Engineering industry document intelligent review system based on large language model

Abstract

The invention relates to the technical field of artificial intelligence and engineering management, in particular to an intelligent review system for engineering industry documents based on a large language model, which comprises a multi-level review point library, a document access module, a document classification module, a large language model processing module and a review report generation module, the multi-level review point library is in four-level classification, covers review dimensions such as multi-service fields, professional classification, project stages and integrity, and provides review standards; the document access module supports multi-type document and multi-channel access; the document classification module performs automatic classification based on a BERT model of engineering corpus fine adjustment; the large language model processing module matches the examination point list, performs slice examination and outputs opinions; the review report generation module summarizes the opinions to generate a structured report, so that the whole-process intelligent review of the document is realized; according to the method, the problems of low efficiency, long time consumption of complicated documents, non-uniform standard and poor result consistency caused by difference of standard understanding of reviews in traditional manual review are solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and engineering management technology, specifically relating to an intelligent document review system for the engineering industry based on a large language model. Background Technology

[0002] In the engineering industry, document review is a crucial step in ensuring project quality, safety, and compliance. It is widely applicable across various engineering sectors, including water conservancy, energy, urban construction, industrial engineering, and digital engineering, covering different stages such as project design, construction, and operation and maintenance. It is of great significance for the smooth progress of engineering projects. Existing technologies, such as some systems, achieve preliminary review through keyword matching or rule engines, but suffer from insufficient flexibility (unable to handle semantic ambiguity) and poor adaptability (difficult to cover the complex regulations of the engineering industry). Specifically, these include: (1) Inefficient: Relying on manual word-by-word verification, the review of a complex construction plan may take several days or even weeks; (2) Inconsistent standards: Different reviewers have different understandings of the specifications, resulting in poor consistency in review results; (3) Incomplete coverage: Manual review is prone to missing details (such as parameter conflicts, format errors), especially in documents with multiple disciplines; (4) High dependence on experience: There is a high degree of dependence on senior experts, and small and medium-sized institutions cannot afford high-cost review resources.

[0003] In view of this, the present invention is hereby proposed. Summary of the Invention

[0004] To address the aforementioned technical problems in existing technologies, this invention provides an intelligent document review system for the engineering industry based on a large language model. This system solves the problems of low efficiency and time-consuming manual review of complex documents, and eliminates the issues of inconsistent review standards and poor consistency of results caused by differences in the understanding of specifications among different reviewers.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: An intelligent document review system for the engineering industry based on a large language model includes: A multi-level review point library is used to provide the review standards required for reviewing engineering documents; The document access module is used to access documents awaiting review from the engineering industry. The document classification module is used to automatically classify the documents to be reviewed that are accessed by the document access module, and is configured with a manual verification interface to verify and modify the automatic classification results; The large language model processing module is configured to automatically match the review point list in the multi-level review point library based on the classification result of the document classification module, process the to-be-reviewed document, and then carry out in-depth review, and output a review opinion; The review report generation module is configured to summarize the review opinions output by the large language model processing module, and generate a review report.

[0006] Further, the multi-level review point library is a four-level classification structure, specifically including: The first business field is divided into water conservancy engineering, energy engineering, urban construction, industrial engineering, and digital engineering; The second professional classification is further divided based on each first business field, wherein the water conservancy engineering corresponds to hydrogeology, hydraulic structure, water conservancy machinery, and water conservancy electrical; the energy engineering corresponds to energy power, electrical system, and new energy technology; the urban construction corresponds to building structure, water supply and drainage, heating ventilation air conditioning, and road and bridge; the industrial engineering corresponds to mechanical manufacturing, automation control, and industrial building; and the digital engineering corresponds to software engineering, data analysis, and Internet of Things technology. The third project stage is divided into a design stage, a construction stage, an operation and maintenance stage, and an acceptance stage. The fourth review dimension is divided into integrity, consistency, accuracy, and compliance.

[0007] Further, the review point list matched by the large language model processing module includes a general review point list and a special review point list. The general review point list is applicable to all engineering documents and covers integrity, consistency, accuracy, and compliance review. The special review point list is set based on the first business field, the second professional classification, and the third project stage corresponding to the to-be-reviewed document.

[0008] Further, the to-be-reviewed document types supported by the document access module include text documents, drawing documents, data table documents, and scanned document types. The text documents include Word documents and PDF documents. The drawing documents include CAD drawings and BIM model related files. The data table documents include Excel documents. The scanned document type is an image format document converted into editable text by OCR technology.

[0009] Further, the access channels supported by the document access module include a local upload channel, a database access channel, and a third-party platform access channel. The local upload channel supports batch uploading and has upload progress display and error prompt functions. The database access channel is connected with the enterprise internal database through configuration of connection parameters and query statements to obtain and synchronize the documents to be reviewed. The third-party platform access channel is integrated with the engineering management platform and the cloud storage platform through an API interface.

[0010] Further, the automatic classification of the document classification module is based on a text classification algorithm fine-tuned by engineering industry expertise and corpus, which is based on a BERT model, analyzes the keywords, sentence structure and semantic relationship of the document to be reviewed, and outputs the corresponding first-level business field, second-level professional classification and third-level project phase of the document. If the document to be reviewed has a classification label, the document classification module will also compare and verify the self-provided label with the automatic classification result. If they are inconsistent, the self-provided label will be used as the standard, and the classification difference will be prompted. The artificial verification interface supports user modification of the classification result, and the modification record is used for subsequent model optimization training.

[0011] Further, the large language model processing module processes the document to be reviewed in slices, uses sliding window technology, sets window size and step, and there is an overlapping part between adjacent slices to ensure the coherence of the context semantics.

[0012] Further, for the document to be reviewed containing charts, the chart and its related explanatory text are taken as a whole during the slicing process.

[0013] Further, in the deep review process of the large language model processing module, for the review points that match successfully, the consistency of the content of the document to be reviewed and the review point standard is analyzed. If there are non-conforming items, the review opinion generated includes problem description, relevant standard basis and improvement suggestion.

[0014] Further, the review report generated by the review report generation module is a structured report, including document basic information, review result summary, detailed review opinion, and review summary and suggestion. The review report supports PDF and Word format output. It also supports adding new business fields and review points, and fine-tuning the large language model to adapt to new review standards.

[0015] Compared with the prior art, the above-mentioned engineering industry document intelligent review system based on a large language model provided by the present application comprises a multi-level review point library, a document access module, a document classification module, a large language model processing module and a review report generation module; the multi-level review point library adopts a four-level classification structure, covers multiple business fields, professional classifications, project stages and integrity, consistency, accuracy, compliance review dimensions, and provides review standard basis; the document access module supports multiple types of engineering documents and multiple channel access; the document classification module automatically classifies based on a BERT model fine-tuned by an engineering industry corpus, and is configured with a manual verification interface; the large language model processing module matches general and special review point lists, deeply reviews and outputs opinions after slicing the document; and the review report generation module generates a structured report by summarizing the opinions and supports PDF and Word format output. The system realizes intelligent review of engineering documents throughout the whole process; the present application solves the problems of low efficiency of traditional manual review, long time consumption of complex documents, eliminates the problems of non-uniform review standards and poor consistency of results caused by differences in understanding of specifications by different reviewers, makes up for the defects of easy omission of parameter conflicts, format errors and other details (especially for multi-specialty cross-document) and incomplete coverage by manual review, reduces the high dependence of review work on senior experts to alleviate the cost pressure of small and medium-sized institutions, and improves the conditions that the existing preliminary automatic review system relying on keyword matching or rule engine is difficult to handle semantic ambiguity, adapt to complex specifications, and has poor flexibility and adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The architecture diagram of the intelligent review system provided by the embodiment of the present application is shown in the figure. Figure 2 The flowchart of automatic classification of the document classification module provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions of the present application will be described clearly in conjunction with the accompanying drawings. Obviously, the described embodiments are not all the embodiments of the present application, and all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0018] It should be noted that, unless otherwise specified, the relative arrangement, numerical expression of the components and steps set forth in these embodiments should not be understood as limiting the scope of the present application.

[0019] The following description of example embodiments is merely illustrative in nature and is in no way intended to limit the application or its application or uses. Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered part of the specification if applicable.

[0020] Embodiment one Reference Figure 1 , Figure 1 An architecture diagram of an intelligent review system for engineering industry documents based on a large language model is proposed for the present application, and the specific steps can include: M1, a multi-level review point library, is used to provide the review standards required for engineering document review; the multi-level review point library is a four-level classification structure, which specifically includes: M11, a first business field, is divided into water conservancy engineering, energy engineering, urban construction, industrial engineering, and digital engineering. Water conservancy engineering covers reservoirs, river regulation, irrigation systems, and other projects; energy engineering includes thermal power, hydroelectric power, wind power, and photovoltaic energy development and utilization projects; urban construction involves building engineering, municipal engineering, transportation engineering, and other urban infrastructure construction; industrial engineering includes various industrial plant construction, production line planning, and other projects; digital engineering focuses on informationization and intelligentization projects in the engineering field, such as smart construction site systems and engineering management software; M12, a second professional classification, is further divided into professionals according to each business field. In water conservancy engineering, professional classification includes hydrogeology, hydraulic structure, water conservancy machinery, and water conservancy electrical equipment; energy engineering can be divided into energy power, electrical system, and new energy technology professionals; the professional classification of urban construction includes building structure, water supply and drainage, heating ventilation and air conditioning, and road and bridge; industrial engineering includes mechanical manufacturing, automation control, and industrial building professionals; digital engineering involves software engineering, data analysis, and Internet of Things technology professionals; M13, a third project stage, is divided into design stage, construction stage, operation and maintenance stage, and acceptance stage. The design stage includes project conceptual design, preliminary design, and construction drawing design documents; the construction stage involves construction organization design, special construction scheme, and engineering progress report; the operation and maintenance stage covers equipment maintenance manual, fault handling report, and operation monitoring data; the acceptance stage includes completion acceptance report and quality evaluation report documents; M14, a fourth review dimension, includes integrity, consistency, accuracy, and compliance. Integrity review checks whether the document contains all necessary components and information; consistency checks whether the data, terminology, format, etc. within the document and between different documents are consistent; accuracy ensures that the data, calculations, descriptions, etc. in the document are accurate; compliance verifies whether the document complies with national laws and regulations, industry standards, and enterprise internal regulations.

[0021] Take the water conservancy project-construction phase-civil engineering structure specialty as an example, in the integrity dimension, whether the construction scheme contains the description of the key construction links such as foundation treatment, concrete pouring and dam filling; in the consistency dimension, whether the expression of concrete strength grade in construction drawings, material list and construction scheme is consistent; in the accuracy dimension, whether the calculation results of dam stress and strain are correct; in the compliance dimension, whether the construction technology conforms to the relevant standards such as “Water Conservancy and Hydropower Engineering Construction Quality Inspection and Evaluation Regulations”.

[0022] M2, document access module, used for accessing the engineering industry documents to be reviewed, the system supports multiple document types access through different channels; the system supports common engineering document types, including but not limited to: M21, text type: such as Word document (.doc,.docx) for writing construction scheme, technical report, etc.; PDF document (.pdf) is often used to publish standard specifications, design drawing instructions, etc.; M22, drawing type: CAD drawing (.dwg,.dxf) is the most commonly used drawing format in engineering design, which accurately expresses engineering structure, size, layout, etc.; BIM model related files (such as.ifc format), which can provide three-dimensional visual building information model, contain rich engineering data and information; M23, data table type: Excel document (.xls,.xlsx) is often used to record engineering data, such as bill of quantities, material procurement table, engineering progress data, etc., which is convenient for data sorting, calculation and analysis; M24, scanned document type: for some paper documents, after scanning and converting to image format (such as.jpg,.png), OCR technology is used to convert them into editable text information, which is included in the system review range. (2) At the same time, the system supports multi-channel access mode: M25, local upload: users can upload documents stored locally through the file upload function of the system interface. Batch upload is supported to improve the efficiency of document access, and upload progress display and error prompt functions are provided to facilitate users to understand the upload status in time. For example, users can upload multiple construction drawing files at a time. M26, database access: the system can be connected with the enterprise internal database to read the stored engineering documents directly from the database. By configuring database connection parameters and query statements, automatic acquisition and synchronous update of documents are realized.

[0023] M27, Third-party platform access: Through the open API interface, the system can be integrated with common engineering management platforms (such as Luban engineering management platform, etc.), cloud storage platforms (such as Baidu cloud, Tencent cloud, etc.). Users can directly push documents to the review system in the third-party platform, or directly obtain documents on the third-party platform in the review system, realizing seamless data circulation. For example, after completing the construction progress plan in the Luban engineering management platform, the document can be pushed to the review system for review.

[0024] M3, Document classification module, for automatically classifying the documents to be reviewed accessed by the document access module, and configuring a manual verification interface to verify and modify the automatic classification results. Referring to Figure 2 , specifically comprising: M31, Document automatic classification: the system uses a text classification algorithm based on deep learning, combined with professional knowledge and corpus in the engineering industry for training. Based on the BERT model, the model is fine-tuned to accurately understand the semantics and structure of engineering documents. The model input is the text content of the document, which learns and analyzes the features such as keywords, sentence structure, semantic relationship in the text, and outputs the first-level business field, second-level professional classification and third-level project stage of the document.

[0025] For example, input a document about the construction scheme of a pressure steel pipe of a hydropower station, the model automatically preliminarily judges that it belongs to the water conservancy engineering-water structure professional-construction stage through the analysis of keywords such as "hydropower station", "pressure steel pipe" and "construction scheme" and related technical description in the document. M32, Classification label verification and manual intervention: If the document itself has a classification label, the system will automatically compare and verify the self-label with the automatic classification result. When the two are inconsistent, the system will give priority to the document self-label and highlight the user classification difference (such as red font) on the system interface, and the user will further confirm and handle it.

[0026] After classification, the system provides the function of manual verification and modification of document type. Users can check the automatic classification results, if there is a misjudgment, they can manually modify the classification information of the document, and save the modification record for subsequent model optimization training. For example, the user finds that a document originally labeled as "energy engineering-thermal power professional-design stage" should actually belong to "energy engineering-thermal power professional-construction stage", the user can directly modify it in the system.

[0027] M4, the large language model processing module, is used to automatically match the review point list in the multi-level review point library based on the classification results of the document classification module, process the document to be reviewed, conduct in-depth review, and output review opinions; the review point list matched by the large language model processing module includes a general review point list and a proprietary review point list.

[0028] M41, the General Review Point Checklist, applies to all engineering documents and covers reviews of completeness, consistency, accuracy, and compliance. Specifically, it includes: M411. Completeness: The document should have a complete structure including a title, table of contents, preface, main body, conclusion, and references, with no parts missing. For example, a construction organization design document must include basic chapters such as project overview, construction deployment, construction schedule, and resource allocation plan. M412. Consistency: The content within the document should be consistent, the terminology should be uniform, the units of measurement should be consistent, and the format should conform to standards. For example, when describing length, both "meter" and "foot" cannot be used simultaneously; in terms of document formatting, the font, font size, and layout of headings at all levels should be consistent. M413. Accuracy: The text in the document is accurate and clear, free of typos and grammatical errors; data citations are accurate, and calculation results are correct. For example, the calculations of various costs in the project budget document should be accurate, and the market price data cited should be reliable; M414. Compliance: Documents must comply with relevant national and industry laws, regulations, standards, and specifications. For example, architectural engineering documents must comply with the "Construction Law," the "Regulations on the Administration of Construction Project Quality," and various architectural design and construction specifications.

[0029] M42, the proprietary review point list, is based on the primary business area, secondary professional classification, and tertiary project stage of the document to be reviewed, making it more targeted. Specifically, it includes: M421, Water Conservancy Engineering - Hydrogeology - Design Stage: In terms of accuracy, review whether the groundwater level data in the hydrogeological survey report is accurate and whether the geological structure analysis is reasonable; in terms of compliance, check whether the survey methods comply with the "Code for Geological Survey of Water Conservancy and Hydropower Engineering". M422, Energy Engineering - Electrical Systems - Construction Phase: In terms of completeness, confirm whether the construction drawings include necessary drawings such as electrical main wiring diagrams, equipment layout diagrams, and cable laying diagrams; in terms of consistency, verify the consistency of equipment parameters in the drawings and equipment list; in terms of compliance, check whether the electrical installation process complies with the "Code for Construction and Acceptance of Electrical Installations".

[0030] M5, the review report generation module, is used to summarize the review comments output by the large language model processing module and generate a review report. This module first needs to automatically match the corresponding review point list based on the document classification results (first-level business domain + second-level professional classification + third-level project stage). For example, if the document is classified as "water conservancy project - construction stage - structural engineering", the system will automatically match the specific review points (such as dam body pouring strength verification) and general review points under this third-level classification, and then combine the large language model to realize hierarchical review. The specific hierarchical review process and report generation logic are as follows: M51, Document Slicing: To enable the large language model to process engineering documents more efficiently, the system uses a sliding window technique to slice the documents. Slicing takes into full account the characteristics of engineering documents, such as text length and density of technical terms, while also considering the input limitations of the large language model to set appropriate window sizes and step sizes.

[0031] Taking common engineering document processing as an example, the window size is set to 2000 characters, and the step size is set to 500 characters. Starting from the beginning of the document, each 2000-character text segment is extracted as an independent slice, and a 500-character overlap is maintained between adjacent slices. This is done to ensure the semantic continuity between slices and avoid semantic misunderstandings caused by slice breaks. For engineering documents containing charts, special treatment is given during slice processing; the charts and their related explanatory text are treated as a whole.

[0032] For example, when processing a construction plan document with a concrete mix proportion chart, the chart itself and the text below the chart explaining the design basis and usage instructions are divided into a single slice to prevent the chart information from being separated from the text explanation, which would affect the completeness of the review. M52. Slice Content Summary and Preliminary Classification: After document slicing, a large language model is used to summarize and preliminarily classify the content of each slice. The large language model deeply analyzes the semantics of the slice text, extracts key information, and generates concise summaries, allowing reviewers to quickly grasp the core content of the slice. Simultaneously, the large language model also refers to the review criteria classification system—namely, completeness, consistency, accuracy, and compliance—to preliminarily classify the slice content and determine the review dimensions that the slice may involve.

[0033] For example, for a slice describing a construction process, the summary generated by the large language model might be "Introduces the bored pile technology and construction process used in the foundation construction of a certain bridge". Through semantic analysis of the content of this slice, it is preliminarily determined that it may involve two review dimensions: accuracy and compliance. The accuracy review mainly focuses on whether the description of the construction process is accurate, while the compliance review focuses on whether the construction process meets the relevant specifications.

[0034] M53. Review Point Matching and In-Depth Review: Based on the initial classification results of the slices, the system will compare the slice content with the previously matched review point list one by one. For review points that match successfully after comparison, in-depth review will be conducted using a large language model. The large language model will strictly follow the specific requirements of the review point to conduct a detailed analysis of the slice content and determine whether it meets the review standards. Taking the review point of concrete strength grade as an example, the large language model will carefully check whether the description of the concrete strength grade in the slice is accurate, check whether the strength grade is consistent with the requirements in the design drawings, and confirm whether it complies with relevant standards and specifications.

[0035] If issues are identified during the review process, the large language model will generate detailed review comments. These comments will include a description of the problem, clearly pointing out the specific issues in the slice; relevant standards and references, explicitly citing the corresponding national laws, regulations, or industry standards; and possible improvement suggestions, providing direction for correcting the problem. For example, if a slice describes a concrete strength grade of C35, but the design drawings require C40, this does not meet the design requirements. The review comments would state, "The concrete strength grade described in the slice is C35, but the design drawings require C40, which does not meet the design requirements. According to the 'Code for Acceptance of Construction Quality of Concrete Structures' GB50204-2015, the concrete strength grade should meet the design specifications." M54. Review Report Generation: Once all slices have been reviewed, the system will summarize the review comments for all slices and generate a comprehensive and detailed review report. The review report uses a structured format and covers several key sections.

[0036] First, there is the basic information of the document, including the document name, the project it belongs to, the document type, and the document classification results, so that reviewers can quickly understand the basic situation of the document; second, there is an overview of the review results, which clearly states whether the review passed or failed, as well as the total number of issues found during the review process. Then comes the detailed review comments, which will be categorized and listed according to the slice order or review dimensions. Each issue will include a specific description, the slice location, the corresponding standard basis, and improvement suggestions. Finally, the review summary and recommendations are presented, summarizing the entire document review process and providing overall suggestions for improvement based on the issues identified.

[0037] Furthermore, the review report supports multiple output formats, including PDF and Word, allowing users to view, print, and share it according to their specific needs. For example, a review report might detail the location of each problematic section in the document, explaining the corresponding review points and problem descriptions, while also referencing specific clauses of relevant standards. At the end of the report, it might provide a summary and recommendations such as, "This document has accuracy and compliance issues in the description of construction processes and some technical parameters; it is recommended to modify the relevant content and re-examine it."

[0038] In summary, the present invention has the following advantages: 1. By automating the entire process (batch access from multiple channels, automatic BERT classification, and automatic review of large models), manual word-by-word verification is replaced, shortening the review cycle and solving the problem of time-consuming manual review; 2. By using a four-level classification review point library to cover multiple fields and all stages, combined with large model semantic understanding, it can accurately identify detailed problems (such as parameter conflicts), avoid human omissions, and solve the problem of insufficient flexibility of rule engines; 3. Automatic judgment is made by using a general review point list and a proprietary review point list to eliminate differences in the understanding of the specifications by different personnel and ensure consistent review results; 4. By supporting multiple document types and multiple access channels, it allows for the addition of new business areas, fine-tuning of models to adapt to new industry standards, and adaptation to dynamic needs.

[0039] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent document review system for the engineering industry based on a large language model, characterized in that, include: A multi-level review point library is used to provide the review standards required for reviewing engineering documents; The document access module is used to access documents awaiting review from the engineering industry. The document classification module is used to automatically classify the documents to be reviewed that are accessed by the document access module, and is configured with a manual verification interface to verify and modify the automatic classification results; The large language model processing module is used to automatically match the list of review points in the multi-level review point library based on the classification results of the document classification module, process the document to be reviewed, conduct in-depth review, and output review opinions. The review report generation module is used to summarize the review comments output by the large language model processing module and generate a review report.

2. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The multi-level review point database has a four-level classification structure, specifically including: The primary business areas are divided into water conservancy engineering, energy engineering, urban construction, industrial engineering, and digital engineering. The secondary professional classification is further subdivided based on each primary business area. Among them, water conservancy engineering corresponds to hydrogeology, hydraulic structure, water conservancy machinery, and water conservancy electrical engineering; energy engineering corresponds to energy power, electrical system, and new energy technology; urban construction corresponds to building structure, water supply and drainage, heating, ventilation and air conditioning, and road and bridge; industrial engineering corresponds to mechanical manufacturing, automation control, and industrial building; and digital engineering corresponds to software engineering, data analysis, and Internet of Things technology. The three-tier project phase is divided into the design phase, construction phase, operation and maintenance phase, and acceptance phase. The four-level review dimensions are completeness, consistency, accuracy, and compliance.

3. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The list of review points matched by the large language model processing module includes a general review point list and a proprietary review point list. The general review checklist applies to all engineering documents and covers reviews of completeness, consistency, accuracy, and compliance. The proprietary review point list is set based on the primary business area, secondary professional category, and tertiary project stage corresponding to the document to be reviewed.

4. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The document access module supports the following document types for review: text documents, drawing documents, data table documents, and scanned documents. The text documents include Word documents and PDF documents; The drawing documents include CAD drawings and BIM model-related files; The data table documents include Excel documents; The scanned document type is an image format document that has been converted into editable text using OCR technology.

5. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The document access module supports access channels including local upload channels, database access channels, and third-party platform access channels; The local upload channel supports batch uploads and features upload progress display and error message display. The database access channel connects with the enterprise's internal database by configuring connection parameters and query statements to obtain and synchronize documents to be reviewed. The third-party platform access channel is integrated with the engineering management platform and cloud storage platform through API interfaces.

6. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The automatic classification of the document classification module is based on a text classification algorithm fine-tuned with engineering industry expertise and corpus. The text classification algorithm is based on the BERT model and outputs the first-level business domain, second-level professional category and third-level project stage of the document by analyzing the keywords, sentence structure and semantic relationships of the document to be reviewed. If the document to be reviewed has its own category tags, the document classification module will also compare and verify the built-in tags with the automatic classification results. If the two are inconsistent, the built-in tags shall prevail and the classification difference shall be indicated. The manual verification interface allows users to modify the classification results, and the modification records are used for subsequent model optimization and training.

7. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The large language model processing module processes the document to be reviewed by slicing, using sliding window technology, setting the window size and step size, and having overlapping parts between adjacent slices to ensure the semantic coherence of the context.

8. The intelligent document review system for the engineering industry based on a large language model according to claim 7, characterized in that, For documents to be reviewed that contain charts, the slicing process extracts the charts and their related explanatory text as a whole.

9. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, During the in-depth review process of the large language model processing module, for the successfully matched review points, the conformity between the content of the document to be reviewed and the review point standards is analyzed. If there are non-conformities, the generated review opinions include a problem description, relevant standard basis, and improvement suggestions.

10. The intelligent document review system for the engineering industry based on a large language model according to claim 1, characterized in that, The review report generation module generates a structured report, which includes basic document information, an overview of the review results, detailed review comments, and a review summary and recommendations. The review report supports PDF and Word format output. It also supports adding new business areas and review points, and adapts to new review standards by fine-tuning the large language model.

Images