International standard engineering document collaborative writing method and system based on artificial intelligence

Abstract

This invention discloses an AI-based collaborative writing method and system for international standard engineering documents, comprising: automatically generating an initial modular framework tree for the document using a pre-built modular knowledge base based on project attributes and selected international standards; analyzing the user's multimodal interaction data through a neural symbolic fusion engine to construct and continuously update their dynamic cognitive mirror; during user editing, intelligent reasoning is performed by integrating project context, module semantics, and cognitive mirror to generate a personalized writing assistance report; in multi-person collaborative scenarios, by real-time monitoring of editing conflicts and multi-agent game theory simulation, potential consensus solutions and a cognitive fusion graph revealing the consensus formation logic are simulated and generated; document evolution suggestions are generated based on the simulation results, and individual cognitive mirrors are continuously optimized using user feedback; simultaneously, by analyzing group decision-making patterns, group cognitive blind spots are identified and feedback is provided, achieving a collaborative writing closed loop from individual intelligent assistance to group cognitive enhancement.

Description

Technical Field

[0001] This invention relates to the field of engineering technical document writing technology, and in particular to an artificial intelligence-based collaborative writing method and system for international standard engineering documents. Background Technology

[0002] Collaborative writing of international standard engineering documents, especially in fields such as Building Information Modeling (BIM) management that follow complex standard systems like ISO 19650, is a highly specialized, multi-party, and rigorous collective intellectual activity. Currently, this work mainly relies on a combination of general-purpose collaborative office software, document version control systems, and instant messaging tools. Traditional version control systems can track the change history of text lines, but they cannot understand the logical connections between standard clauses, the consistency requirements of terminology, and the inherent connections between technical specifications across chapters. This results in merge requests only resolving surface textual conflicts, while deeper semantic contradictions and logical inconsistencies are often only discovered during later review, causing a large amount of rework. Existing online document editing platforms support real-time editing and commenting by multiple users, but they lack the ability to perceive and guide the unique structure of international standard documents. Writers need to manually integrate a large number of scattered standard texts, historical project cases, and localized specifications, which is time-consuming, laborious, and prone to errors.

[0003] In recent years, although general artificial intelligence (AI) technologies have made progress in text generation and assisted writing, resulting in tools capable of grammar checking, style optimization, and even content continuation, these technologies have significant limitations when applied to the writing of international standard engineering documents. Existing tools cannot model and reconcile deep-seated cognitive differences based on professional knowledge and experience. The collaborative process still heavily relies on time-consuming and unstructured meetings and email discussions, leading to inefficient consensus formation and the failure to effectively capture and reuse valuable collective decision-making logic and compromise paths. Therefore, this invention proposes an AI-based collaborative writing method and system for international standard engineering documents, thereby freeing expert groups from tedious information integration and repetitive communication, allowing them to focus on higher-value technological innovation and decision-making, ultimately improving the quality, efficiency, and internal consistency of international standard engineering document writing. Summary of the Invention

[0004] This invention overcomes the shortcomings of existing technologies and provides a collaborative writing method and system for international standard engineering documents based on artificial intelligence. Its main purpose is to improve the writing quality, efficiency and internal consistency of international standard engineering documents.

[0005] To achieve the above objectives, the first aspect of this invention provides a collaborative writing method for international standard engineering documents based on artificial intelligence, comprising: Obtain the basic attribute information of the target project and one or more selected international standards and specifications, and generate an initial document modular framework tree that matches the project attributes and standards and specifications based on a pre-built modular knowledge base; The system acquires multimodal interaction data generated by users during collaborative editing, parses the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generates the user's dynamic cognitive signature, and constructs the user's cognitive image. Based on the user's editing interaction data on a specific document module in the current project, the basic attribute information of the target project, the contextual semantics of the corresponding document module, and the user's cognitive image are integrated to perform writing reasoning, generate a writing assistance report, and push it out. In multi-user collaborative editing scenarios, the system monitors the editing content of different users on the same module or a cluster of logically related modules in real time. Combining the cognitive mirrors of related users, it conducts multi-cognitive mirror game simulation to generate potential consensus solutions and cognitive fusion graphs. Based on the potential consensus scheme and cognitive fusion graph, document evolution suggestions are generated, and the cognitive mirror is optimized based on user decision feedback. At the same time, the group cognitive blind spot information is identified and fed back by analyzing the group decision-making pattern in the collaborative process.

[0006] In this solution, the acquisition of basic attribute information of the target project and one or more selected international standards and specifications, and the generation of an initial document modular framework tree matching the project attributes and standards and specifications based on a pre-built modular knowledge base, specifically includes: Obtain basic attribute information of the target project and one or more selected international standards and specifications. The basic attribute information includes project type, project scale, project location, and core asset types involved in the project. The basic attribute information of the project is concatenated and vectorized with the selected standard specification identifier to form a multidimensional query vector, which is then input into a pre-built modular knowledge base. The knowledge base is organized in the form of an attribute graph, whose nodes store atomic document modules and metadata decomposed according to standard specifications, and whose edges are encoded with mandatory logical relationships specified by standards, semantic concept association relationships, and co-occurrence dependency relationships based on historical project statistics. Based on the query vector, graph retrieval is performed in the attribute graph to locate a set of candidate document module nodes that are related to the selected standard and whose applicable conditions intersect with the project attributes. An initial matching score is calculated for each module and the module is instantiated according to the project size attribute. Starting with the instantiated module with the highest matching degree, the graph is traversed according to the mandatory logical relationship edges in the attribute graph to deduce and link all the necessary supporting modules, forming a basic skeleton that meets the minimum requirements of the standard. Optional supplementary modules are recommended based on conceptual association and statistical co-occurrence relationships. Based on the basic skeleton and optional supplementary modules, a directed graph of module relationships is assembled. The directed graph of module relationships is then converted into a hierarchical tree view to generate an initial document modular framework tree.

[0007] In this solution, the step of acquiring multimodal interaction data generated by users during collaborative editing, parsing the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generating the user's dynamic cognitive signature, and constructing the user's cognitive image specifically includes: By integrating event listeners into the collaborative editing interface, the system captures and aggregates multimodal interactive data streams generated when users operate on specific modules in the initial document modular framework tree. These data streams include text content entered or modified by the user in the editor, annotations and comments added to specific content blocks, and discussion text posted in the project discussion area. After the multimodal interactive data stream is preprocessed by the neural symbol fusion engine, it is parsed in parallel through a dual-channel Transformer encoder network with a shared embedding layer. The dual channels include a continuous semantic encoding channel and a discrete logic extraction channel. In the continuous semantic encoding channel, the input sequence is deep context-encoded through stacked multi-head self-attention layers to capture semantic dependencies that span long distances between words and sentences, and the high-dimensional continuous semantic vector representing the user's complete intent and technical preferences is aggregated and output through pooling layers. In the discrete logic extraction channel, the attention mechanism is guided by the embedded differentiable logic constraint layer to identify logical relationship patterns and drive the pointer network to mark the relevant logical entities, thus constructing a structured discrete logic graph with logical entities as nodes and logical relationships as edges. Based on the newly generated high-dimensional continuous semantic vector and the target user's historical semantic vector sequence, the temporal evolution trajectory and expression style of the user's focus are analyzed to generate an evolution trajectory feature vector; the newly generated structured discrete logic graph is input into the graph neural network, and the argument structure topology is learned through the message passing mechanism to analyze the logical rigor, standard citation preference and decision tendency of the user's current interaction to generate a logical pattern feature vector. The evolution trajectory feature vector and the logical pattern feature vector are concatenated to form a high-dimensional joint feature vector. This vector is then subjected to dimensionality reduction and normalization through a fully connected layer to generate an updatable multidimensional dynamic cognitive signature tensor. Based on this dynamic cognitive signature tensor, a user cognitive mirror model is instantiated and continuously updated.

[0008] In this solution, the step of generating a writing assistance report based on user editing interaction data of a specific document module in the current project, integrating the basic attribute information of the target project, the contextual semantics of the corresponding document module, and the user's cognitive image, and then pushing it out, specifically includes: When the system detects an editing event of a user on a specific document module in the project, it obtains the functional definition, hierarchy in the tree structure, and semantics of the associated modules of the target module from the initial document modular framework tree, and generates the context semantic encoding of the target module. Obtain the basic attribute information vector of the target project, and retrieve the quantitative feature vectors of knowledge mastery, editing preferences and decision-making tendencies related to the technical field of the target module from the user's cognitive mirror, and generate the user cognitive mirror feature vector; Using the node corresponding to the target module in the attribute graph of the modular knowledge base as the query center, multi-hop traversal and retrieval are performed along the mandatory logical relationship edge, semantic association relationship edge and historical co-occurrence relationship edge encoded in the graph to obtain a set of candidate knowledge fragments with source weight and context relevance score; The candidate knowledge fragment set, the context semantic encoding of the target module, the project attribute vector, and the user cognitive mirror feature vector are input into the neural symbolic reasoning model based on the encoder-decoder architecture. The candidate knowledge fragments and their relationship with the graph are jointly encoded by the encoder of the graph attention network to generate a deep context-aware vector representation of each fragment. The deep context-aware vector representation of each segment is input into a decoder based on an embedded controllable attention mechanism for text generation. During the decoding and generation process, the user's cognitive mirror feature vector is used for real-time modulation and guidance, and the attention distribution of the encoded segment is dynamically calculated. Logical reasoning and text sequence generation are then performed based on the attention distribution. The final output is a draft text that is highly adapted to the user's individual cognitive level and project context, generating a writing assistance report, which is then pushed to the user's editing interface through the collaborative editing platform's interface.

[0009] In this solution, under the scenario of multi-user collaborative editing, the editing content of different users on the same module or a logically related cluster of modules is monitored in real time. Combined with the cognitive mirrors of related users, multi-cognitive mirror game theory is performed to simulate and generate potential consensus solutions and cognitive fusion graphs. Specifically, this includes: In multi-user collaborative editing mode, the editing event streams of different users are captured in real time through event listening and status tracking, and conflict domain dynamic identification is performed. The differential algorithm is used to locate the substantial overlapping modifications to the same document module. By leveraging the logical dependencies in the initial document modular framework tree, distributed modifications to module clusters with parent-child, sequential, or strong reference relationships are identified, and the identification results are aggregated into conflict domain description objects. For each conflict domain, the current cognitive mirror of all users is retrieved, and the original edited content of each user is analyzed in depth. By deconstructing the core claims, supporting reasons and bottom-line boundaries implied in each editing behavior through the technical stance vector, argumentation logic pattern and decision preference characteristics encapsulated in the cognitive mirror, the intentional claim object is formed. Based on the intent claim object and its associated cognitive mirror, a multi-agent cognitive game inference environment is initialized through the standard mandatory logical relationships and project business rules in the modular knowledge base. The policy network of each agent is parameterized and initialized by the corresponding user cognitive mirror feature vector. The simulation employs a multi-agent reinforcement learning framework based on an actor-critic architecture. Each agent iteratively generates and executes action strategies, including advocacy adjustment, argument transformation, or proposing alternatives, based on the current environmental state and the historical actions of other agents. The process involves iterative computation until an equilibrium state that satisfies all hard constraints and maximizes the capacity to accommodate the core claims of each agent converges. From this converged equilibrium state, an executable, conflict-free editing sequence is obtained as a potential consensus scheme. The evolution trajectory of each agent strategy vector, the key decision points where the claims were modified, and the fusion weight and logical reconstruction relationship of the final potential consensus solution to the initial claims are recorded synchronously throughout the entire process of deduction. A cognitive fusion graph is constructed, in which nodes represent claims and decision events, and edges are labeled with influence and fusion relationships.

[0010] In this solution, the step of generating document evolution suggestions based on the potential consensus scheme and cognitive fusion graph, optimizing the cognitive mirror based on user decision feedback, and identifying and feeding back information on group cognitive blind spots by analyzing group decision-making patterns during the collaboration process specifically includes: Obtain potential consensus schemes and cognitive fusion graphs, extract specific editing operations, modified text content and their logical reasoning chains contained in potential consensus schemes, and obtain consensus content information; By identifying key compromises, adopted core arguments, and shelved minor disagreements through cognitive fusion graph identification, graph logical information is obtained. This information, along with the consensus content information, is then input into a natural language generation model to transform it into document evolution suggestions that include recommended modifications. When a user makes a decision to adopt, modify and then adopt or reject the document evolution suggestion, the system captures and associates the feedback behavior, the original suggestion, the user's cognitive image snapshot and the project context, and regards the user's decision as new empirical evidence of their cognitive preferences. The system then adjusts the feature vector parameters in the user's cognitive image through the backpropagation optimization algorithm. When optimizing individual cognitive mirrors, we perform group cognitive pattern analysis on all collaborative interaction historical data in the current project. Through cluster analysis and anomaly detection algorithms, we identify potential systemic blind spots in group decision-making and obtain information on group cognitive blind spots. Based on information about blind spots in group cognition, a group cognition insight report is generated and pushed to project managers or relevant users through a collaborative platform, indicating areas where external review, supplementary research, or special discussions are needed, thereby enhancing group cognition.

[0011] A second aspect of the present invention provides an artificial intelligence-based collaborative writing system for international standard engineering documents. The system includes: a memory, a processor, and a communication interface. The memory contains a program for an artificial intelligence-based collaborative writing method for international standard engineering documents. When the program for an artificial intelligence-based collaborative writing method for international standard engineering documents is executed by the processor, it implements the steps of the artificial intelligence-based collaborative writing method for international standard engineering documents as described above. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0013] Figure 1 A flowchart of the first method for a collaborative writing method of international standard engineering documents based on artificial intelligence, provided as an embodiment of the present invention; Figure 2 A flowchart of a second method for collaborative writing of international standard engineering documents based on artificial intelligence, provided as an embodiment of the present invention; Figure 3 A block diagram of an international standard engineering document collaborative writing system based on artificial intelligence, provided as an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0016] Figure 1 A flowchart of the first method for a collaborative writing method of international standard engineering documents based on artificial intelligence, provided as an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart for a collaborative writing method of international standard engineering documents based on artificial intelligence, including: S102: Obtain the basic attribute information of the target project and one or more selected international standard specifications; based on the pre-built modular knowledge base, generate an initial document modular framework tree that matches the project attributes and standard specifications. S104, acquire the multimodal interaction data generated by the user during the collaborative editing process, parse the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generate the user's dynamic cognitive signature and construct the user's cognitive image; S106, Based on the user's editing interaction data on a specific document module in the current project, the basic attribute information of the target project, the contextual semantics of the corresponding document module, and the user's cognitive image are integrated to perform writing reasoning, generate a writing assistance report and push it out; S108, in multi-user collaborative editing scenarios, monitors in real time the editing content of different users on the same module or a cluster of modules with logical connections, combines the cognitive mirrors of related users, performs multi-cognitive mirror game deduction, and simulates and generates potential consensus solutions and cognitive fusion graphs. S110, generate document evolution suggestions based on the potential consensus scheme and cognitive fusion graph, optimize the cognitive mirror based on user decision feedback, and identify and feedback group cognitive blind spot information by analyzing the group decision-making pattern in the collaborative process.

[0017] Furthermore, in a preferred embodiment of the present invention, the step of obtaining the basic attribute information of the target project and one or more selected international standard specifications, and generating an initial document modular framework tree matching the project attributes and standard specifications based on a pre-built modular knowledge base, specifically includes: Obtain basic attribute information of the target project and one or more selected international standards and specifications. The basic attribute information includes project type, project scale, project location, and core asset types involved in the project. The basic attribute information of the project is concatenated and vectorized with the selected standard specification identifier to form a multidimensional query vector, which is then input into a pre-built modular knowledge base. The knowledge base is organized in the form of an attribute graph, whose nodes store atomic document modules and metadata decomposed according to standard specifications, and whose edges are encoded with mandatory logical relationships specified by standards, semantic concept association relationships, and co-occurrence dependency relationships based on historical project statistics. Based on the query vector, graph retrieval is performed in the attribute graph to locate a set of candidate document module nodes that are related to the selected standard and whose applicable conditions intersect with the project attributes. An initial matching score is calculated for each module and the module is instantiated according to the project size attribute. Starting with the instantiated module with the highest matching degree, the graph is traversed according to the mandatory logical relationship edges in the attribute graph to deduce and link all the necessary supporting modules, forming a basic skeleton that meets the minimum requirements of the standard. Optional supplementary modules are recommended based on conceptual association and statistical co-occurrence relationships. Based on the basic skeleton and optional supplementary modules, a directed graph of module relationships is assembled. The directed graph of module relationships is then converted into a hierarchical tree view to generate an initial document modular framework tree.

[0018] It's important to note that the core of this step lies in automatically transforming abstract project requirements and international standards into concrete, actionable document writing frameworks through a modular knowledge base. The "basic attribute information," such as project type (e.g., "large transportation hub"), project scale, and region, is the key input for filtering and customizing document content. Vectorization technology integrates these attributes with user-selected standard identifiers (e.g., "ISO 19650-2") to form a comprehensive query vector that simultaneously expresses project characteristics and specification requirements. The "pre-built modular knowledge base" deconstructs and semantically reorganizes the text of international standards (e.g., the ISO 19650 series), forming an attribute graph with "atomic document modules" as nodes. Each module corresponds to an independently manageable concept or requirement unit within the standard, such as "Common Data Environment (CDE) configuration requirements" or "Model Level of Detail (LOD) delivery specifications." The edges in the graph define the rich relationships between these modules: for example, "mandatory logical relationships" ensure the order and dependency specified in the standard (e.g., module A must be completed before module B); "semantic conceptual relationships" reveal the inherent connections between modules in terms of technical themes; and "co-occurrence dependencies based on historical project statistics" include industry best practices, such as in "large transportation hub" projects, where "progress management module" and "interface coordination module" often need to be defined in detail together.

[0019] Furthermore, the query vector is used to perform intelligent retrieval and reasoning within the graph. Instead of simply returning all relevant modules, it prioritizes them by calculating an "initial matching score" and pre-populates the parameters within each module based on attributes such as "project size" (e.g., automatically setting more frequent model collaboration review cycles for large projects). Then, using the core module as an anchor, the graph is traversed along the edges of "mandatory logical relationships" to construct a basic document skeleton that meets minimum compliance requirements, ensuring the document's structural completeness. Based on this, "optional supplementary modules" that can improve document quality and project adaptability are intelligently recommended according to semantic and statistical correlations. For example, for a "transportation hub" project, modules such as "human flow simulation data delivery requirements" are automatically suggested. Finally, all selected modules and their relationships are integrated into a "directed graph of module relationships" and converted into a clear, hierarchical tree view, i.e., the "initial document modular framework tree," thus presenting the user with a customized document writing blueprint that has been intelligently analyzed and assembled.

[0020] Furthermore, in a preferred embodiment of the present invention, the step of acquiring multimodal interaction data generated by the user during collaborative editing, parsing the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generating the user's dynamic cognitive signature, and constructing a user cognitive image specifically includes: By integrating event listeners into the collaborative editing interface, the system captures and aggregates multimodal interactive data streams generated when users operate on specific modules in the initial document modular framework tree. These data streams include text content entered or modified by the user in the editor, annotations and comments added to specific content blocks, and discussion text posted in the project discussion area. After the multimodal interactive data stream is preprocessed by the neural symbol fusion engine, it is parsed in parallel through a dual-channel Transformer encoder network with a shared embedding layer. The dual channels include a continuous semantic encoding channel and a discrete logic extraction channel. In the continuous semantic encoding channel, the input sequence is deep context-encoded through stacked multi-head self-attention layers to capture semantic dependencies that span long distances between words and sentences, and the high-dimensional continuous semantic vector representing the user's complete intent and technical preferences is aggregated and output through pooling layers. In the discrete logic extraction channel, the attention mechanism is guided by the embedded differentiable logic constraint layer to identify logical relationship patterns and drive the pointer network to mark the relevant logical entities, thus constructing a structured discrete logic graph with logical entities as nodes and logical relationships as edges. Based on the newly generated high-dimensional continuous semantic vector and the target user's historical semantic vector sequence, the temporal evolution trajectory and expression style of the user's focus are analyzed to generate an evolution trajectory feature vector; the newly generated structured discrete logic graph is input into the graph neural network, and the argument structure topology is learned through the message passing mechanism to analyze the logical rigor, standard citation preference and decision tendency of the user's current interaction to generate a logical pattern feature vector. The evolution trajectory feature vector and the logical pattern feature vector are concatenated to form a high-dimensional joint feature vector. This vector is then subjected to dimensionality reduction and normalization through a fully connected layer to generate an updatable multidimensional dynamic cognitive signature tensor. Based on this dynamic cognitive signature tensor, a user cognitive mirror model is instantiated and continuously updated.

[0021] It's important to note that this step aims to transform user behavior during collaborative editing into a dynamic digital model that represents their professional knowledge and decision-making patterns. The event listener captures not only the text itself, but also the interactive intent within the specific engineering standard context. For example, when a user edits in the "Detail Requirements" module, their input, critical comments on a clause, or arguments supporting a particular technical approach in the discussion area collectively constitute a multimodal data stream reflecting their professional knowledge and judgment. The neural symbolic fusion engine is the core processing unit of this step, its key being parallel parsing achieved through a dual-channel Transformer encoder network. The continuous semantic encoding channel is analogous to understanding the "context and emotion" of user expression; it captures the overall intent and technical inclination of the text through a deep attention mechanism, such as identifying a user's strong support for the claim of "using the highest level of model accuracy." The discrete logic extraction channel acts as the "skeleton and joints" of the user's argument. Its built-in differentiable logic constraint layer can identify standardized logical expressions such as "According to ISO 19650-3 Clause 5.2, therefore..." and accurately locates the cited standard clauses and specific technical parameters through a pointer network, thereby constructing a structured logical relationship graph. After obtaining the semantic vector and logic graph, deep feature extraction is performed through temporal modeling and graph structure analysis. Analyzing the temporal evolution of the user's focus, for example, reveals a shift in the discussion focus from "geometric precision" to "reliability of information delivery," indicating a shift in their thinking focus. Simultaneously, graph neural networks are used to analyze the logic graph to evaluate the rigor of the user's argument, such as whether the cited standard clauses are accurate and whether the argument chain is complete. Finally, the trajectory features reflecting the "thought process" are fused with the pattern features of the "argument skeleton," generating a concise yet information-rich dynamic cognitive signature tensor through dimensionality reduction. This tensor is the mathematical core of the user's cognitive mirror, which enables the system to quantify the knowledge depth, style preferences and decision-making inertia of different users when dealing with different modules such as "sustainability assessment" or "safety compliance", thus laying the foundation for subsequent personalized assistance and intelligent collaboration.

[0022] Furthermore, in a preferred embodiment of the present invention, the step of generating a writing assistance report and pushing it based on the user's editing interaction data of a specific document module in the current project, integrating the basic attribute information of the target project, the contextual semantics of the corresponding document module, and the user's cognitive image, specifically includes: When the system detects an editing event of a user on a specific document module in the project, it obtains the functional definition, hierarchy in the tree structure, and semantics of the associated modules of the target module from the initial document modular framework tree, and generates the context semantic encoding of the target module. Obtain the basic attribute information vector of the target project, and retrieve the quantitative feature vectors of knowledge mastery, editing preferences and decision-making tendencies related to the technical field of the target module from the user's cognitive mirror, and generate the user cognitive mirror feature vector; Using the node corresponding to the target module in the attribute graph of the modular knowledge base as the query center, multi-hop traversal and retrieval are performed along the mandatory logical relationship edge, semantic association relationship edge and historical co-occurrence relationship edge encoded in the graph to obtain a set of candidate knowledge fragments with source weight and context relevance score; The candidate knowledge fragment set, the context semantic encoding of the target module, the project attribute vector, and the user cognitive mirror feature vector are input into the neural symbolic reasoning model based on the encoder-decoder architecture. The candidate knowledge fragments and their relationship with the graph are jointly encoded by the encoder of the graph attention network to generate a deep context-aware vector representation of each fragment. The deep context-aware vector representation of each segment is input into a decoder based on an embedded controllable attention mechanism for text generation. During the decoding and generation process, the user's cognitive mirror feature vector is used for real-time modulation and guidance, and the attention distribution of the encoded segment is dynamically calculated. Logical reasoning and text sequence generation are then performed based on the attention distribution. The final output is a draft text that is highly adapted to the user's individual cognitive level and project context, generating a writing assistance report, which is then pushed to the user's editing interface through the collaborative editing platform's interface.

[0023] It's important to note that this step describes how to achieve truly personalized intelligent writing assistance, the core of which lies in integrating specific task contexts with user cognitive characteristics for targeted reasoning. When a user begins editing a specific module, such as the "Information Security Management System" section, the system doesn't process the module's text in isolation, but rather constructs a multi-dimensional decision context. This context integrates the module's position and function within the overall document framework (e.g., belonging to the "Management Requirements" main section), the project's fundamental attributes (e.g., the high security requirements involved in a "large airport project"), and personalized characteristics extracted from the user's cognitive mirror (e.g., their past "risk aversion tendency" on security issues and familiarity with specific standard clusters). Based on this rich context, intelligent retrieval is performed in a modular knowledge base. The retrieval process centers on the graph node corresponding to the module, exploring along predefined relational paths. It not only searches for explicitly mandated related clauses in the standards but also discovers semantically related concepts (e.g., "access control" and "data encryption") and references common practice combinations from similar historical projects, thus forming a set of candidate knowledge fragments that are both compliant and practically valuable. Subsequently, the encoder employs a graph attention network to jointly encode the retrieved knowledge fragments and their complex relationships, ensuring that the vector representation of each fragment contains its semantic position within the current knowledge network. During the decoding and generation phase, a controllable attention mechanism uses user cognitive mirror features as regulatory signals. For example, if the mirror indicates a shallow understanding of standards related to "business continuity," the system assigns higher explanatory weights to fragments related to this topic and automatically supplements background information when generating suggestions; if the user's decision-making tendency vector shows a structured expression of their preferences, the generated text becomes more coherent. Through this dynamic modulation, the decoder's logical reasoning and text generation achieve a deep integration of general knowledge, project requirements, and individual cognitive characteristics. The final output draft text, used to create supplementary reports, is no longer simply filling in a standardized template, but rather a product of deep adaptation. The report not only includes the recommended text, but may also include a concise explanation of the reasoning behind it, such as "This recommendation incorporates the requirements of Section X of ISO 19650-5 and references Solution Y, which you have adopted in similar projects, while also taking into account the high security level of this project," thus providing users with transparent, reliable and highly usable intelligent assistance.

[0024] Furthermore, in a preferred embodiment of the present invention, the step of monitoring the editing content of different users on the same module or a logically related module cluster in real time in a multi-user collaborative editing scenario, and combining the cognitive mirrors of related users to perform multi-cognitive mirror game deduction, simulating and generating potential consensus schemes and cognitive fusion graphs, specifically includes: In multi-user collaborative editing mode, the editing event streams of different users are captured in real time through event listening and status tracking, and conflict domain dynamic identification is performed. The differential algorithm is used to locate the substantial overlapping modifications to the same document module. By leveraging the logical dependencies in the initial document modular framework tree, distributed modifications to module clusters with parent-child, sequential, or strong reference relationships are identified, and the identification results are aggregated into conflict domain description objects. For each conflict domain, the current cognitive mirror of all users is retrieved, and the original edited content of each user is analyzed in depth. By deconstructing the core claims, supporting reasons and bottom-line boundaries implied in each editing behavior through the technical stance vector, argumentation logic pattern and decision preference characteristics encapsulated in the cognitive mirror, the intentional claim object is formed. Based on the intent claim object and its associated cognitive mirror, a multi-agent cognitive game inference environment is initialized through the standard mandatory logical relationships and project business rules in the modular knowledge base. The policy network of each agent is parameterized and initialized by the corresponding user cognitive mirror feature vector. The simulation employs a multi-agent reinforcement learning framework based on an actor-critic architecture. Each agent iteratively generates and executes action strategies, including advocacy adjustment, argument transformation, or proposing alternatives, based on the current environmental state and the historical actions of other agents. The process involves iterative computation until an equilibrium state that satisfies all hard constraints and maximizes the capacity to accommodate the core claims of each agent converges. From this converged equilibrium state, an executable, conflict-free editing sequence is obtained as a potential consensus scheme. The evolution trajectory of each agent strategy vector, the key decision points where the claims were modified, and the fusion weight and logical reconstruction relationship of the final potential consensus solution to the initial claims are recorded synchronously throughout the entire process of deduction. A cognitive fusion graph is constructed, in which nodes represent claims and decision events, and edges are labeled with influence and fusion relationships.

[0025] It's important to note that this technical step aims to address the most complex and deep-seated conflicts in collaborative editing. Its core lies in transforming superficial text editing differences into a game-like and fusion process of underlying cognitive models. Specifically, when the system runs in collaborative editing mode, it not only tracks surface-level text modifications in real time but also identifies implicit conflicts across modules through the logical dependencies encoded in the initial document modular framework tree. For example, when a structural engineer modifies parameters in the "Load Standards" module, but the architect fails to make corresponding adjustments in the "Construction Details" module, which references the same module, the system identifies this distributed modification as a conflict domain requiring coordination. After identifying the conflict domain, the system doesn't simply compare or merge texts but initiates deep cognitive analysis. It invokes the cognitive mirrors of the users involved to deconstruct the technical stances and argumentation logic behind their editing behavior. For instance, one user's modification might be based on strict adherence to local fire codes (such as the UK's "Approved Document B"), while another user's modification might emphasize the flexibility of internationally accepted standards (such as ISO). The system parses these differing claims into structured "intent claim objects."

[0026] Subsequently, these claimants and their underlying cognitive agents are placed in a simulated cognitive game environment for deduction. The rules of this environment are derived from standard mandatory logic and project business constraints, ensuring the compliance of the deduction process. The "behavioral pattern" of each agent is dynamically shaped by its corresponding user cognitive mirror, and the deduction process uses multi-agent reinforcement learning to simulate multi-party negotiation. For example, the system may simulate agent A (representing a structural engineer) accepting an alternative proposed by agent B (representing an architect) after several rounds of interaction, citing a new research result, provided that agent B agrees to strengthen the safety redundancy description in another related clause. The actor-critic architecture multi-agent reinforcement learning framework, through its "centralized training, decentralized execution" paradigm, provides a systematic solution to the multi-agent collaborative decision-making problem. Its core advantage lies in its ability to use global information to guide individual strategy optimization and effectively address key challenges such as credit allocation, environmental non-stationarity, and scalability through technologies such as advantage functions and communication mechanisms. Through iterative computation until an equilibrium state that satisfies all hard constraints and maximizes the preservation of the core concerns of all parties is reached, from which an executable "potential consensus solution" is extracted. Simultaneously, based on the strategic evolution of each agent throughout the entire deduction process, the key nodes advocating compromise, and the fusion logic of the final solution, a systematic record was created and constructed into a "cognitive fusion map." This map allows for the visualization and tracing of the complete cognitive path from the initial conflict to the formation of consensus. It not only outputs a conflict-free text merging result, but more importantly, it provides the team with a "cognitive map" for understanding the nature of the conflict, the decision-making logic, and the consensus-building process, greatly improving the decision-making quality and efficiency of collaborative writing of complex engineering documents.

[0027] Figure 2 A flowchart of a second method for collaborative writing of international standard engineering documents based on artificial intelligence, provided as an embodiment of the present invention; like Figure 2 As shown, this invention provides a second method flowchart for a collaborative writing method of international standard engineering documents based on artificial intelligence, including: S202, obtain the potential consensus scheme and cognitive fusion graph, extract the specific editing operations, modified text content and their logical reasoning chain contained in the potential consensus scheme, and obtain consensus content information; S204, by identifying the key compromise points, adopted core arguments, and shelved minor disagreements through cognitive fusion graph, the graph logic information is obtained and input into the natural language generation model along with the consensus content information to transform it into document evolution suggestions including recommended modification items; S206, when a user makes a decision feedback on the document evolution suggestion, such as adopting, modifying and then adopting or rejecting it, the system captures and associates the feedback behavior, the original suggestion, the user's cognitive image snapshot and the project context, regards the user's decision as a new empirical evidence of their cognitive preferences, and adjusts the feature vector parameters in the user's cognitive image through the backpropagation optimization algorithm. S208, during the optimization of individual cognitive mirrors, performs group cognitive pattern analysis on all collaborative interaction historical data in the current project, and identifies possible systemic blind spots in group decision-making through cluster analysis and anomaly detection algorithms, thereby obtaining information on group cognitive blind spots; S210 generates a group cognition insight report based on information about blind spots in group cognition. This report is then pushed to project managers or relevant users through a collaborative platform, indicating areas where external review, supplementary research, or thematic discussions are needed, thereby enhancing group cognition.

[0028] It's important to note that this step constructs a complete learning loop, evolving from intelligent collaboration to collective wisdom. First, the "potential consensus solutions" and "cognitive fusion maps" generated through game theory are deeply processed, transforming them into actionable guidelines understandable and actionable by human experts. For example, the system not only generates a list of suggested modifications to "adjust structural load parameters from X to Y," but also uses natural language to generate a model with a clear explanation: "This adjustment integrates structural engineer A's conservative approach based on safety redundancy with architect B's requirements for spatial clearance. The compromise lies in using an intermediate value and adding monitoring clauses. Simulation verification of extreme conditions has been temporarily shelved due to significant controversy and is recommended for further discussion." This makes the document evolution suggestion itself a "decision memo" recording the negotiation logic. When users provide feedback on the suggestions, this feedback is considered a valuable opportunity for cognitive calibration. For example, if a user known for their "innovation adoption" tendency rejects a technologically advanced solution, the system will use a backpropagation algorithm to fine-tune implicit parameters in their cognitive image, such as "risk assessment weight" or "trust level towards a specific vendor," making their digital avatar more closely resemble their actual decision-making patterns. This continuous optimization ensures that the cognitive image can dynamically track the growth of expert experience and the evolution of perspectives.

[0029] Simultaneously, at the macro level, group cognition diagnosis is conducted. This involves analyzing records of all discussions, decisions, and consensus reached within the project, and using algorithms to identify potential systemic biases within the team. For example, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) density clustering algorithm is used. This algorithm automatically discovers clusters based on the density distribution of data points in the feature space. Its advantage lies in not requiring pre-specified number of categories and effectively identifying discrete noise points. By vectorizing all decision points in the project (such as the adoption, modification, or rejection of specific standard clauses) according to their technical attributes, participating experts, and decision results, and inputting them into the DBSCAN algorithm, the algorithm automatically clusters distinct decision-making patterns when the team addresses "structural security" and "sustainability" issues. It may also identify a few novel proposals that reference cutting-edge technologies but were not adopted as "noise points" or potential innovations far removed from the main clusters. Furthermore, based on the identification of group patterns, the Isolation Forest anomaly detection algorithm is employed to detect systemic blind spots. This algorithm is suitable for high-dimensional data. It quickly identifies "abnormal" samples that are significantly different from other data points by randomly partitioning the feature space. Various features in the group decision-making pattern (such as the age distribution of cited standards, the source type of the argument, and the repeatedly shelved issue areas) are constructed into multi-dimensional feature vectors and input into an isolated forest for training and testing. For example, in a year-long project collaboration history, all standard clauses relied upon for decisions regarding "information security" were published before a certain point in time, forming an anomalous cluster significantly different from the team's common pattern of frequently citing the latest standards in other technical fields. This accurately pinpoints the specific blind spot in group cognition: "outdated information security standard knowledge base." Finally, based on the analysis results, a "group cognition insight report" is generated, which helps resolve current conflicts, proactively reveals the team's capability gaps, promotes organizational learning and improves collective cognition, and makes the collaborative writing process itself a platform for accumulating and passing on organizational wisdom.

[0030] Figure 3 An embodiment of the present invention provides an artificial intelligence-based collaborative writing system 3 for international standard engineering documents. The system includes a memory 301, a processor 302, and a communication interface 303. The memory 301 contains a program for a collaborative writing method of international standard engineering documents based on artificial intelligence. When the program for a collaborative writing method of international standard engineering documents based on artificial intelligence is executed by the processor 302, it implements the steps of the collaborative writing method of international standard engineering documents based on artificial intelligence as described above.

[0031] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A collaborative writing method for international standard engineering documents based on artificial intelligence, characterized in that, include: Obtain the basic attribute information of the target project and one or more selected international standards and specifications, and generate an initial document modular framework tree that matches the project attributes and standards and specifications based on a pre-built modular knowledge base; The system acquires multimodal interaction data generated by users during collaborative editing, parses the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generates the user's dynamic cognitive signature, and constructs the user's cognitive image. Based on the user's editing interaction data on a specific document module in the current project, the basic attribute information of the target project, the contextual semantics of the corresponding document module, and the user's cognitive image are integrated to perform writing reasoning, generate a writing assistance report, and push it out. In multi-user collaborative editing scenarios, the system monitors the editing content of different users on the same module or a cluster of logically related modules in real time. Combining the cognitive mirrors of related users, it conducts multi-cognitive mirror game simulation to generate potential consensus solutions and cognitive fusion graphs. Based on the potential consensus scheme and cognitive fusion graph, document evolution suggestions are generated, and the cognitive mirror is optimized based on user decision feedback. At the same time, the group cognitive blind spot information is identified and fed back by analyzing the group decision-making pattern in the collaborative process.

2. The method for collaborative writing of international standard engineering documents based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the basic attribute information of the target project and one or more selected international standards and specifications, and generating an initial document modular framework tree that matches the project attributes and standards and specifications based on a pre-built modular knowledge base, specifically includes: Obtain basic attribute information of the target project and one or more selected international standards and specifications. The basic attribute information includes project type, project scale, project location, and core asset types involved in the project. The basic attribute information of the project is concatenated and vectorized with the selected standard specification identifier to form a multidimensional query vector, which is then input into a pre-built modular knowledge base. The modular knowledge base is organized in the form of an attribute graph. Its nodes store atomic document modules and metadata decomposed according to standard specifications. Its edges are encoded with mandatory logical relationships specified by the standard, semantic concept association relationships, and co-occurrence dependency relationships based on historical project statistics. Based on the query vector, graph retrieval is performed in the attribute graph to locate a set of candidate document module nodes that are related to the selected standard and whose applicable conditions intersect with the project attributes. An initial matching score is calculated for each module and the module is instantiated according to the project size attribute. Starting with the instantiated module with the highest matching degree, the graph is traversed according to the mandatory logical relationship edges in the attribute graph to deduce and link all the necessary supporting modules, forming a basic skeleton that meets the minimum requirements of the standard. Optional supplementary modules are recommended based on conceptual association and statistical co-occurrence relationships. Based on the basic skeleton and optional supplementary modules, a directed graph of module relationships is assembled. The directed graph of module relationships is then converted into a hierarchical tree view to generate an initial document modular framework tree.

3. The method for collaborative writing of international standard engineering documents based on artificial intelligence according to claim 1, characterized in that, The process of acquiring multimodal interaction data generated by users during collaborative editing, parsing the multimodal interaction data in parallel into continuous semantic vectors and discrete logic graphs, generating dynamic cognitive signatures of users, and constructing user cognitive images specifically includes: By integrating event listeners into the collaborative editing interface, the system captures and aggregates multimodal interactive data streams generated when users operate on specific modules in the initial document modular framework tree. These data streams include text content entered or modified by the user in the editor, annotations and comments added to specific content blocks, and discussion text posted in the project discussion area. After the multimodal interactive data stream is preprocessed by the neural symbol fusion engine, it is parsed in parallel through a dual-channel Transformer encoder network with a shared embedding layer. The dual channels include a continuous semantic encoding channel and a discrete logic extraction channel. In the continuous semantic encoding channel, the input sequence is deep context-encoded through stacked multi-head self-attention layers to capture semantic dependencies that span long distances between words and sentences, and the high-dimensional continuous semantic vector representing the user's complete intent and technical preferences is aggregated and output through pooling layers. In the discrete logic extraction channel, the attention mechanism is guided by the embedded differentiable logic constraint layer to identify logical relationship patterns and drive the pointer network to mark the relevant logical entities, thus constructing a structured discrete logic graph with logical entities as nodes and logical relationships as edges. Based on the newly generated high-dimensional continuous semantic vector and the target user's historical semantic vector sequence, the temporal evolution trajectory and expression style of the user's focus are analyzed to generate an evolution trajectory feature vector; the newly generated structured discrete logic graph is input into the graph neural network, and the argument structure topology is learned through the message passing mechanism to analyze the logical rigor, standard citation preference and decision tendency of the user's current interaction to generate a logical pattern feature vector. The evolution trajectory feature vector and the logical pattern feature vector are concatenated to form a high-dimensional joint feature vector. This vector is then subjected to dimensionality reduction and normalization through a fully connected layer to generate an updatable multidimensional dynamic cognitive signature tensor. Based on this dynamic cognitive signature tensor, the user cognitive mirror model is instantiated and continuously updated.

4. The method for collaborative writing of international standard engineering documents based on artificial intelligence according to claim 1, characterized in that, The process involves using user editing interaction data for specific document modules within the current project, integrating the target project's basic attribute information, the contextual semantics of the corresponding document modules, and the user's cognitive mirror to perform writing reasoning, generating a writing assistance report for delivery, specifically including: When the system detects an editing event of a user on a specific document module in the project, it obtains the functional definition, hierarchy in the tree structure, and semantics of the associated modules of the target module from the initial document modular framework tree, and generates the context semantic encoding of the target module. Obtain the basic attribute information vector of the target project, and retrieve the quantitative feature vectors of knowledge mastery, editing preferences and decision-making tendencies related to the technical field of the target module from the user's cognitive mirror, and generate the user cognitive mirror feature vector; Using the node corresponding to the target module in the attribute graph of the modular knowledge base as the query center, multi-hop traversal and retrieval are performed along the mandatory logical relationship edge, semantic association relationship edge and historical co-occurrence relationship edge encoded in the graph to obtain a set of candidate knowledge fragments with source weight and context relevance score; The candidate knowledge fragment set, the context semantic encoding of the target module, the project attribute vector, and the user cognitive mirror feature vector are input into the neural symbolic reasoning model based on the encoder-decoder architecture. The candidate knowledge fragments and graph relationships are jointly encoded by the encoder of the graph attention network to generate a deep context-aware vector representation of each fragment. The deep context-aware vector representation of each segment is input into a decoder based on an embedded controllable attention mechanism for text generation. During the decoding and generation process, the user's cognitive mirror feature vector is used for real-time modulation and guidance, and the attention distribution of the encoded segment is dynamically calculated. Logical reasoning and text sequence generation are then performed based on the attention distribution. The final output is a draft text that is highly adapted to the user's individual cognitive level and project context, generating a writing assistance report, which is then pushed to the user's editing interface through the collaborative editing platform's interface.

5. The method for collaborative writing of international standard engineering documents based on artificial intelligence according to claim 1, characterized in that, In the scenario of multi-user collaborative editing, the editing content of different users on the same module or a logically related cluster of modules is monitored in real time. Combined with the cognitive mirrors of related users, multi-cognitive mirror game theory is performed to simulate and generate potential consensus schemes and cognitive fusion graphs. Specifically, this includes: In the multi-user collaborative editing mode, the editing event streams of different users are captured in real time through event listening and status tracking, and the conflict domain is dynamically identified. The differential algorithm is used to locate the substantial overlapping modifications to the same document module. By leveraging the logical dependencies in the initial document modular framework tree, distributed modifications to module clusters with parent-child, sequential, or strong reference relationships are identified, and the identification results are aggregated into conflict domain description objects. For each conflict domain, the current cognitive mirror of all users is retrieved, and the original edited content of each user is analyzed in depth. By deconstructing the core claims, supporting reasons and bottom-line boundaries implied in each editing behavior through the technical stance vector, argumentation logic pattern and decision preference characteristics encapsulated in the cognitive mirror, the intentional claim object is formed. Based on the intent claim object and its associated cognitive mirror, a multi-agent cognitive game inference environment is initialized through the standard mandatory logical relationships and project business rules in the modular knowledge base. The policy network of each agent is parameterized and initialized by the corresponding user cognitive mirror feature vector. The simulation employs a multi-agent reinforcement learning framework based on an actor-critic architecture. Each agent iteratively generates and executes action strategies, including advocacy adjustment, argument transformation, or proposing alternatives, based on the current environmental state and the historical actions of other agents. The process involves iterative computation until an equilibrium state that satisfies all hard constraints and maximizes the capacity to accommodate the core claims of each agent converges. From this converged equilibrium state, an executable, conflict-free editing sequence is obtained as a potential consensus scheme. The evolution trajectory of each agent strategy vector, the key decision points where the claims were modified, and the fusion weight and logical reconstruction relationship of the final potential consensus solution to the initial claims are recorded synchronously throughout the entire process of deduction. A cognitive fusion graph is constructed, in which nodes represent claims and decision events, and edges are labeled with influence and fusion relationships.

6. The method for collaborative writing of international standard engineering documents based on artificial intelligence according to claim 1, characterized in that, The process of generating document evolution suggestions based on the potential consensus scheme and cognitive fusion graph, optimizing the cognitive mirror based on user decision feedback, and identifying and feeding back information on group cognitive blind spots by analyzing group decision-making patterns during the collaboration process specifically includes: Obtain the potential consensus scheme and cognitive fusion graph, extract the specific editing operations, modified text content and their logical reasoning chain contained in the potential consensus scheme, and obtain the consensus content information; By identifying key compromises, adopted core arguments, and shelved minor disagreements through cognitive fusion graphs, the graph logic information is obtained and input into a natural language generation model along with the consensus content information to transform it into document evolution suggestions that include recommended modifications. When a user makes a decision to adopt, modify and then adopt or reject the document evolution suggestion, the system captures and associates the feedback behavior, the original suggestion, the user's cognitive image snapshot and the project context, and regards the user's decision as new empirical evidence of their cognitive preferences. The system then adjusts the feature vector parameters in the user's cognitive image through the backpropagation optimization algorithm. When optimizing individual cognitive mirrors, we perform group cognitive pattern analysis on all historical collaborative interaction data in the current project. Through cluster analysis and anomaly detection algorithms, we identify potential systemic blind spots in group decision-making and obtain information on group cognitive blind spots. Based on information about blind spots in group cognition, a group cognition insight report is generated and pushed to project managers or relevant users through a collaborative platform, indicating areas where external review, supplementary research, or special discussions are needed, thereby enhancing group cognition.

7. A collaborative writing system for international standard engineering documents based on artificial intelligence, characterized in that, The system includes: a memory, a processor, and a communication interface. The memory contains a program for collaborative writing of international standard engineering documents based on artificial intelligence. When the program for collaborative writing of international standard engineering documents based on artificial intelligence is executed by the processor, it implements the steps of the collaborative writing method for international standard engineering documents based on artificial intelligence as described in any one of claims 1-6.

Images