Construction technology document intelligent review method based on knowledge-guided retrieval

Abstract

The invention relates to a construction technology document intelligent review method based on knowledge-guided retrieval, and the method comprises the following steps: obtaining query sentences in a construction technology document, inputting the query sentences into a BERT-based encoder, and screening a preliminary retrieval set by the BERT-based encoder based on global similarity; the method comprises the following steps: encoding a query sentence by using field enhanced BERT, calculating term importance, term rare and consistency indexes of quantitatively extracted key information marked in each document block of a knowledge base, calculating a total key information score of each document block, and screening a preliminary retrieval set based on the total key information score; and obtaining a trained and enhanced screening model preliminary retrieval set, and inputting the preliminary retrieval set into the trained and enhanced screening model to obtain an examination result. Compared with the prior art, the method has the advantages that the problems that in an existing construction technology document review technology, knowledge in the building field is insufficient, and the technology violation retrieval precision is limited are solved.

Description

Technical Field

[0001] This invention relates to the intersection of building information modeling and artificial intelligence, and in particular to an intelligent review method for construction technical documents based on knowledge-guided retrieval. Background Technology

[0002] Construction technical document review is a crucial step in building engineering quality control, primarily comprising three stages: error identification, information retrieval, and revision generation. Traditional manual review methods rely on the reviewer's professional knowledge and experience, involving manual comparison and analysis of building regulations. However, this method suffers from high subjectivity, is time-consuming, and prone to errors, making it unsuitable for large-scale document review needs.

[0003] Natural Language Processing (NLP) technology offers a new solution for automated review of construction documents. Traditional NLP methods, such as text classification and named entity recognition models, typically only generate predefined symbolic outputs, limiting their capabilities in architectural document review tasks. While Large Language Models (LLMs) possess strong semantic reasoning capabilities, they are prone to illusion problems in complex reasoning due to a lack of architectural domain expertise.

[0004] Retrieval Enhancement Generation (RAG) methods enhance the domain capabilities of LLMs by retrieving relevant entries from a knowledge-based corpus. Existing RAG technologies suffer from two main drawbacks: first, they are built upon a general corpus and lack specialized knowledge in the construction field; second, retrieval based on overall semantic similarity is primarily suitable for identifying topic relevance and concept matching, but struggles to identify fine-grained technical violations requiring precise comparison, such as deviations in construction parameters, errors in process flow, and inconsistencies in technical standards.

[0005] Therefore, there is currently no intelligent system specifically designed for reviewing construction technical documents. Existing general document processing systems cannot meet the specialized review needs of the construction industry, and there is an urgent need to develop specialized intelligent review methods for the construction industry. Summary of the Invention

[0006] The purpose of this invention is to address the problems of insufficient knowledge in the construction field and limited accuracy in retrieval of technical violations in existing construction technical document review technologies, and to provide an intelligent review method for construction technical documents based on knowledge-guided retrieval to significantly improve the accuracy and efficiency of the review.

[0007] The objective of this invention can be achieved through the following technical solutions: A knowledge-guided retrieval-based intelligent review method for construction technical documents, comprising the following steps: Retrieves query sentences from construction technical documents. The query sentences are input into a BERT-based encoder, which then filters the initial retrieval set based on global similarity. ; The query sentence is encoded using domain-enhanced BERT to generate its contextual representation. The trained bidirectional LSTM model is used to represent the context. The process is performed to obtain weighted key information items. ,in Context representation Key information at the i-th importance level, express The weights are calculated, and the term importance, term rarity, and consistency index of the extracted key information tags in each document block of the knowledge base are quantified. Based on the term importance, term rarity, and consistency index, the weighted key information items and the matching degree between each document block are calculated, and the total key information score of each document block is calculated. Based on the total key information score, the preliminary search set is screened. ; Obtain the enhanced screening model after training is completed; Preliminary search set and preliminary search set Input the enhanced screening model after training to obtain the review results.

[0008] Furthermore, the importance of the terminology is as follows: in, This indicates the key information extracted by quantification. The mark, For the j-th document block, This is a document length adjustment factor, where f represents key information. The frequency of the marker appearing in the j-th document block, where i is the importance level of the i-th block.

[0009] Furthermore, the document length adjustment factor is: in, This indicates the average length of the document. For the j-th document block The length.

[0010] Furthermore, the rarity of the term is: in, Indicates the total number of segments. This indicates that it contains key information. The number of segments marked.

[0011] Furthermore, the consistency index is: Specifically, a sliding window is generated on the j-th document block. ,in It is the window length. It is an indicator function, where t represents the t-th marker.

[0012] Furthermore, the degree of matching is: ( in, Let τ represent the τ-th tag. M represents the total number of tags.

[0013] Furthermore, the total key information score is: in, For querying sentences.

[0014] Furthermore, the global similarity is cosine similarity.

[0015] Furthermore, the preliminary search set and preliminary search set The specific steps to obtain the review results by inputting the trained and enhanced screening model are as follows: The enhanced screening model performs knowledge consistency matching and redundant knowledge removal on the candidate knowledge block set, and outputs a refined Top-N result set after removing redundancy as the core basis for error analysis and revision generation. Based on the core basis of error analysis and revision generation, the candidate knowledge block set is corrected to obtain the review result.

[0016] Furthermore, the candidate knowledge block set constitutes the preliminary retrieval set. and preliminary search set The intersection of.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention enhances the information retrieval capabilities of a large language model by leveraging key information extracted from queries, enabling it to incorporate both holistic semantic information and core fine-grained element information during the retrieval process. This method overcomes the limitation of traditional information retrieval systems that rely solely on sentence similarity, achieving deep semantic understanding and refinement. Attached Figure Description

[0018] Figure 1 This is a screenshot of the interface of an automated document review system. Figure 2 Graph of the data preparation process; Figure 3 Here is the overall workflow diagram for the construction technical document review system; Figure 4A comprehensive workflow diagram for a knowledge-guided generative retrieval framework; Figure 5 A hierarchical diagram for architectural queries; Figure 6 Impact diagrams to guide key information; Figure 7 This is a graph showing the degree of attention paid to fine-grained information in the final search results; Figure 8 This is a diagram illustrating the entire automated review process for Case 1. Figure 9 This is a diagram illustrating the entire automated review process for Case 2. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0020] This invention provides an intelligent review method and system for construction technical documents based on knowledge-guided retrieval. By constructing a high-quality architectural knowledge base and using a generative knowledge-guided retrieval framework, it aims to solve the problems of insufficient knowledge in the architectural field and limited accuracy in retrieval of technical violations in existing construction technical document review technologies, thereby significantly improving the accuracy and efficiency of the review process.

[0021] To achieve the above effects, this invention provides an intelligent review method for construction technical documents based on generative knowledge-guided retrieval. This method obtains more accurate knowledge retrieval prompts by integrating semantic similarity and key query element matching, and uses these prompts to guide a large language model in conducting compliance review and correction of the document to be reviewed, ultimately presenting a revised overall document. The proposed method specifically includes the following steps: a. Data collection and preprocessing: Collect construction technical documents and relevant standards and specifications, and build a review standards knowledge base; b. Review Query Element Generation: Receives construction technical document input and generates review questions based on the context; c. Generative knowledge-guided retrieval: Based on the constructed knowledge base and the question to be reviewed, the semantics of the whole sentence and key element information are integrated to achieve more accurate knowledge prompts.

[0022] d. Error Analysis and Revision Generation: Based on the retrieved knowledge tips, the document to be reviewed is analyzed to identify potential errors in the construction technical documents and generate targeted revision suggestions, ultimately achieving automated correction.

[0023] The process of receiving construction technical document input and automatically generating review questions based on contextual information includes two components: (1) The first stage involves extracting the core events and key technical features of the document; (2) The second stage is to generate fine-grained compliance review issues based on core content.

[0024] The steps for knowledge retrieval using a generative knowledge-guided retrieval framework include four components: (1) The global similarity retrieval branch calculates the semantic similarity between the query and the knowledge block to obtain the preliminary retrieval set. ; (2) The retrieval branch guided by key element knowledge identifies key element information in the query and generates importance weights for each element based on contextual information, resulting in a key information subgraph of the query. Using this as guiding knowledge, refined core knowledge block matching is achieved, yielding a preliminary retrieval set. ; (3) Use the query, the corresponding key information subgraph, and two preliminary retrieval sets to construct training samples, enhance the compliance knowledge screening capability of the large language model, and realize the injection of fine-grained retrieval constraint knowledge obtained in component (2) into the model; (4) The enhanced screening model is used to analyze and rearrange the candidate knowledge block set, and finally the refined Top-N search results are output as the basis for error analysis and revision.

[0025] In extracting the core events of the document, key components such as core events, key technologies, and operational processes are extracted based on the initial prompts and organized into a concise summary. This process clarifies the main focus of the review and lays the foundation for generating more targeted questions in the next stage.

[0026] In generating fine-grained compliance review questions based on core content, a second hint guides the model to generate fine-grained, multi-perspective questions based on the summarized core content. These questions focus on potential constraints, procedural risks, and compliance issues. Queries do not rely on fixed templates but are dynamically generated based on document content, enhancing the system's ability to identify potential defects and support targeted searches.

[0027] The global similarity retrieval branch primarily processes the input query sentence. First, the query is encoded into a sentence-level embedding representation using an embedding model encoder. Second, to measure the overall similarity between the query and each text block in the knowledge base, the cosine similarity between the query's embedding representation and the embedding representation of each text block is calculated, forming a sentence-level score vector. Finally, based on the sentence-level score vector, all knowledge base blocks are ranked, and the top N blocks are retrieved to form a preliminary retrieval set. .

[0028] The key element knowledge-guided retrieval branch mainly includes three sub-modules: key information extraction, indicator calculation, and key information score calculation.

[0029] In enhancing the model's compliance knowledge screening capabilities, the fine-grained retrieval knowledge mentioned earlier is injected into the large language model, strengthening the model's emphasis on fine-grained information, thereby enabling it to acquire more accurate retrieval knowledge.

[0030] In the knowledge reordering and refinement process, an enhanced filtering model is used to perform knowledge consistency matching and redundant knowledge removal on the retrieved candidate knowledge block set. First, the relevance and information redundancy of each knowledge block to the query question are evaluated; second, they are prioritized based on the evaluation scores; finally, a refined Top-N result set after removing redundancy is output. This result will serve as the core basis for subsequent error analysis and revision generation.

[0031] The steps of generating a compliance review report based on the search results through error analysis and revision include four components: (1) Locate potential violation areas in the document using the sliding window method; (2) Compare the technical parameters of the search knowledge reference with those of the original text; (3) Based on the detected errors and relevant knowledge references, generate revision suggestions that include references to regulatory clauses; (4) Output an audit report with error type annotations.

[0032] This invention provides an intelligent review method and system for construction technical documents based on knowledge-guided retrieval. Through innovative technical solutions, it effectively addresses the problems of low accuracy, poor efficiency, and error-proneness in traditional review methods. Experiments demonstrate that this system significantly improves both review quality and efficiency, possessing significant theoretical value and broad application prospects.

[0033] This invention provides an intelligent review method and system for construction technical documents based on knowledge-guided retrieval. By constructing a high-quality building knowledge base and using a generative knowledge-guided retrieval framework, it aims to solve the problems of insufficient knowledge in the building field and limited accuracy in retrieving technical violations in existing construction technical document review technologies, thereby significantly improving the accuracy and efficiency of the review.

[0034] To achieve the above effects, this invention provides an intelligent review method for construction technical documents based on generative knowledge-guided retrieval. This method obtains more accurate knowledge retrieval prompts by integrating semantic similarity and key query element matching, and uses these prompts to guide a large language model in conducting compliance review and correction of the document to be reviewed, ultimately presenting a revised overall document. The proposed method specifically includes the following steps: a. Data collection and preprocessing: Collect construction technical documents and relevant standards and specifications, and build a review standards knowledge base; b. Review Query Element Generation: Receives construction technical document input and generates review questions based on the context; c. Generative knowledge-guided retrieval: Based on the constructed knowledge base and the question to be reviewed, the semantics of the whole sentence and key element information are integrated to achieve more accurate knowledge prompts.

[0035] d. Error Analysis and Revision Generation: Based on the retrieved knowledge tips, the document to be reviewed is analyzed to identify potential errors in the construction technical documents and generate targeted revision suggestions, ultimately achieving automated correction.

[0036] The process of receiving construction technical document input and automatically generating review questions based on contextual information includes two components: (1) The first stage involves extracting the core events and key technical features of the document; (2) The second stage is to generate fine-grained compliance review issues based on core content.

[0037] The steps for knowledge retrieval using a generative knowledge-guided retrieval framework include four components: (1) The global similarity retrieval branch calculates the semantic similarity between the query and the knowledge block to obtain the preliminary retrieval set. ; (2) The retrieval branch guided by key element knowledge identifies key element information in the query and generates importance weights for each element based on contextual information, resulting in a key information subgraph of the query. Using this as guiding knowledge, refined core knowledge block matching is achieved, yielding a preliminary retrieval set. ; (3) Use the query, the corresponding key information subgraph, and two preliminary retrieval sets to construct training samples, enhance the compliance knowledge screening capability of the large language model, and realize the injection of fine-grained retrieval constraint knowledge obtained in component (2) into the model; (4) The enhanced screening model is used to analyze and rearrange the candidate knowledge block set, and finally the refined Top-N search results are output as the basis for error analysis and revision.

[0038] The steps of generating a compliance review report based on the search results through error analysis and revision include four components: (1) Locate potential violation areas in the document using the sliding window method; (2) Compare the technical parameters of the search knowledge reference with those of the original text; (3) Based on the detected errors and relevant knowledge references, generate revision suggestions that include references to regulatory clauses; (4) Output an audit report with error type annotations.

[0039] This invention provides an intelligent review method and system for construction technical documents based on knowledge-guided retrieval. Through innovative technical solutions, it effectively addresses the problems of low accuracy, poor efficiency, and error-proneness in traditional review methods. Experiments demonstrate that this system significantly improves both review quality and efficiency, possessing significant theoretical value and broad application prospects.

[0040] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for intelligent review of construction technical documents based on knowledge-guided retrieval, characterized in that, The method includes the following steps: Retrieves query sentences from construction technical documents. The query sentences are input into a BERT-based encoder, which then filters the initial retrieval set based on global similarity. ; The query sentence is encoded using domain-enhanced BERT to generate its contextual representation. The trained bidirectional LSTM model is used to represent the context. The process is performed to obtain weighted key information items. ,in Context representation Key information at the i-th importance level, express The weights are calculated, and the term importance, term rarity, and consistency index of the extracted key information tags in each document block of the knowledge base are quantified. Based on the term importance, term rarity, and consistency index, the weighted key information items and the matching degree between each document block are calculated, and the total key information score of each document block is calculated. Based on the total key information score, the preliminary search set is screened. ; Obtain the enhanced screening model after training is completed; Preliminary search set and preliminary search set Input the enhanced screening model after training to obtain the review results.

2. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 1, characterized in that, The importance of the terminology is as follows: in, This indicates the key information extracted by quantification. The mark, For the j-th document block, This is a document length adjustment factor, where f represents key information. The frequency of the marker appearing in the j-th document block, where i is the importance level of the i-th block.

3. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 2, characterized in that, The document length adjustment factor is: in, This indicates the average length of the document. For the j-th document block The length.

4. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 3, characterized in that, Terminology rarity is: in, Indicates the total number of segments. This indicates that it contains key information. The number of segments marked.

5. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 4, characterized in that, The consistency index is: Specifically, a sliding window is generated on the j-th document block. ,in It is the window length. It is an indicator function, where t represents the t-th marker.

6. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 5, characterized in that, The degree of matching is: ( in, Let τ represent the τ-th tag. M represents the total number of tags.

7. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 6, characterized in that, The total key information score is: in, For querying sentences.

8. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 1, characterized in that, The global similarity is cosine similarity.

9. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 1, characterized in that, Preliminary search set and preliminary search set The specific steps to obtain the review results by inputting the trained and enhanced screening model are as follows: The enhanced screening model performs knowledge consistency matching and redundant knowledge removal on the candidate knowledge block set, and outputs a refined Top-N result set after removing redundancy as the core basis for error analysis and revision generation. Based on the core basis of error analysis and revision generation, the candidate knowledge block set is corrected to obtain the review result.

10. The intelligent review method for construction technical documents based on knowledge-guided retrieval according to claim 9, characterized in that, The candidate knowledge block set is the preliminary retrieval set. and preliminary search set The intersection of.

Images