :

[0015] C i ={c i1 ,c i2 ,…c ij ,…c im}; For each clause c ij Perform attribute recognition to extract clause type, scope of application and mandatory level attribute information. The attribute recognition formula is A ij =M a (c ij ), where M a Represents the clause attribute recognition model, A ij Indicates terms c ij A collection of properties.

[0016] Use pre-trained language model to train standard terms c ij Encode and get the vector representation v of the terms ij,in represents a k-dimensional vector.

[0017] Furthermore, the pre-trained language model is used to train the standard terms c ij The encoding is based on the Transformer architecture, expressed as:

[0018] h l =Transformer l (h l-1 ),l∈{1,2,…,L}, where h l Represents the hidden state of the lth layer, L represents the number of model layers; use the trained language model for each term c ij Encoding is performed; the training phase uses a large-scale corpus in the standard document field and incorporates the design of clause relationship prediction and clause type classification tasks.

[0019] Furthermore, the construction of a standard clause association graph based on the standard clause knowledge representation specifically includes:

[0020] Calculate any two terms c ij with c pq The semantic similarity sim(v ij ,v pq ), the calculation formula is:

[0021] Among them, v pq c pq The vector representation of .

[0022] Identify the reference relationship between terms and construct the reference relationship set Cite: Cite = {(c ij ,c pq ,w ij→pq )}, where c ij is the reference clause, c pq Indicates the referenced clause, w ij→pq Represents the citation weight, which is calculated based on the citation frequency. The formula is: where f ij→pq Indicates terms c ij Reference clause c pq Frequency; ∑ x,y f ij→xy Indicates terms c ij Total number of citations.

[0023] Analyze the co-occurrence of terms in the application scenario and calculate the co-occurrence intensity coc(c ij ,c pq ):

[0024] Where P(cij ,c pq ) represents the probability of co-occurrence of terms, P(c ij ) and P(c pq ) represent the probability of the two terms appearing alone.

[0025] Furthermore, based on the above three relationships, the comprehensive correlation between the terms r(c ij ,c pq ):

[0026] r(c ij ,c pq )=0.4sim(v ij ,v pq )+0.3w ij→pq +0.3coc(c ij ,c pq );

[0027] Construct a clause association graph G = (V, E, W), where V is a node set representing standard clauses, E is an edge set representing the association relationship between clauses, and W is an edge weight set, where the weight value is equal to the corresponding comprehensive association degree.

[0028] Furthermore, we construct a correlation subgraph for the clauses in different standard systems and calculate the cross-standard correlation between the clauses in different standard systems. in Indicates clause c in standard system A i , Indicates clause c in standard system B j , and denote their respective vector representations, It indicates the strength of the cross-standard reference relationship between two clauses. Based on the cross-standard correlation, the association sub-graphs of different standard systems are connected to form a complete standard clause association graph.

[0029] Furthermore, applying a graph mining algorithm to perform a structured analysis on the association graph to identify key terms and terms association paths includes: using a graph neural network to perform representation learning on the standard terms association graph to obtain enhanced node representations. Where L represents the number of network layers.

[0030] Based on the node representation, the standard terms are divided into several categories using clustering algorithms, and the centrality index of each node in the graph is calculated, including degree centrality Closeness Centrality and betweenness centrality, where d(v) represents the degree of node v, |V| represents the total number of nodes, and d(v,u) represents the shortest path length from node v to node u.

[0031] Calculate the term importance score Imp(v) based on the centrality index:

[0032] Imp(v)=0.4·C D (v)+0.3·C C (v)+0.3·C B (v), where C B (v) represents betweenness centrality. Based on the importance score, the top three terms are selected as key terms, and the association paths between them are analyzed as key association paths.

[0033] Furthermore, based on the association graph and the structured analysis results, an intelligent auxiliary writing model is constructed to receive the user input demand q and map it to the vector space to obtain the demand vector v q , retrieve the clause set Rel(q) related to the demand semantics in the association graph: Rel(q) = {c i |sim(v q ,v i )>θ}, where θ represents the similarity threshold, θ=0.5; calculate the recommendation score of the terms Among them, δ represents the weight of graph structure information, δ = 0.3.

[0034] Based on the recommendation scores, a ranked list of recommended terms is generated; based on the recommended terms and user needs, standard text content is generated; the terminology consistency, logical consistency and structural consistency of the standard content are tested, and a consistency report is generated.

[0035] Furthermore, the generation of standard text content based on recommended terms and user needs includes: hierarchical generation of standard framework generation, chapter structure generation and term content generation; selecting the template most similar to the user needs from a pre-established standard template library as the basic framework; and normalizing the use of terms in the generated text content to ensure terminology consistency.

[0036] Furthermore, the Transformer architecture and the number of network layers L of the graph neural network are consistent, and L is not less than 3.

[0037] In the second aspect, based on the same inventive concept, the present invention provides an artificial intelligence-based standard document auxiliary writing system for executing the method of the first aspect, the system comprising: a standard analysis module for acquiring a standard document corpus, analyzing the standard document corpus, extracting standard clauses and constructing a standard clause knowledge representation; a graph construction module for constructing a standard clause association graph based on the standard clause knowledge representation, the association graph being used to characterize the association relationship between standard clauses; a graph analysis module for applying a graph mining algorithm to perform structured analysis on the association graph to identify key clauses and clause association paths; an intelligent assistance module for constructing an intelligent auxiliary writing model based on the association graph and the structured analysis results, receiving user input, and performing standard clause recommendations, automatic content generation and consistency verification based on the user input to assist users in completing standard document writing.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention constructs a structured knowledge representation of standard clauses by performing in-depth semantic analysis and knowledge extraction on the standard document corpus, constructs a standard clause association graph based on the standard clause knowledge representation, and uses a graph mining algorithm to perform structured analysis on the association graph. It can intelligently identify key clauses and association paths, and provide users with the most valuable standard reference content. Compared with traditional retrieval systems, this method can infer indirectly related but important reference value clauses based on the graph structure, can perform personalized clause recommendations and content generation according to user input requirements, and perform multi-dimensional consistency verification. It significantly improves the efficiency and quality of standard document writing, especially when dealing with complex standard writing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a standard document auxiliary writing method based on artificial intelligence of the present invention;

[0041] Figure 2 This is a schematic diagram of the composition of an artificial intelligence-based standard document auxiliary writing system of the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only part of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Example 1

[0044] like Figure 1FIG. 1 is a method for assisting in writing standard documents based on artificial intelligence according to the present invention, wherein the method comprises the following steps:

[0045] Step S1: Obtain a standard document corpus, analyze the standard document corpus, extract standard clauses and construct a standard clause knowledge representation.

[0046] A standards document corpus is a data set containing a large number of standard documents, which can come from various sources such as national standards, industry standards, group standards, or enterprise standards. For example, in the field of information technology, it can include GB / T 22239 Information Security Technology Series Standards and ISO / IEC 27000 Information Security Management System Series Standards.

[0047] The standard document corpus used in this embodiment contains more than 10,000 standard documents in different fields, covering multiple industries such as information technology, manufacturing, and service industries.

[0048] The analysis of the standard document corpus, extraction of standard clauses and construction of standard clause knowledge representation specifically includes:

[0049] Get the standard document corpus D = {d1, d2, ...d i ,…,d n}, where d i Represents the i-th standard file, preprocesses the standard file, identifies the clause boundaries and obtains the boundary identification set B of the i-th standard file i :

[0050] B i ={(s1,e1),(s2,e2),…,(s m ,e m )},s m and e m Respectively represent standard files d i The starting and ending positions of the mth clause identified in .

[0051] For example, for the standard document d1 "Basic Requirements for Information Security Technology and Network Security Level Protection", a BiLSTM-CRF-based sequence labeling model is used to identify clause boundaries, and the boundary identification set B1 = {(102, 250), (251, 390), (391, 520), ...} is obtained, where (102, 250) indicates that the 102nd to 250th characters in the document constitute a complete clause.

[0052] According to the identified terms boundary set B i , from the standard document d i Extract the corresponding clause text set C i :

[0053] C i ={c i1 ,c i2 ,…c ij ,…c im}; For each clause c ij Perform attribute recognition to extract clause type, scope of application and mandatory level attribute information. The attribute recognition formula is A ij =M a (c ij}, where M a Represents the clause attribute recognition model, A ij Indicates terms c ij A collection of properties.

[0054] For standard document d1, based on boundary identification set B1, the extracted clause text set C1 may include clauses such as "5.1.1 Network Architecture," "5.1.2 Communication Transmission," and "5.1.3 Boundary Protection." For clause "5.1.2 Communication Transmission," attribute identification is performed using clause attribute identification model Ma, resulting in its attribute set A12 = {Type: "Technical Requirements," Scope of Application: "Network Security," Mandatory Level: "Mandatory"}.

[0055] Use pre-trained language model to train standard terms c ij Encode and get the vector representation v of the terms ij ,in represents a k-dimensional vector.

[0056] The pre-trained language model is used to ij The encoding is based on the Transformer architecture, expressed as:

[0057] h l =Transformer l (h l-1 ),l∈{1,2,…,L}, where h l Represents the hidden state of the lth layer, L represents the number of model layers, L is not less than 3; use the trained language model for each clause c ij Encoding is performed; the training phase uses a large-scale corpus in the standard document field and incorporates the design of clause relationship prediction and clause type classification tasks.

[0058] The pre-trained language model uses the BERT (Bidirectional Encoder Representations from Transformers) architecture and is pre-trained and fine-tuned on a 100GB standard document corpus. For example, for clause c12 "5.1.2 Communication Transmission", the encoded vector representation v12∈R768, i.e., k=768, captures the semantic features of the clause and provides a basis for subsequent similarity calculation and graph construction.

[0059] The Transformer architecture uses a multi-head self-attention mechanism, with L set to 12 layers, a hidden layer dimension of 768, and 12 attention heads. In addition to the basic masked language model (MLM) task, two specific tasks are added during the training phase: (1) a clause relationship prediction task, which predicts whether two clauses have a reference, subordinate, or related relationship; and (2) a clause type classification task, which classifies clauses into different types such as definition, requirement, and test method.

[0060] Step S2: constructing a standard clause association graph based on the standard clause knowledge representation, wherein the association graph is used to represent the association relationship between standard clauses.

[0061] The Standard Clause Association Graph is a specialized knowledge graph whose nodes represent standard clauses and edges represent the relationships between clauses. By constructing such a graph, the clause knowledge scattered across different standard documents can be structured, revealing implicit connections between clauses and providing knowledge support for standard writing.

[0062] Constructing a standard clause association map specifically includes:

[0063] Calculate any two terms c ij with c pq The semantic similarity sim(v ij ,v pq ), the calculation formula is:

[0064] Among them, v pq c pq For example, for clause c12 "5.1.2 Communication Transmission" and clause c35 "6.2.1 Data Transmission Security", calculate the cosine similarity between their vector representations v12 and v35: It indicates that the two terms have a high semantic similarity.

[0065] Identify the reference relationship between terms and construct the reference relationship set Cite: Cite = {(c ij ,c pq ,w ij→pq )}, where cij is the reference clause, c pq Indicates the referenced clause, w ij→pq Represents the citation weight, which is calculated based on the citation frequency. The formula is: where f ij→pq Indicates terms c ij Reference clause c pq Frequency; ∑ x,y f ij→xy Indicates terms c ij Total number of citations.

[0066] Reference relationships can be identified by combining rule matching and pattern recognition. For example, when clause c12 contains the phrase "according to clause 6.2.1" or "see 6.2.1", it can be identified that c12 references c35. Assume that the frequency of c12 citing c35 is f12→35=2, and the total frequency of c12 citing all other clauses is ∑ x,y f12→xy=5, then the reference weight w12→35=2 / 5=0.4.

[0067] Analyze the co-occurrence of terms in the application scenario and calculate the co-occurrence intensity coc(c ij ,c pq ):

[0068] Where P(c ij ,c pq ) represents the probability of co-occurrence of terms, P(c ij ) and P(c pq ) represent the probability of the two terms appearing alone;

[0069] Based on the above three relationships, the comprehensive correlation between the terms r(c ij ,c pq ):

[0070] r(c ij ,c pq )=0.4sim(v ij ,v pq )+0.3w ij→pq +0.3coc(c ij ,c pq ); Continuing with the above example, the comprehensive correlation between clauses c12 and c35 is r(c12,c35)=0.4×0.82+0.3×0.4+0.3×0.1=0.478.

[0071] Construct a clause association graph G = (V, E, W), where V is a node set representing standard clauses, E is an edge set representing the association relationship between clauses, and W is an edge weight set, where the weight value is equal to the corresponding comprehensive association degree.

[0072] In this example, the constructed clause association graph G contains approximately 50,000 nodes (standard clauses) and over 200,000 weighted edges (clause association relationships). To improve the quality of the graph, only edges with an association greater than a threshold of 0.3 are retained, and weak associations are filtered out, making the graph more refined and effective.

[0073] Construct a correlation subgraph for clauses in different standard systems and calculate the cross-standard correlation between clauses in different standard systems in Indicates clause c in standard system A i , Indicates clause c in standard system B j , and denote their respective vector representations, It indicates the strength of the cross-standard reference relationship between two clauses. Based on the cross-standard correlation, the association sub-graphs of different standard systems are connected to form a complete standard clause association graph.

[0074] Cross-standard correlation analysis helps to discover the correspondence and compatibility points between different standard systems, facilitating the coordinated implementation and compatible writing of standards. For example, for clause ciA "A.14.1.2 Secure Communication Services" in ISO / IEC 27001 Information Security Management System and clause cjB "5.1.2 Communication Transmission" in GB / T 22239 Network Security Level Protection, the cross-standard correlation degree cross(ciA,cjB) is calculated to be 0.65, the semantic similarity of the two clauses is 0.75, and the cross-standard reference relationship strength of the two clauses is 0.5 (obtained by analyzing the standard mapping table or cross-reference document).

[0075] Step S3: Applying a graph mining algorithm to perform structured analysis on the association graph to identify key terms and term association paths;

[0076] Applying graph mining algorithms to perform structural analysis on the association graph to identify key terms and terms association paths includes: using graph neural networks to learn the representation of standard terms association graphs to obtain enhanced node representations. i (L) , where L represents the number of network layers. The Transformer architecture and the number of network layers L of the graph neural network are consistent.

[0077] The graph neural network used is a graph convolutional network (GCN), which learns from graph-structured data through a message-passing mechanism. Specifically, for a node vi in ​​the graph (representing a clause ci), its initial representation is the clause's vector representation vi. After passing through L layers of GCN, the enhanced node representation ziL is obtained. This representation not only contains the node's own characteristics but also incorporates information from its neighboring nodes, better reflecting the clause's position and role in the entire graph.

[0078] Based on the node representation, the standard terms are divided into several categories using clustering algorithms, and the centrality index of each node in the graph is calculated, including degree centrality Closeness Centrality and betweenness centrality, where d(v) represents the degree of node v, |V| represents the total number of nodes, and d(v,u) represents the shortest path length from node v to node u.

[0079] Using the K-means clustering algorithm, all clauses are divided into 10 categories based on the enhanced node representation (ziL) of the clauses, including: "definition category," "scope category," "technical requirements category," "management requirements category," and "test method category." For clauses in the cybersecurity standard, the calculated centrality metrics are as follows: The degree centrality (CD) for clause "5.1.2 Communication Transmission" is 0.015 (connected to 150 other clauses, with a total of 10,000 nodes), the closeness centrality (CC) is 0.48, and the betweenness centrality (CB) is 0.025.

[0080] Calculate the term importance score Imp(v) based on the centrality index:

[0081] Imp(v)=0.4·C D (v)+0.3·C C (v)+0.3·C B (v), where C B (v) represents betweenness centrality. Based on the importance score, the top three terms are selected as key terms, and the association paths between them are analyzed as key association paths.

[0082] These key clauses hold a core position within the entire standard system, interconnected with numerous other clauses and exerting a wide impact, necessitating significant consideration during standard development. By calculating the importance scores of 50,000 clauses, the three with the highest scores were selected as key clauses: "Information Security Management System Requirements," "Basic Requirements for Cybersecurity Multi-Level Protection," and "Basic Requirements for Data Security Management." The collection of clauses directly (path length 1) or indirectly (path length 2) linked to these key clauses constitutes a key linkage path, providing a reference framework for standard development.

[0083] Step S4, based on the association graph and structured analysis results, build an intelligent auxiliary writing model, receive user input, recommend standard terms, automatically generate content and perform consistency verification based on the user input, and assist the user in completing the standard document writing.

[0084] Based on the association graph and structured analysis results, an intelligent auxiliary writing model is constructed to receive the user's input demand q and map it to the vector space to obtain the demand vector v q , retrieve the clause set Rel(q) related to the demand semantics in the association graph: Rel(q) = {c i |sim(v q ,v i )>θ}, where θ represents the similarity threshold, θ=0.5; calculate the recommendation score of the terms Among them, δ represents the weight of graph structure information, δ = 0.3.

[0085] When a user enters a requirement q, "Security requirements related to data transmission encryption need to be formulated," the system first maps it into a vector space to generate a requirement vector vq. It then searches the association graph for clauses with a similarity greater than a threshold θ = 0.5, obtaining the set of related clauses Rel(q) = {c35 "6.2.1 Data transmission security", c12 "5.1.2 Communication transmission", c87 "7.1.3 Encryption management", ...}.

[0086] For each clause ci, its recommendation score is calculated. For example, the recommendation score of c35 is score(c35|q)=(1-0.3)0.85+0.3×0.68=0.8, where 0.85 is the semantic similarity between c35 and requirement q, and 0.68 is the recommendation score calculated based on the graph structure. Based on the recommendation score, the system generates a ranked clause recommendation list and recommends the clause with the highest score to the user first.

[0087] Based on the recommendation scores, a ranked list of recommended terms is generated; based on the recommended terms and user needs, standard text content is generated; the terminology consistency, logical consistency and structural consistency of the standard content are tested, and a consistency report is generated.

[0088] When the user selects the recommended terms, the system generates standard text content based on these terms and user needs; for example, if the user selects "6.2.1 Data Transmission Security" and "5.1.2 Communication Transmission" as references, the system will automatically generate new standard terms: "6.3.1 Data Transmission Encryption Requirements: Important data should be protected by encryption technology during transmission to ensure the confidentiality and integrity of data transmission. The encryption algorithm should comply with national password management requirements, and the key length should be no less than 128 bits."

[0089] The system will also check the generated content for terminology consistency, logical consistency, and structural consistency, and generate a consistency report to point out possible problems and improvement suggestions; for example, it detects that: "key length" should be changed to "key strength" to maintain terminology consistency with other terms.

[0090] The generation of standard text content based on recommended terms and user needs includes: hierarchical generation of standard framework generation, chapter structure generation and term content generation; selecting the template most similar to the user needs from a pre-established standard template library as the basic framework; and normalizing the use of terms in the generated text content to ensure terminology consistency.

[0091] During the hierarchical generation process, the system first generates a standard framework to determine the overall structure and main chapters of the document; then generates a detailed chapter structure based on the framework, including titles and clause numbers at all levels; and finally generates specific clause content.

[0092] For example, the standard framework generated for "Information System Security Design Specifications" includes major chapters such as "1. Scope", "2. Normative Reference Documents", "3. Terms and Definitions", "4. General Requirements", and "5. Technical Requirements".

[0093] When selecting a template, the system selects the "Information Security Design Technical Specifications" template that is most similar to the user's requirement "Data Security Design Specifications" as the basic framework from a library of more than 500 standard templates.

[0094] For terminology standardization, the system maintains a term library containing more than 5,000 professional terms and their standard definitions to ensure that the terminology used in the generated content complies with the standards, such as the unified use of "data transmission encryption" instead of "data encryption transmission" or "transmission data encryption".

[0095] Example 2

[0096] like Figure 2 Figure 2 shows a schematic diagram of the AI-based standard document authoring system. The system comprises a standard analysis module, a graph construction module, a graph analysis module, and an intelligent assistance module, all connected in sequence. The system utilizes a distributed architecture to support large-scale standard data processing and analysis, with flexible scaling of computing resources based on demand.

[0097] The standard analysis module is used to obtain the standard document corpus, analyze the standard document corpus, extract standard terms and construct the standard terms knowledge representation.

[0098] The standard analysis module consists of a data acquisition submodule, a text preprocessing submodule, a clause identification submodule, an attribute extraction submodule, and a vector encoding submodule. The data acquisition submodule is responsible for collecting standard documents from various sources, including national standard databases, industry standard libraries, and enterprise standard libraries. The text preprocessing submodule cleans, segments, and removes stop words from the original standard documents to improve the accuracy of subsequent analysis. The clause identification submodule uses a deep learning model to identify clause boundaries within standard documents. The attribute extraction submodule extracts attribute information such as the clause type and scope of application. The vector encoding submodule converts the clause text into a vector representation.

[0099] A graph construction module is used to construct a standard clause association graph based on the standard clause knowledge representation, and the association graph is used to represent the association relationship between standard clauses.

[0100] The graph construction module consists of a similarity calculation submodule, a citation relationship identification submodule, a co-occurrence analysis submodule, a correlation calculation submodule, and a graph storage submodule. The similarity calculation submodule is responsible for calculating the semantic similarity between clause vectors; the citation relationship identification submodule identifies citation relationships between clauses through rule matching and text analysis; the co-occurrence analysis submodule statistically analyzes the co-occurrence of clauses in application scenarios; the correlation calculation submodule combines the results of the first three submodules to calculate the comprehensive correlation between clauses; and the graph storage submodule uses graph database technology to store and manage the constructed correlation graph. This module also includes cross-standard correlation analysis, which can analyze the corresponding relationships between clauses in different standard systems.

[0101] The graph analysis module is used to apply a graph mining algorithm to perform structured analysis on the association graph and identify key terms and term association paths.

[0102] The graph analysis module includes a graph representation learning submodule, a cluster analysis submodule, a centrality calculation submodule, an importance assessment submodule, and a path analysis submodule. The graph representation learning submodule uses graph neural network technology to learn the structured representation of clause nodes; the cluster analysis submodule classifies clauses based on their node representations; the centrality calculation submodule calculates various centrality metrics to assess the importance of clauses in the graph; the importance assessment submodule uses a combination of multiple metrics to assess the overall importance of clauses; and the path analysis submodule analyzes the association paths between clauses and discovers indirect relationships between them.

[0103] The intelligent assistance module is used to build an intelligent assistance writing model based on the association map and structured analysis results, receive user input, recommend standard terms, automatically generate content and check consistency based on user input, and assist users in completing standard document writing.

[0104] The intelligent assistance module consists of a user interaction submodule, a demand analysis submodule, a clause recommendation submodule, a content generation submodule, and a consistency check submodule. The user interaction submodule provides a user-friendly interface, accepting user input and displaying system output. The demand analysis submodule performs semantic understanding and vector mapping of user input requirements. The clause recommendation submodule calculates and recommends relevant clauses based on the requirement vectors and association graphs. The content generation submodule uses a pre-trained language model to generate standard text content based on the recommended clauses and user requirements. The consistency check submodule performs multi-dimensional consistency checks on the generated content to ensure that it complies with standard writing specifications.

[0105] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A standard document auxiliary writing method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1, obtaining a standard document corpus, analyzing the standard document corpus, extracting standard clauses and constructing a knowledge representation of standard clauses; Step S2: constructing a standard clause association graph based on the standard clause knowledge representation, wherein the association graph is used to represent the association relationship between the standard clauses; Step S3: Applying a graph mining algorithm to perform structured analysis on the association graph to identify key terms and term association paths; Step S4, based on the association graph and structured analysis results, build an intelligent auxiliary writing model, receive user input, recommend standard terms, automatically generate content and perform consistency verification based on the user input, and assist the user in completing the standard document writing.

2. The artificial intelligence-based standard document auxiliary writing method according to claim 1 is characterized in that: The analysis of the standard document corpus, extraction of standard clauses and construction of standard clause knowledge representation specifically includes: Get the standard document corpus D = {d1, d2, ...d i ,…,d n }, where d i Represents the i-th standard file, preprocesses the standard file, identifies the clause boundaries and obtains the boundary identification set B of the i-th standard file i : B i ={(s1,e1),(s2,e2),…,(s m ,e m )},s m and e m Respectively represent standard files d i The start and end positions of the mth clause identified in ; According to the identified terms boundary set B i , from the standard document d i Extract the corresponding clause text set C i : C i ={c i1 ,c i2 ,…c ij ,…c im }; For each clause c ij Perform attribute recognition to extract clause type, scope of application and mandatory level attribute information. The attribute recognition formula is A ij =M a (c ij ), where M a Represents the clause attribute recognition model, A ij Indicates terms c ij The attribute set of Use pre-trained language model to train standard terms c ij Encode and get the vector representation v of the terms ij ,in represents a k-dimensional vector.

3. The artificial intelligence-based standard document auxiliary writing method according to claim 2, characterized in that: The pre-trained language model is used to ij The encoding is based on the Transformer architecture, expressed as: h l =Transformer l (h l-1 ),l∈{1,2,…,L}, where h l Represents the hidden state of the lth layer, L represents the number of model layers; use the trained language model for each term c ij Encoding; The training phase uses a large-scale corpus in the standard document domain and incorporates the design of clause relationship prediction and clause type classification tasks.

4. The artificial intelligence-based standard document auxiliary writing method according to claim 1, characterized in that: The construction of the standard clause association graph based on the standard clause knowledge representation specifically includes: Calculate any two terms c ij with c pq The semantic similarity sim(v ij ,v pq ), the calculation formula is: Among them, v pq c pq Vector representation of ; Identify the reference relationship between terms and construct the reference relationship set Cite: Cite = {(c ij ,c pq ,w ij→pq )}, where c ij is the reference clause, c pq Indicates the referenced clause, w ij→pq Represents the citation weight, which is calculated based on the citation frequency. The formula is: where f ij→pq Indicates terms c ij Reference clause c pq Frequency; ∑ x,y f ij→xy Indicates terms c ij Total number of citations; Analyze the co-occurrence of terms in the application scenario and calculate the co-occurrence intensity coc(c ij ,c pq ): Where P(c ij ,c pq ) represents the probability of co-occurrence of terms, P(c ij ) and P(c pq ) represent the probability of the two terms appearing alone; Based on the above three relationships, the comprehensive correlation between the terms r(c ij ,c pq ): r(c ij ,c pq )=0.4sim(v ij ,v pq )+0.3w ij→pq +0.3coc(c ij ,c pq ); Construct a clause association graph G = (V, E, W), where V is a node set representing standard clauses, E is an edge set representing the association relationship between clauses, and W is an edge weight set, where the weight value is equal to the corresponding comprehensive association degree.

5. The artificial intelligence-based standard document auxiliary writing method according to claim 4 is characterized in that: Construct a correlation subgraph for clauses in different standard systems and calculate the cross-standard correlation between clauses in different standard systems in Indicates clause c in standard system A i , Indicates clause c in standard system B j , and denote their respective vector representations, It indicates the strength of the cross-standard reference relationship between two clauses. Based on the cross-standard correlation, the association sub-graphs of different standard systems are connected to form a complete standard clause association graph.

6. The artificial intelligence-based standard document auxiliary writing method according to claim 1 is characterized in that: Applying graph mining algorithms to perform structured analysis on the association graph to identify key terms and terms association paths includes: using graph neural networks to learn representations of standard terms association graphs to obtain enhanced node representations. Where L represents the number of network layers; Based on the node representation, the standard terms are divided into several categories using clustering algorithms, and the centrality index of each node in the graph is calculated, including degree centrality Closeness Centrality and betweenness centrality, where d(v) represents the degree of node v, |V| represents the total number of nodes, and d(v,u) represents the shortest path length from node v to node u; Calculate the term importance score Imp(v) based on the centrality index: Imp(v)=0.4·C D (v)+0.3·C C (v)+0.3·C B (v), where C B (v) represents betweenness centrality. Based on the importance score, the top three terms are selected as key terms, and the association paths between them are analyzed as key association paths.

7. The artificial intelligence-based standard document auxiliary writing method according to claim 1 is characterized in that: Based on the association graph and structured analysis results, an intelligent auxiliary writing model is constructed to receive the user's input demand q and map it to the vector space to obtain the demand vector v q , retrieve the clause set Rel(q) related to the demand semantics in the association graph: Rel(q) = {c i |sim(v q ,v i )>θ}, where θ represents the similarity threshold, θ=0.5; calculate the recommendation score score(c i |q): Among them, δ represents the weight of graph structure information, δ = 0.3; Based on the recommendation scores, a ranked list of recommended terms is generated; based on the recommended terms and user needs, standard text content is generated; the terminology consistency, logical consistency and structural consistency of the standard content are tested, and a consistency report is generated.

8. The artificial intelligence-based standard document auxiliary writing method according to claim 7, characterized in that: The generation of standard text content based on recommended terms and user needs includes: hierarchical generation of standard framework generation, chapter structure generation and term content generation; selecting the template most similar to the user needs from a pre-established standard template library as the basic framework; and normalizing the use of terms in the generated text content to ensure terminology consistency.

9. The artificial intelligence-based standard document auxiliary writing method according to claim 7, characterized in that: The Transformer architecture and the number of network layers L of the graph neural network are consistent, and L is not less than 3.

10. An artificial intelligence-based standard document auxiliary writing system for executing the method according to any one of claims 1 to 9, characterized in that: The system includes: a standard analysis module for acquiring a standard document corpus, analyzing the standard document corpus, extracting standard clauses and constructing a standard clause knowledge representation; a graph construction module for constructing a standard clause association graph based on the standard clause knowledge representation, wherein the association graph is used to characterize the association relationship between standard clauses; a graph analysis module for applying a graph mining algorithm to perform structured analysis on the association graph to identify key clauses and clause association paths; and an intelligent assistance module for constructing an intelligent assistance writing model based on the association graph and the structured analysis results, receiving user input, and performing standard clause recommendations, automatic content generation, and consistency verification based on the user input to assist users in completing standard document writing.

Images