Construction scheme document architecture adaptive adjustment method, system and electronic device

Abstract

The application provides a construction scheme document architecture adaptive adjustment method and system and electronic equipment, and relates to the technical field of document processing. The method comprises the following steps: obtaining a first compilation document and all review suggestions for the first compilation document; constructing a graph structure and building a knowledge graph based on the graph structure; updating the optimized graph structure based on the current review suggestion, updating the knowledge graph based on the updated graph structure; matching the current review suggestion with the nodes of the updated knowledge graph to obtain a matching result; adjusting the first compilation document based on the matching result; updating the updated graph structure based on the adjusted first compilation document, updating the knowledge graph based on the updated graph structure; and repeating the above steps until all review suggestions are executed to obtain a second compilation document after adaptive adjustment. The application realizes re-understanding of the adjusted document after each review suggestion is processed, and realizes dynamic content migration.

Description

Technical Field

[0001] This invention belongs to the field of document processing technology, and in particular relates to a method, system and electronic equipment for adaptive adjustment of construction plan document structure. Background Technology

[0002] The drafting of large and complex documents such as construction plans typically involves multiple stages, including initial drafting, review feedback, and secondary drafting. After the initial drafting, numerous review suggestions are often received, covering aspects such as content logic, technical details, and chapter structure. Secondary drafting is based on these review suggestions and usually involves adjusting the structure, repositioning paragraphs, adding details, and refining the language. When the initial draft is substantial, maintaining the continuity of context and semantics in the secondary draft becomes a key focus of document revision.

[0003] Traditional secondary compilation methods rely on manual review of the initial document content, followed by revisions and adjustments based on reviewer feedback. When faced with construction plans containing numerous chapters such as project overview, construction deployment, construction schedule, and resource allocation plans, manual review based on experience struggles to accurately identify the semantic connections between the content before and after adjustments. This can easily lead to logical breaks in the migrated content, resulting in insufficient or incorrect content coherence.

[0004] Furthermore, when dealing with a large number of review comments, manual processing typically requires first gaining an overall understanding of the initially drafted document, and then addressing each comment individually based on personal interpretation. After processing one comment, the document content is not re-understood; instead, the next comment is processed based on the initial understanding. The drawback of this approach is:

[0005] (1) After some of the review comments have been processed, the first draft document will be adjusted multiple times and in multiple places. Subsequent review comments are still aimed at the first draft document, which will cause problems that are difficult to effectively respond to and greatly increase the probability of operators making mistakes.

[0006] (2) The semantic relationship of the first draft document after multiple and multiple adjustments has also changed in the context, but this point was ignored in the handling of subsequent review comments. Summary of the Invention

[0007] In order to overcome the above-mentioned deficiencies of the prior art, the construction scheme document architecture adaptive adjustment method, system and electronic equipment are provided, the first compiled document is analyzed by intelligent semantic association to construct a graph structure, and then a dynamically updated knowledge graph is constructed, the knowledge graph is updated once by using the review suggestions, and the knowledge graph is updated twice by using the adjusted document, the adjusted document is re-understood after each review suggestion processing, the dynamic content migration is realized based on the matching of the review suggestions and the nodes in the knowledge graph, the document editing efficiency, quality and accuracy are improved, and the secondary compilation cost is reduced.

[0008] In order to achieve the above object, one or more embodiments of the present application provide the following technical scheme:

[0009] The first aspect of the present application provides a construction scheme document architecture adaptive adjustment method.

[0010] The construction scheme document architecture adaptive adjustment method comprises the following steps:

[0011] The first compiled document and all review suggestions for the first compiled document are obtained;

[0012] A graph structure is constructed based on the first compiled document, the node features in the graph structure are optimized by using a graph neural network model, an optimized graph structure is obtained, and a knowledge graph is constructed based on the optimized graph structure;

[0013] The types of the review suggestions are identified, the review suggestions are sorted based on a preset priority, and an instruction tree is formed;

[0014] The optimized graph structure is updated once based on the current review suggestion in the instruction tree, an updated graph structure is obtained, and the knowledge graph is updated once based on the updated graph structure;

[0015] The current review suggestion is matched with the nodes of the knowledge graph updated once, and a matching result is obtained;

[0016] The first compiled document is adjusted based on the matching result;

[0017] The updated graph structure is updated twice by using the adjusted first compiled document, and the knowledge graph is updated twice based on the updated graph structure;

[0018] The next review suggestion in the instruction tree is obtained, the updated graph structure is updated again by using the next review suggestion, and the matching and the document adjustment are performed again, and the cycle is repeated until all the review suggestions are executed, and a secondary compiled document adjusted adaptively is obtained.

[0019] The second aspect of the present application provides a construction scheme document architecture adaptive adjustment system.

[0020] The construction scheme document architecture adaptive adjustment system comprises:

[0021] The data acquisition module is configured to acquire the first compilation document and all review suggestions on the first compilation document.

[0022] The knowledge graph construction module is configured to construct a graph structure based on the first compilation document, optimize node features in the graph structure by using a graph neural network model to obtain an optimized graph structure, and construct a knowledge graph based on the optimized graph structure.

[0023] The review suggestion analysis module is configured to identify the suggestion type of the review suggestion, sort the review suggestion based on a preset priority, and form an instruction tree.

[0024] The one-time update module is configured to perform one-time update on the optimized graph structure based on the current review suggestion in the instruction tree to obtain one-time updated graph structure, and perform one-time update on the knowledge graph based on the one-time updated graph structure.

[0025] The dynamic matching module is configured to match the current review suggestion with the nodes of the one-time updated knowledge graph to obtain a matching result.

[0026] The adjustment module is configured to adjust the first compilation document based on the matching result.

[0027] The secondary update module is configured to perform secondary update on the one-time updated graph structure by using the adjusted first compilation document, and perform secondary update on the knowledge graph based on the secondary updated graph structure.

[0028] The cycle module is configured to acquire the next review suggestion in the instruction tree, perform again update on the secondary updated graph structure by using the next review suggestion, and perform again matching and document adjustment, and the cycle is repeated until all review suggestions are executed to obtain a secondary compilation document that is adaptively adjusted.

[0029] The third aspect of the present application provides an electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the construction scheme document architecture adaptive adjustment method according to the first aspect of the present application.

[0030] The above one or more technical solutions have the following beneficial effects:

[0031] The application provides a construction scheme document architecture adaptive adjustment method and system and electronic equipment, constructs a graph structure through intelligent semantic correlation analysis of first compilation documents, and further constructs a knowledge graph; review suggestions are processed in turn according to priority, the knowledge graph is updated once by using the review suggestions, dynamic content migration is realized based on matching of the review suggestions and nodes in the knowledge graph, the documents are adjusted, and document editing efficiency, quality and accuracy are improved, and secondary compilation cost is reduced; and the knowledge graph is updated again by using the adjusted documents, the adjusted documents are re-understood based on the review suggestions each time, and the document context correlation, logicality and coherence can be better understood when the next review suggestion is processed, and more accurate content migration processing is made.

[0032] The application adjusts the next review suggestion adaptively after processing the previous review suggestion, ensures correctness of entities in the suggestion content of the next review suggestion, and thus makes the adjustment more targeted and ensures efficiency and accuracy when processing the review suggestion each time.

[0033] When the graph structure is constructed, the graph structure is constructed according to the overall logical order of multi-modal fusion features→multi-head attention (multi-dimensional correlation extraction)→multi-dimensional correlation features→GRU (long-distance information optimization)→optimized node state features→calculation of feature similarity→construction of the graph structure (nodes+edges), multi-dimensional semantic information and context node correlation information can be better reflected, and semantic positioning of the nodes in the entire document can be accurately reflected.

[0034] When the knowledge graph is updated by using the review suggestion, entities and relationships are extracted from the suggestion content, it is judged whether content addition, content deletion and content modification exist in the suggestion content based on the extracted entities and relationships, the nodes are updated based on whether the content addition, content deletion and content modification exist, and the node features and edges are updated, and more accurate subsequent matching process can be realized.

[0035] The application sets an instruction tree, sorts all review instructions based on a preset priority, presents the sorted review instructions through the instruction tree, and executes each review instruction in turn according to the priority order, so that the time for realizing document adjustment is shortened and the time cost is reduced.

[0036] The application breaks through the limitation of a traditional static knowledge graph, captures the correlation between construction scheme review suggestions and existing document content in real time, dynamically updates knowledge graph nodes and edges, and realizes dynamic knowledge graph construction.

[0037] The application combines the context logic of the construction scheme chapter, such as the sequence of the construction process, the dependency relationship of each sub-project, evaluates the feasibility of content migration, and realizes the context correlation analysis mechanism. For the content with high semantic correlation degree, the cross-chapter migration and format adaptation are automatically executed. For the content with logical conflict or low correlation degree, prompt information and reference modification scheme are generated to assist the editor in decision-making, and ensure the logical coherence of the document after adjustment.

[0038] Advantages of additional aspects of the application will be given in part in the following description, some will become apparent from the following description, or be understood by practicing the application. BRIEF DESCRIPTION OF DRAWINGS

[0039] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application and explanations thereof serve to explain the application, and do not constitute improper limitations on the application.

[0040] Figure 1 The method flowchart of example one.

[0041] Figure 2 The node feature optimization flowchart in the structure of example one.

[0042] Figure 3 The knowledge graph one-time update flowchart of example one.

[0043] Figure 4 The overall structure diagram of example two. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0045] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application.

[0046] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0047] Example one

[0048] As the core guidance document of engineering construction, the complexity and professionalism of construction scheme are particularly prominent. A complete construction scheme usually contains dozens of chapters such as engineering overview, construction deployment, progress plan, technical measures, etc., which deeply integrates multi-disciplinary knowledge system such as building structure, material science and construction technology. Its compilation period is often as long as several weeks or even months, and needs to be completed by a professional team composed of project managers, technical engineers, safety officers and other professionals.

[0049] The scheme preparation process needs to go through multiple rounds of optimization iterations of "first compilation-review feedback-second compilation-second review", among which the problems raised in the initial review stage after the first compilation are the most concentrated, and the corresponding modification workload is also the largest. The second compilation is significantly different from the first compilation: the former is mainly based on the content of the initial draft for optimization, which is manifested in adjusting the architecture system, restructuring the paragraph layout, supplementing technical details, and standardizing language expression, etc., which is different from the creative content construction of the first compilation. However, when the document volume is large, the second compilation is difficult to accurately control the coherence of the context semantic context, and it is difficult for humans to accurately identify the content association logic before and after the chapter adjustment based on experience, which is easy to cause logical discontinuity of the migrated content and cause the problem of insufficient correlation between the preceding and subsequent texts.

[0050] To solve the above problems, the embodiment provides a document architecture adaptive adjustment technology based on large model semantic analysis, and discloses a construction scheme document architecture adaptive adjustment method, which plays a role in the automatic adjustment link before the second compilation of the construction scheme document, improves the document editing efficiency and quality through intelligent semantic association analysis and dynamic content migration, and reduces the second compilation cost.

[0051] As shown in Figure 1 The construction scheme document architecture adaptive adjustment method comprises the following steps:

[0052] Obtain the first compilation document and all review suggestions for the first compilation document;

[0053] Construct a graph structure based on the first compilation document, optimize the node features in the graph structure using a graph neural network model, obtain an optimized graph structure, and construct a knowledge graph based on the optimized graph structure;

[0054] Identify the suggestion type of the review suggestion, sort the review suggestions based on the preset priority, and form an instruction tree;

[0055] Update the optimized graph structure based on the current review suggestion in the instruction tree to obtain an updated graph structure, and update the knowledge graph based on the updated graph structure;

[0056] Match the current review suggestion with the nodes of the updated knowledge graph to obtain a matching result;

[0057] Adjust the first compilation document based on the matching result;

[0058] Update the updated graph structure using the adjusted first compilation document, and update the knowledge graph based on the updated graph structure;

[0059] The next review suggestion in the instruction tree is obtained, the graph structure after the secondary update is updated again using the next review suggestion, and matching and document adjustment are performed again, and the cycle is repeated until all review suggestions are executed, to obtain the secondary compiled document after adaptive adjustment.

[0060] (I) input driving data

[0061] The execution of the method of the embodiment first needs to obtain a construction scheme document and a review suggestion. The construction scheme document is the full text of the construction scheme compiled for the first time (including the contents of chapters such as project overview, construction deployment, and progress plan), which is the basic data source for all subsequent processing. The review suggestion is the modification opinion fed back in the initial review link, covering adjustment requirements such as content logic, chapter architecture, technical details (such as “add a green construction chapter” and “adjust the description position of pile foundation process parameters”), which drives the adjustment direction of the secondary compilation of the document.

[0062] (II) document analysis

[0063] Next, the first compiled document will be analyzed: the construction scheme document is structurally analyzed to extract hierarchical structures such as chapters, sections, and paragraphs, and to identify key information in the text (such as technical parameters, process steps, and terminology definitions). The main role of this step is to convert unstructured text into structured data (JSON structure) to facilitate subsequent graph neural network modeling.

[0064] (III) building a graph structure

[0065] Further, a graph structure is built based on the first compiled document, specifically including:

[0066] The technical parameters, process steps, and technical terms in the first compiled document are taken as nodes of the graph structure;

[0067] When determining whether to connect edges between the nodes of the graph structure:

[0068] The first compiled document is extracted by a text encoder and an image recognition model to obtain fusion features;

[0069] The multi-head attention mechanism is used to extract semantic associations between nodes from multiple preset dimensions, to obtain multi-dimensionally fused semantic association features, and the preset dimensions include construction process steps, technical parameters, quality requirements, and certificates and qualifications;

[0070] The GRU gating mechanism is used to process long-distance semantic associations of the multi-dimensionally fused semantic association features to obtain node features output by the GRU gating mechanism;

[0071] The similarity between the node features output by the GRU gating mechanism is calculated, and based on the similarity, it is determined whether to add an upper edge between the corresponding nodes, thereby completing the construction of the graph structure.

[0072] The overall logic of the above graph structure construction is: multi-modal fusion features -> multi-head attention (multi-dimensional association extraction) -> multi-dimensional association features -> GRU (long-distance information optimization) -> optimized node state features -> calculate feature similarity -> construct graph structure (nodes + edges).

[0073] This embodiment is based on the multi-head attention mechanism and the GRU gating recurrent unit, uses multi-head attention to extract multi-dimensional association features, and then uses GRU to optimize long-distance information transmission to construct a graph structure model of the document content. Through the logic of "multi-head attention to extract multi-dimensional association -> GRU to optimize long-distance information transmission -> determine edges based on optimized features", the precise modeling of the document graph structure is finally realized, providing reliable semantic association basis for subsequent dynamic knowledge graph updating and document adjustment. The configuration of the multi-head comes from the pre-configured dimension configuration module, which identifies the relevant dimensions in advance, and will not be described here. The specific process is as follows:

[0074] Step 1: Multi-modal feature fusion as input.

[0075] After the document content (text, image, and other multi-modal data) is extracted by the text encoder and image recognition model, the fusion feature vector is generated through a hybrid fusion strategy (initial splicing, intermediate layer attention fusion).

[0076] Step 2: Multi-head attention mechanism to extract multi-dimensional semantic associations.

[0077] The fusion feature vector is first input into the multi-head attention mechanism to capture the semantic associations between nodes from pre-set dimensions (such as construction process steps, technical parameters, quality requirements, certificates, and qualifications):

[0078] Each attention head focuses on a dimension (such as "process step sequence"), and by calculating the attention scores of the query vector (Q), key vector (K), and value vector (V), the node association features in that dimension (such as the sequence dependency relationship between step A and step B) are output.

[0079] After that, the outputs of multiple attention heads are spliced and linearly transformed to obtain multi-dimensional fusion semantic association features; the multi-dimensional fusion semantic association features contain the association strength between nodes in different dimensions.

[0080] For each node, a multi-dimensional feature vector is output, which integrates semantic information in different attention heads (dimensions) (such as "sequence association in the process step dimension" and "value dependency in the technical parameter dimension").

[0081] For example, when processing the section "Bridge Pile Foundation Construction", the output feature vector contains not only the step sequence association of "drilling → hole cleaning → pouring", but also the parameter dependency relationship of "concrete strength grade C30" and "curing time 7 days".

[0082] Step 3: GRU gating unit optimizes long-distance information transmission.

[0083] The multi-dimensional fused semantic association features output by multi-head attention are input into the GRU gating unit to optimize information transmission, focusing on cross-section and long-distance node association:

[0084] Reset gate filters historical information (such as process code "A" mentioned earlier) and retains context related to the current node;

[0085] The update gate controls the fusion ratio of new information (such as the specific description of "code A" in the current node) and historical information, avoiding redundancy or loss;

[0086] Based on the output of the reset gate and the update gate, the candidate state of the node is calculated, and the final state is generated by fusing the historical state, realizing efficient information transmission between long-distance nodes;

[0087] The final output is the optimized node state feature, which not only contains the multi-dimensional semantic information of the node itself, but also integrates the associated information of the context nodes, ensuring the consistency of long-distance context (such as the semantic uniformity of "code A" throughout the text), and accurately reflecting the semantic positioning of the node in the entire document (such as the logical dependency between a paragraph and a cross-section paragraph).

[0088] Step 4: Build graph structure edges based on GRU output.

[0089] The node features output by the GRU gating mechanism are used to calculate the similarity between nodes. When the similarity exceeds a preset threshold (such as 0.7), it is determined that there is a semantic association between the nodes, and an edge is added in the graph structure, finally forming a complete document graph structure.

[0090] For any two nodes (such as paragraph A and paragraph B), calculate the similarity (such as cosine similarity) of their GRU output feature vectors. If the similarity exceeds a preset threshold (such as 0.7), it is determined that there is a semantic association between the two, and an edge is added in the graph;

[0091] The weight of the edge can be further quantified based on the similarity value (such as similarity 0.9 corresponding to weight 0.9), which is used for subsequent evaluation of the association strength.

[0092] (Four) Graph Neural Network Optimization

[0093] The graph neural network model aggregates the adjacency information between nodes in the graph structure to optimize the node features in the graph structure.

[0094] GNN processing logic for knowledge graph:

[0095] The input of the graph neural network (GNN) is the initial graph structure of the knowledge graph (nodes, edges, initial features), and its core function is to optimize the feature vector of each node by aggregating the adjacent information of the nodes (such as the subordinate relationship between "concrete strength" and "pouring process", and the association between "green construction" and "resource allocation").

[0096] For example, before GNN processing, the features of "green construction" may only contain the literal meaning; after GNN processing, its features will integrate the information of adjacent nodes such as "resource allocation plan" and "environmental materials", more accurately reflecting its semantic connotation in the field of construction scheme. Its role is to provide more accurate node feature basis for subsequent similarity calculation.

[0097] (Five) Construction of knowledge graph

[0098] Based on the above graph structure, a knowledge graph of the construction scheme field is dynamically constructed, with nodes including terms, processes, parameters, etc. in the document, and edges representing the association between them (such as the subordinate relationship between "concrete strength" and "pouring process"). Using semantic analysis of large language models, combined with semantic similarity calculation (calculating the relationship between review suggestions and knowledge graph nodes, matching instructions to specific locations), the knowledge graph is dynamically updated to reflect the association between review suggestions and existing content (such as adding a "green construction" node and associating it to the "resource allocation plan" chapter), providing field knowledge support for subsequent adjustment strategies, and ensuring the logical accuracy of content migration.

[0099] The core of dynamic updating of the knowledge graph is to "maintain the consistency of the knowledge graph with the construction scheme content and review logic", and its update sources can indeed be divided into two parts:

[0100] (1) Dynamically adjusted document itself

[0101] When the first drafted construction scheme document is modified (such as adding a process, adjusting parameters, supplementing chapters), the knowledge graph needs to be updated synchronously for nodes and edges. For example, the document adds a "mass concrete temperature control process", and the knowledge graph needs to add a new process node and associate it with existing nodes such as "pouring temperature parameters" and "curing measures".

[0102] (2) Input review suggestions as adjustment instructions

[0103] Instructions are external trigger update signals (such as "increase green construction requirements", "optimize concrete strength grade"), and knowledge graph needs to find the association point with existing knowledge based on the semantics of the instructions and update. For example, the instruction "green construction should be included in resource allocation" needs to add a "green construction" node and establish an association edge with the "resource allocation plan" chapter.

[0104] This embodiment updates the knowledge graph from two angles. Document adjustment is the update of "content results", and instructions are the driving force of "adjustment reasons". The knowledge graph needs to reflect both "results" and "reasons and results", ensuring that subsequent adjustment strategies can trace the logic.

[0105] Instruction-driven knowledge graph update is essentially "converting natural language into structured operations of knowledge graph", which relies on semantic analysis to determine "what to update", relies on semantic similarity and domain rules to determine "where to associate", and ultimately realizes the dynamic mapping of knowledge graph to the "instruction-adjustment-knowledge association" full link, providing traceable and logically consistent domain knowledge support for subsequent construction plan adjustment strategies.

[0106] (Six) Review suggestion analysis

[0107] The semantic analysis of the review suggestions identifies the suggestion type (such as architecture adjustment, content supplement, logic optimization) and the specific adjustment object (such as a chapter, a paragraph).

[0108] Among them, with the help of large model semantic analysis, natural language suggestions are converted into machine-understandable instructions (such as "move the 'pile foundation construction technology' subsection from the 'construction deployment' chapter to the 'technical measures' chapter"), and are associated with the nodes in the knowledge graph. When associating with the knowledge graph, all parsed instructions and the order of the instructions need to be considered to determine whether the instructions are based on the current preliminary knowledge graph changes or based on the changes in the knowledge graph.

[0109] Further, after identifying the suggestion type of the review suggestions, the review suggestions are sorted based on the preset priority to form an instruction tree. Specifically, it includes:

[0110] Semantic analysis of review suggestions to identify suggestion types;

[0111] Set the priority in descending order of architecture adjustment instructions, cross-chapter adjustment instructions, chapter adjustment instructions, and content optimization instructions;

[0112] Based on the set priority of the suggestion type, all review suggestions are sorted in descending order of priority to form an instruction tree.

[0113] The instructions are initially divided into architecture adjustment instructions, chapter adjustment instructions, optimization instructions, etc. All instructions depend on semantic analysis to form a complete instruction tree, which is executed from top to bottom.

[0114] The priority of the architecture adjustment instruction is greater than that of the chapter adjustment instruction, and the priority of the content optimization instruction is the lowest. Within the architecture adjustment instruction, the priority is set according to the first-level outline adjustment instruction, the second-level outline adjustment instruction, and the cross-chapter adjustment instruction, and the dependency relationship is set.

[0115] In order to make each review suggestion targeted to adjust the document, the operation object of the review suggestion needs to be considered. After the first draft document has been adjusted multiple times and in multiple places by the previous review instructions, the first draft document after multiple adjustments has changed in context and semantic association compared to the original first draft document. The subsequent review suggestion needs to be matched with the current document again to ensure the accuracy of content adjustment.

[0116] In addition, the secondary adjustment instruction (which depends on the previous instruction) will fine-tune after the previous instruction is adjusted to ensure the correctness of the instruction. For example, the second instruction is originally to optimize the content of the second chapter, but the first instruction is to exchange the positions of the second chapter and the first chapter. After the execution of the first instruction, the second instruction becomes adjusting the content of the first chapter.

[0117] (Seven) Dynamic matching and adjustment

[0118] Based on the semantic analysis results of the knowledge graph and the graph neural network model, the association degree between the review suggestion and the document content is finally calculated to realize matching.

[0119] In order to realize dynamic matching, the following two update processes are designed before each review suggestion matching.

[0120] (1) The first update process is:

[0121] Based on the current review suggestion in the instruction tree, the optimized graph structure is updated once to obtain an updated graph structure, which specifically includes:

[0122] The review suggestions in the instruction tree are executed in order from high to low priority;

[0123] The semantic analysis of the current review suggestion in the instruction tree is performed to obtain the suggestion content, and the entities and relationships are extracted from the suggestion content;

[0124] Based on the extracted entities and relationships, it is judged whether there is content addition, content deletion, and content modification in the suggestion content;

[0125] updating the nodes based on whether there is content addition, content deletion and content modification: if there is content addition, a new node is created in the graph structure; if there is content deletion, the corresponding node is deleted in the graph structure; if there is content modification, the node features of the corresponding node are updated in the graph structure;

[0126] recomputing the similarity between nodes and updating whether to connect edges between nodes, thereby completing an update of the optimized graph structure.

[0127] Then, the knowledge graph is updated based on the updated graph structure, and the current review suggestion is matched with the nodes of the updated knowledge graph to obtain a matching result, which specifically includes:

[0128] extracting the feature vector of the review suggestion;

[0129] calculating the similarity between the feature vector of the review suggestion and the feature vector of the node in the updated knowledge graph, and screening the nodes with a similarity greater than a set threshold as the matching result matched with the review suggestion.

[0130] Further, based on the matching result, the first drafted document is adjusted, specifically including:

[0131] screening the nodes with a similarity greater than a set threshold as high correlation nodes, and screening the nodes with a similarity less than a set threshold as low correlation nodes;

[0132] taking the content of the first drafted document corresponding to the high correlation nodes as a part to be adjusted, planning a migration path or optimizing (such as moving cross-chapter content) for the part to be adjusted, and generating a prompt (such as "the parameter description here is inconsistent with the progress plan, which needs to be confirmed manually") for the low correlation nodes or conflicting content;

[0133] outputting adjustment instructions (such as migration position, format adaptation requirements) for the part to be adjusted. The execution order of the migration instructions depends on the instruction dependency relationship;

[0134] performing intelligent migration or optimization (such as paragraph copying, chapter reorganization) of the part to be adjusted according to the adjustment instructions, and automatically adapting the format (such as title level, numbering rules);

[0135] optimizing the content logic after migration, and verifying the context coherence (such as ensuring that the terms before and after the migrated paragraph are consistent) through a graph neural network model;

[0136] outputting the adjusted construction plan and an adjustment report containing adjustment tracks and semantic correlation degree evaluation results, facilitating problem troubleshooting.

[0137] (2) The second update process is:

[0138] The second update of the graph structure after the first update is performed by using the adjusted first compilation document, specifically including:

[0139] The nodes of the graph structure after the first update are re-determined based on the adjusted first compilation document, the node features are calculated, and it is determined whether to connect edges between the nodes, so as to realize the second update of the graph structure after the first update.

[0140] In the matching process, in general, the review suggestion will first be parsed into structured features (such as keywords, entities, relationships) and semantic features (such as intent, contextual meaning). For example, the instruction "increase the humidity parameter of concrete maintenance" will extract the entity "concrete maintenance" "humidity parameter", the relationship "increase", and the semantic intent "supplement process parameters". The node feature representation in the knowledge graph (such as "concrete maintenance" "paving process") not only contains the literal features (terms themselves), but also contains the deep semantic features generated by the pre-training model or GNN (such as the position of the node in the graph, the associated relationship, the domain classification, etc.). For example, the features of "concrete maintenance" may include the "post-construction processing" category it belongs to, and the adjacent node information such as "temperature parameter" "maintenance duration". The core of the similarity calculation is to find the most matched node by calculating the similarity between the "instruction (review suggestion) feature vector" and the "node feature vector". For example, the "humidity parameter" in the instruction will be compared with the "parameter class" sub-node of the "concrete maintenance" node in the knowledge graph to determine the semantic association strength.

[0141] After the preliminary processing of the review suggestion, similarity calculation is performed with the document content (knowledge graph formed by combining the document, neural network), and a specific suggestion instruction is matched to a specific knowledge graph node.

[0142] In this embodiment, the semantic features of the instruction after parsing are used as the basis for similarity calculation with the knowledge graph node features optimized by the graph neural network, and the association degree of the review suggestion and the document content is finally determined by combining the association relationship weight of the nodes in the graph. Among them, the GNN optimized node features are used to improve the matching accuracy, and the essence of the association degree is the comprehensive matching result of "instruction semantics" and "graph node semantics + structural relationship".

[0143] (Eight) Instruction execution in sequence

[0144] For multiple review suggestions in the instruction tree, after the current review suggestion is processed, the next one needs to be executed in sequence. The overall process can be simply described as:

[0145] 1. Identify the instruction;

[0146] 2. Layer the instruction (adjust the instruction of the architecture, adjust the instruction within the chapter, and optimize the content of the instruction).

[0147] 3. Architecture adjustment instruction priority is highest, while subdivided into cross-chapter adjustment instructions, chapter-in architecture adjustment instructions;

[0148] 4. According to the dependency relationship of cross-chapter adjustment instructions, chapter-in architecture adjustment instructions, chapter-in adjustment instructions (chapter-in paragraph), content optimization instructions in turn, a complete instruction tree is formed;

[0149] 5. The instruction is matched on the knowledge graph and graph neural network;

[0150] 6. Execute instruction A instruction, adjust the document;

[0151] 7. Identify the subsequent instruction adjustment caused by the change of instruction A;

[0152] 8. Execute B instruction loop 6, 7 steps;

[0153] 9. All instruction content adjustment is completed.

[0154] After the current review suggestion is executed, the next review suggestion in the instruction tree is obtained, and the graph structure after the secondary update is updated again using the next review suggestion, and matching and document adjustment are performed again, and the cycle is repeated until all review suggestions are executed. The self-adaptive adjusted secondary compiled document is obtained, which specifically includes:

[0155] Based on the adjustment result of the last review suggestion on the first compiled document, the suggestion content of the next review suggestion in the instruction tree is adjusted accordingly to ensure the correctness of the entity in the suggestion content of the next review suggestion;

[0156] Extract the entity and relationship from the adjusted suggestion content of the next review suggestion;

[0157] Based on the extracted entity and relationship, update the nodes of the secondary updated graph structure, and then update the edges between the nodes to complete the secondary update of the secondary updated graph structure;

[0158] Based on the graph structure updated again, the secondary updated knowledge graph is updated again;

[0159] Extract the feature vector of the adjusted next review suggestion;

[0160] Calculate the similarity between the feature vector of the adjusted next review suggestion and the feature vector of the node in the knowledge graph updated again, and select the node with a similarity greater than a set threshold as the matching result matched with the next review suggestion.

[0161] (Nine) Output

[0162] Adjusted construction plan: output the construction plan document based on the review suggestions to complete the architecture and content adjustment, ensure the logical coherence of the chapters and the completeness of the content.

[0163] Logical verification report: provide semantic association verification results of the adjustment process, including logical rationality analysis of content migration and potential conflict prompts to assist manual review.

[0164] (10) Algorithm implementation

[0165] (1) Multi-head attention calculation algorithm:

[0166] Multiply the input document content vector by the query matrix Q, key matrix K, and value matrix V to get the query vector q, key vector k, and value vector v. Calculate the attention score by the following formula :

[0167]

[0168] where d k is the key vector dimension, and after the Softmax function is converted into a probability distribution, the value vector is weighted and summed to obtain the output vector processed by the attention mechanism.

[0169] (2) GRU message passing algorithm:

[0170] When updating the node state, control the flow of information through the update gate and the reset gate. The update gate is represented as:

[0171]

[0172] Reset gate:

[0173]

[0174] where is the Sigmoid function, W z , W r are weight matrices, h t-1 is the node state at the previous time, x t is the current input, z t , r t are the update gate and the reset gate, respectively.

[0175] Through the candidate state and the final node state, information is effectively transmitted and fused.

[0176] Candidate state:

[0177]

[0178] Final node state:

[0179]

[0180] wherein, represents the candidate state; represents the final node state; W represents a weight matrix.

[0181] (3) Semantic similarity calculation algorithm

[0182] The semantic embedding method based on the pre-trained language model is adopted to map the document content and the review suggestion, and the similarity between vectors is calculated through the cosine similarity formula:

[0183]

[0184] wherein, A and B are semantic embedding vectors, , is the vector norm, which accurately quantifies the content correlation degree. The value of 0.7 is the preset threshold value in the embodiment, which is the best value after the current multiple verification matching degree.

[0185] (4) Dynamic graph update algorithm

[0186] When it is detected that the document content has changed, the knowledge graph is updated according to the similarity calculation result. The corresponding node is created and the feature vector is initialized when the new content is added, and it is judged whether to add the edge through the similarity threshold (0.7 above is the similarity threshold); when the content is deleted, the related node and edge are removed; when the content is modified, the node feature vector is updated, and the graph structure is optimized to remove redundant elements, so as to ensure that the knowledge graph efficiently reflects the document adjustment requirements.

[0187] Next, the drawings will be described as follows:

[0188] (1) As shown in Figure 2 , it is a flow chart of node feature optimization in the graph structure. It is realized based on the multi-head attention mechanism, the gated recurrent unit and the graph neural network model:

[0189] Multi-modal feature fusion: The features output by the text encoder and the image recognition model are combined, and through the hybrid fusion strategy, the correlation degree and the calculation comprehensiveness are improved based on the attention mechanism fusion in the early feature extraction, the middle layer and the output layer, so as to accurately process the multi-modal content such as the text of the construction scheme and the drawings.

[0190] The input document content is extracted based on the image recognition model and the text encoder, specifically, the text feature vector is extracted by the text editor, the multi-text feature vector is spliced, and the image feature vector recognized by the image recognition model is fused to obtain the multi-modal fusion feature.

[0191] Multi-head attention mechanism: pre-set dimension range, identify the dimensions used in the document, then capture the semantic association of the document content through multiple dimensions, each dimension (such as construction process steps, technical parameters, quality requirements, certificates and qualifications) is an attention head. Each attention head can accurately locate the key content. For example: when processing the "bridge pile foundation construction process" chapter, some attention heads focus on the sequence logic between process steps, and some focus on the value range of technical parameters.

[0192] The multi-modal fusion features are input into the multi-head attention mechanism, the multi-dimensional features of the fusion features are extracted and spliced, and the output of the multi-head attention is obtained.

[0193] Gated recurrent unit (GRU): incorporate gated recurrent unit into message passing process, control information flow through update gate and reset gate. In the construction scheme document, chapters such as "construction schedule plan" and "resource allocation plan" are closely related. The gating mechanism of the gated recurrent unit (GRU) can prevent information loss or redundancy during transmission, ensuring efficient propagation of contextual information between long-distance nodes and accurate capture of cross-chapter semantic associations. For example: if a certain construction process appears in the second chapter and is given the code A, then all subsequent chapters with the A code represent a certain construction process.

[0194] The output of the multi-head attention is input into the gated recurrent unit, and the long-distance semantic association is processed to obtain the output of the gated recurrent unit. Calculate the similarity between the node features of the GRU gated recurrent unit output, and determine whether to add an upper edge between the corresponding nodes based on the similarity, thereby completing the construction of the graph structure.

[0195] Graph neural network model: optimize node features based on graph neural network model for graph structure, and finally obtain document semantic representation vector.

[0196] (2) As shown in Figure 3 , it is a flowchart of knowledge graph update. All review comments are analyzed for suggestion content using large models and other technologies, entities and relationships are extracted from the suggestion content, and then trigger dynamic update of the knowledge graph. For example, review suggestion 1 is concrete pouring process optimization, and review suggestion 2 is to add green implementation requirements. The above review suggestions will trigger updates.

[0197] When updating specifically, whether to add new content (chapters, sections, etc.), whether to modify content, and whether to delete content are judged layer by layer. When there is new content, create a new node; when there is content modification, update the node features; when there is content deletion, move the node and the edge.

[0198] Recalculate the semantic similarity of the updated nodes, and determine whether to add edges: when the similarity is greater than the threshold, add the upper edge or update the edge, when the similarity is less than the threshold, do not add the edge. Further optimize the graph structure.

[0199] Update the dynamic knowledge graph based on the optimized graph structure.

[0200] Embodiment two

[0201] The embodiment discloses a construction scheme document architecture adaptive adjustment system.

[0202] The construction scheme document architecture adaptive adjustment system comprises:

[0203] The data acquisition module is configured to: acquire the first drafted document and all review suggestions for the first drafted document;

[0204] The knowledge graph construction module is configured to: construct a graph structure based on the first drafted document, optimize the node features in the graph structure by using a graph neural network model to obtain an optimized graph structure, and construct a knowledge graph based on the optimized graph structure;

[0205] The review suggestion analysis module is configured to: identify the suggestion type of the review suggestion, sort the review suggestions based on a preset priority, and form an instruction tree;

[0206] The one-time update module is configured to: perform one-time update on the optimized graph structure based on the current review suggestion in the instruction tree to obtain a one-time updated graph structure, and perform one-time update on the knowledge graph based on the one-time updated graph structure;

[0207] The dynamic matching module is configured to: match the current review suggestion with the nodes of the one-time updated knowledge graph to obtain a matching result;

[0208] The adjustment module is configured to: adjust the first drafted document based on the matching result;

[0209] The secondary update module is configured to: perform secondary update on the one-time updated graph structure by using the adjusted first drafted document, and perform secondary update on the knowledge graph based on the secondary updated graph structure;

[0210] The cycle module is configured to: acquire the next review suggestion in the instruction tree, perform further update on the secondary updated graph structure by using the next review suggestion, and perform matching and document adjustment again, and the cycle is repeated until all review suggestions are executed to obtain a secondary drafted document that is adaptively adjusted.

[0211] It can be understood that the embodiment also comprises a document analysis module, a graph neural network modeling module, and a content migration and optimization module, wherein:

[0212] a document parsing module configured to parse document content;

[0213] a graph neural network modeling module configured to model a graph neural network, including a multi-head attention mechanism and a gated recurrent unit;

[0214] a content migration and optimization module configured to output an adjusted construction scheme and a logical verification report.

[0215] As shown in Figure 4 the first time the prepared document and review suggestions of the embodiment are input into the input layer, and then data processing is performed through the processing layer, specifically including: using the document parsing module to parse the document content, performing feature extraction transmission of the document parsing structure based on the multi-head attention mechanism and the gated recurrent unit, obtaining multi-dimensional features and context association, constructing a graph structure, and using the graph neural network modeling module to optimize the node features in the graph structure; based on the optimized graph structure, constructing a knowledge graph, using the review suggestion parsing module to parse the review suggestions, and dynamically updating the knowledge graph, and performing semantic similarity calculation; using the dynamic matching and adjustment module to match the review suggestions to the knowledge graph nodes with high correlation, screening out the nodes with high correlation, migrating and optimizing the document content corresponding to the nodes with high correlation; and finally outputting the adjusted construction scheme and the logical verification report through the content migration and optimization module.

[0216] Embodiment Three

[0217] An object of the embodiment is to provide an electronic device.

[0218] The electronic device includes a memory, a processor, and a program stored on the memory and executable on the processor, and the processor executes the program to implement the steps in the construction scheme document architecture adaptive adjustment method according to Embodiment 1 of the present disclosure.

[0219] The steps and methods involved in the above embodiments two and three correspond to Embodiment One, and the specific implementation can refer to the relevant description part of Embodiment One. The term "computer readable storage medium" should be understood to include a single medium or multiple media of one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry instruction sets for execution by a processor and cause the processor to perform any method in the present application.

[0220] Those skilled in the art should understand that the modules or steps of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device and executed by a computing device, or they can be respectively made into individual integrated circuit modules, or a plurality of modules or steps among them can be made into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0221] Although the specific embodiments of the present application are described above in combination with the drawings, the description is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for adaptive adjustment of construction plan documentation architecture, characterized in that, The method comprises the following steps: obtaining a first compilation document and all review suggestions on the first compilation document; constructing a graph structure based on the first compilation document, optimizing node features in the graph structure by using a graph neural network model to obtain an optimized graph structure, and constructing a knowledge graph based on the optimized graph structure; identifying the suggestion type of the review suggestion, sorting the review suggestions based on the preset priority, and forming an instruction tree; updating the optimized graph structure based on the current review suggestion in the instruction tree to obtain an updated graph structure, and updating the knowledge graph based on the updated graph structure; matching the current review suggestion with the nodes of the updated knowledge graph to obtain a matching result; adjusting the first compilation document based on the matching result; updating the updated graph structure based on the adjusted first compilation document, and updating the knowledge graph based on the updated graph structure; obtaining the next review suggestion in the instruction tree, updating the updated graph structure based on the next review suggestion, and matching and adjusting the document again, and repeating until all review suggestions are executed, to obtain a second compilation document adjusted adaptively; constructing a graph structure based on the first compilation document, specifically comprising: taking technical parameters, process steps, and technical terms in the first compilation document as nodes of the graph structure; when judging whether to connect edges between the nodes of the graph structure: extracting and fusing basic features of the first compilation document by using a text encoder and an image recognition model to obtain fused features; extracting semantic associations between nodes from multiple preset dimensions by using a multi-head attention mechanism to obtain multi-dimensionally fused semantic association features, wherein the preset dimensions include construction process steps, technical parameters, quality requirements, and certificates and qualifications; processing long-distance semantic associations of the multi-dimensionally fused semantic association features by using a GRU gating mechanism to obtain node features output by the GRU gating mechanism; calculating the similarity between the node features output by the GRU gating mechanism, and determining whether to add edges between the corresponding nodes based on the similarity, thereby completing the construction of the graph structure.

2. The construction plan document architecture self-adapting adjustment method of claim 1, wherein, The graph neural network model aggregates adjacency information between nodes of the graph structure to optimize the node features in the graph structure.

3. The construction plan document architecture adaptation method of claim 1, wherein, Identifying the suggestion type of the review suggestion, sorting the review suggestions based on the preset priority, and forming an instruction tree, specifically comprising: performing semantic analysis on the review suggestion to identify the suggestion type; setting the priority in descending order of the priority of the architectural adjustment instruction, the cross-chapter adjustment instruction, the intra-chapter adjustment instruction, and the content optimization instruction; sorting all review suggestions in descending order of priority based on the set priority of the suggestion type to form an instruction tree.

4. The construction plan document architecture adaptation method of claim 1, wherein, Updating the optimized graph structure based on the current review suggestion in the instruction tree to obtain an updated graph structure, specifically comprising: the review suggestions in the instruction tree are executed in descending order of priority; performing semantic analysis on the current review suggestion in the instruction tree to obtain suggestion content, and extracting entities and relationships from the suggestion content; Based on the extracted entities and relationships, it is determined whether there is content addition, content deletion, and content modification in the suggestion content; Based on whether there is content addition, content deletion, and content modification, the nodes are updated: if there is content addition, a new node is created in the graph structure; if there is content deletion, the corresponding node is deleted in the graph structure; if there is content modification, the node features of the corresponding node are updated in the graph structure; The similarity between nodes is recalculated, and the edges between nodes are updated, thereby completing an update of the optimized graph structure.

5. The construction plan document architecture adaptation method of claim 1, wherein, The current review suggestion is matched with the nodes of the knowledge graph after one update to obtain a matching result, which specifically includes: extracting the feature vector of the review suggestion; Calculate the similarity between the feature vector of the review suggestion and the feature vector of the node in the knowledge graph after one update, and select the nodes with a similarity greater than a certain threshold as the matching results matched with the review suggestion. Or, Based on the matching result, the first drafted document is adjusted, specifically including: selecting the nodes with a similarity greater than a certain threshold as high correlation nodes, and selecting the nodes with a similarity less than a certain threshold as low correlation nodes; the content of the first drafted document corresponding to the high correlation nodes is taken as the part to be adjusted, and the part to be adjusted is planned to migrate or optimized, and a prompt is generated for the low correlation nodes or conflicting content; output an adjustment instruction for the part to be adjusted; According to the adjustment instruction, the intelligent migration or optimization of the part to be adjusted is performed, and the format is automatically adapted; Optimize the logic of the migrated content and verify the context coherence through a graph neural network model.

6. The construction plan document architecture adaptation method of claim 1, wherein, Use the adjusted first drafted document to perform a second update on the graph structure after one update, specifically including: Based on the adjusted first drafted document, the nodes of the graph structure after one update are re-determined, the node features are calculated, and it is determined whether to connect edges between nodes, thereby realizing a second update on the graph structure after one update.

7. The construction plan document architecture adaptation method of claim 4, wherein, Get the next review suggestion in the instruction tree, use the next review suggestion to update the graph structure after two updates again, and perform matching and document adjustment again, specifically including: Based on the adjustment result of the first drafted document according to the previous review suggestion, the suggestion content of the next review suggestion in the instruction tree is adjusted accordingly to ensure the correctness of the entities in the suggestion content of the next review suggestion; extract entities and relationships from the suggestion content of the adjusted next review suggestion; Based on the extracted entities and relationships, the nodes of the graph structure after two updates are updated, and then the edges between the nodes are updated to complete the re-update of the graph structure after two updates; Based on the re-updated graph structure, the knowledge graph after two updates is updated again; extract the feature vector of the adjusted next review suggestion; Calculate the similarity between the feature vector of the adjusted next review suggestion and the feature vector of the node in the knowledge graph after two updates, and select the nodes with a similarity greater than a certain threshold as the matching results matched with the next review suggestion.

8. A construction plan document architecture adaptive adjustment system, characterized by, including: a data acquisition module configured to acquire a first drafted document and all review suggestions for the first drafted document; The knowledge graph construction module is configured to: construct a graph structure based on the first drafted document, optimize node features in the graph structure by using a graph neural network model to obtain an optimized graph structure, and construct a knowledge graph based on the optimized graph structure; The review suggestion analysis module is configured to: identify the suggestion type of the review suggestion, sort the review suggestions based on a preset priority to form an instruction tree; The one-time update module is configured to: perform one-time update on the optimized graph structure based on the current review suggestion in the instruction tree to obtain one-time updated graph structure, and perform one-time update on the knowledge graph based on the one-time updated graph structure; The dynamic matching module is configured to: match the current review suggestion with the nodes of the one-time updated knowledge graph to obtain a matching result; The adjustment module is configured to: adjust the first drafted document based on the matching result; The secondary update module is configured to: perform secondary update on the one-time updated graph structure by using the adjusted first drafted document, and perform secondary update on the knowledge graph based on the secondary updated graph structure; The cycle module is configured to: obtain the next review suggestion in the instruction tree, perform re-update on the secondary updated graph structure by using the next review suggestion, and perform matching and document adjustment again, and the cycle is repeated until all review suggestions are executed to obtain a second drafted document that is adaptively adjusted; The graph structure is constructed based on the first drafted document, and specifically includes: The technical parameters, process steps and technical terms in the first drafted document are taken as nodes of the graph structure; When judging whether edges are connected between the nodes of the graph structure: The first drafted document is respectively extracted by a text encoder and an image recognition model to obtain basic features and fusion features; The multi-head attention mechanism is used to extract semantic associations between nodes from multiple preset dimensions to obtain multi-dimensionally fused semantic association features, and the preset dimensions include construction process steps, technical parameters, quality requirements and certificate qualifications; The long-distance semantic associations of the multi-dimensionally fused semantic association features are processed by using the GRU gating mechanism to obtain node features output by the GRU gating mechanism; The similarity between the node features output by the GRU gating mechanism is calculated, and based on the similarity, it is determined whether to add edges between the corresponding nodes, thereby completing the construction of the graph structure.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized by The processor executes the program to implement the steps in the construction scheme document architecture adaptive adjustment method of any one of claims 1-7.

Images