An information accurate comparison method and system based on project file intelligent checking

Abstract

The present application relates to the field of data processing, and specifically relates to a kind of information accurate comparison method and system based on project file intelligent check.The present application includes: obtaining design file and specification file;Extract key indicators, build associated indicator list;Design file is divided into paragraph unit, and extract keyword set;For any paragraph unit, match its keyword set with each key indicator, determine to be checked indicator, and locate candidate paragraph;For the candidate paragraph corresponding to any to-be-checked indicator, obtain semantic level value, and build level tree;Calculate the structural similarity between the level trees of different candidate paragraphs, and screen paragraph pair;Match the keywords in any structural similar paragraph pair, and screen candidate keyword;Determine the meaning correlation between the candidate keyword and to-be-checked indicator;Generate check result.The present application can improve the accuracy, intelligence and automation level of project file check.

Description

Technical Field

[0001] This invention relates to the field of data processing, and specifically to a method and system for accurate information comparison based on intelligent verification of project documents. Background Technology

[0002] Project document verification refers to the systematic inspection of the compliance, accuracy, and completeness of design drawings, technical documents, construction plans, and other related documents during the engineering construction, design review, and final acceptance stages, in accordance with relevant technical specifications, national standards, and industry regulations. This process aims to ensure that the project design meets mandatory regulatory requirements and avoid engineering quality problems or safety hazards caused by design flaws.

[0003] Currently, automated compliance verification of engineering design drawings primarily relies on rule-based matching methods that match preset keywords with numerical values. These methods construct keyword libraries and rule templates to search and compare text annotations in the drawings. For example, they search for preset keywords such as "energy efficiency," "Level 1," and "COP" in the equipment list. When a keyword is matched, the corresponding numerical comparison rule is triggered. This method, to a certain extent, automates the processing of structured data and is widely used in the field of architectural design compliance review.

[0004] However, existing technologies mainly face the following technical problems: due to the lack of deep understanding of the semantics of the specification text, they cannot identify the implicit logical relationship between the specific performance parameters in the design drawings and the qualitative requirements of the specification, nor can they call on external knowledge bases to perform parameter conversion and threshold comparison, resulting in a large number of compliance defects that should have been discovered being missed, making it difficult to meet the actual needs of intelligent review of engineering design drawings. Summary of the Invention

[0005] This invention provides a method and system for accurate information comparison based on intelligent verification of project documents, in order to solve existing problems.

[0006] The present invention provides a method for accurate information comparison based on intelligent verification of project documents, which adopts the following technical solution: One embodiment of the present invention provides a method for accurate information comparison based on intelligent verification of project documents, the method comprising the following steps: Obtain the design documents to be verified and the specification documents used as the basis for verification; Key indicators are extracted from the standard documents, and a list of related indicators for each key indicator is constructed based on the co-occurrence relationship of each key indicator in the standard clauses. Divide the design file into paragraph units and extract the keyword set for each paragraph unit; For any paragraph unit, its keyword set is matched with each key indicator, the indicators to be verified in the design document are determined from the key indicators, and the candidate paragraphs corresponding to the indicators to be verified are located. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph, and construct a hierarchical tree to represent the semantic hierarchical structure of the paragraph based on the semantic level value; Calculate the structural similarity between the hierarchical trees of different candidate paragraphs, and filter out paragraph pairs with similar structures; Match keywords in any pair of paragraphs with similar structures, and filter out candidate keywords that correspond to the indicators to be verified based on the matching results; The candidate keywords are compared and verified with the preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified. The verification information is extracted from the paragraphs containing candidate keywords whose semantic relevance has been verified, and the verification information is compared with the verification requirements of the indicators to be verified in the specification document to generate verification results.

[0007] Optionally, key indicators are extracted from the specification documents, and a list of associated indicators for each key indicator is constructed based on the co-occurrence relationships of each key indicator in the specification clauses, specifically including: The standard document is divided into clauses, with each clause treated as an independent unit; Perform natural language processing on each independent unit to identify key metrics; Calculate the frequency of each key indicator in all independent units, and the frequency of any two key indicators appearing together in the same independent unit; For any two key indicators, the ratio of the frequency of their co-occurrence to the total frequency of at least one of them is determined as the co-occurrence strength between the two key indicators. For any key indicator, other key indicators whose co-occurrence intensity exceeds a preset co-occurrence threshold are used as associated indicators of that key indicator, and a list of associated indicators of that key indicator is constructed.

[0008] Optionally, for any paragraph unit, its keyword set is matched with each key indicator to determine the indicators to be verified in the design document from the key indicators, and the candidate paragraphs corresponding to the indicators to be verified are located, specifically including: For any given paragraph unit, match its keyword set with each key metric; If the keyword set includes a key indicator, and the paragraph unit includes verification information for that key indicator, then the key indicator is determined to be a non-verification indicator. If the keyword set does not include a key indicator, or if the keyword set includes a key indicator but the paragraph unit does not include verification information for that key indicator, then the key indicator is identified as an indicator to be verified. For any indicator to be verified, each associated indicator in its associated indicator list is matched with the keyword set of each paragraph unit. The paragraph unit containing the successfully matched associated indicator is determined as the candidate paragraph for that indicator to be verified.

[0009] Optionally, for any candidate paragraph corresponding to an indicator to be verified, the semantic level values ​​of each keyword within that paragraph are obtained, and a hierarchical tree representing the semantic hierarchy of that paragraph is constructed based on these semantic level values, specifically including: Obtain a pre-built vocabulary library, which includes at least three words and the semantic level value corresponding to each word. The semantic level value is used to quantify the semantic generalization degree of the words. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph from the thesaurus; Based on the semantic level values ​​of each keyword, the keyword set of the candidate paragraph is clustered from bottom to top to construct the hierarchical tree of the candidate paragraph.

[0010] Optionally, a pre-built vocabulary can be obtained, specifically including: The text in standard documents, qualified design documents, and related field documents is cleaned, segmented into words, and divided into sentences to obtain a corpus; Each corpus in the corpus is processed using a word embedding model to obtain word vectors for each corpus; Obtain at least three clustering density parameters. For each clustering density parameter, perform density clustering on all word vectors based on that clustering density parameter to obtain the clusters under that clustering density parameter. For each cluster under each cluster density parameter, calculate the average of all word vectors in the cluster to obtain the cluster center vector of that cluster; For each corpus under each cluster density parameter, calculate the distance between the word vector of the corpus and the cluster center vectors of all clusters under that density, convert the distance into similarity, and normalize the similarity to obtain the membership degree of the corpus to each cluster. For each corpus under each density parameter, record its maximum membership degree and second maximum membership degree, and calculate the difference between the two; For each corpus, the difference sequence under different density parameters is obtained to form the difference change curve of the corpus; The difference change curve for each corpus is fitted, and the slope of the trend representing the decrease of the difference as the density parameter increases is calculated. Sort the absolute values ​​of the trend slopes of all corpora to determine the maximum and minimum trend slopes; For each corpus, the difference between the absolute value of its trend slope and the minimum value of its trend slope is divided by the difference between the maximum value of its trend slope and the minimum value of its trend slope to calculate the normalized semantic level value. Each corpus and its corresponding semantic level value are stored to construct a vocabulary database.

[0011] Optionally, the structural similarity between the hierarchical trees of different candidate paragraphs is calculated, and paragraph pairs with similar structures are selected, specifically including: For any two candidate paragraphs, calculate the average semantic hierarchy value of their hierarchical trees; Compare the average semantic level values ​​of the two, and identify the candidate paragraph with the lower average semantic level value as the summary paragraph, and identify the candidate paragraph with the higher average semantic level value as the detailed paragraph. The hierarchical tree of the summary paragraph is compared with the hierarchical tree of the detailed paragraph layer by layer, and the distribution similarity of the semantic hierarchy values ​​corresponding to each node is calculated. If the number of levels with distribution similarity exceeding the preset structural similarity threshold accounts for more than the preset proportion threshold of the total number of levels in the hierarchical tree, then the two candidate paragraphs are determined to be a pair of paragraphs with similar structure.

[0012] Optionally, keywords in any pair of paragraphs with similar structures are matched, and candidate keywords corresponding to the indicators to be verified are selected based on the matching results, specifically including: For any pair of paragraphs, obtain the keyword set of the summary paragraph and the keyword set of the detailed paragraph in that pair; A bipartite graph is constructed using the keyword set of the summary paragraph and the keyword set of the detailed paragraph as two vertex sets. The weight of each edge is calculated based on the semantic similarity and semantic level value difference between the two keywords. The KM algorithm is used to solve the maximum weight matching of the bipartite graph, and the matching pairs and the matching degree corresponding to each matching pair are obtained. Based on the matching degree from smallest to largest, a preset number of matching pairs are selected, and the keywords belonging to the detailed paragraphs in each selected matching pair are used as candidate keywords corresponding to the indicators to be verified.

[0013] Optionally, the candidate keywords are compared and verified with a preset knowledge base to determine the semantic relevance between the candidate keywords and the indicators to be verified, specifically including: Obtain a preset knowledge base, which includes standard names, synonyms, abbreviations, and definitions of domain terms; For each candidate keyword, retrieve the corresponding standard term from the preset knowledge base; If the retrieved standard terminology matches the name of the indicator to be verified, it is determined that there is a semantic relationship between the candidate keyword and the indicator to be verified. If the standard term corresponding to the candidate keyword is not found in the preset knowledge base, the semantic similarity between the candidate keyword and the name of the indicator to be verified is calculated. If the semantic similarity exceeds the preset semantic similarity threshold, it is determined that there is a semantic association between the candidate keyword and the indicator to be verified.

[0014] Optionally, verification information is extracted from the paragraphs containing candidate keywords whose semantic relevance has been verified, and the verification information is compared with the verification requirements of the indicators to be verified in the specification document to generate verification results, specifically including: For each candidate keyword that passes the semantic relevance verification, locate the sentence containing the candidate keyword and at least one sentence before and after it, as the scope of information extraction; Extract numerical values ​​and unit combinations from the information extraction range to serve as the verification information corresponding to the candidate keyword; The extracted verification information was converted to units according to the unit requirements of the indicators to be verified in the standard documents; Obtain the numerical requirements and modifiers of the indicator to be verified in the specification document, and compare the converted verification information with the numerical requirements and modifiers; Based on the comparison results, the verification result of the indicator to be verified is determined, including whether the verification is passed, failed, or the information is missing.

[0015] This invention proposes an information accuracy comparison system based on intelligent verification of project documents, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of an information accuracy comparison method based on intelligent verification of project documents.

[0016] The beneficial effects of the technical solution of the present invention are: In this embodiment of the invention, by constructing a list of associated indicators for key indicators and building a paragraph hierarchy tree based on semantic level values, structured semantic matching between design documents and specification documents is achieved, enabling accurate location of candidate paragraphs corresponding to the indicators to be verified. By calculating the similarity of the hierarchy tree structure and filtering structurally similar paragraph pairs, and combining this with a bipartite graph maximum weight matching algorithm to filter candidate keywords, the verification difficulties caused by scattered information and inconsistent expressions in design documents are effectively solved. By comparing and verifying candidate keywords with a preset knowledge base, the implicit logical relationships between performance parameters and specification requirements are accurately identified, avoiding omissions and misjudgments caused by traditional keyword matching. By extracting information from the verified paragraphs and comparing them with specification requirements, accurate verification results are generated. This invention significantly improves the intelligence and accuracy of project document verification, effectively ensuring design compliance and reducing manual review costs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for accurate information comparison based on intelligent verification of project documents, provided in one embodiment of the present invention; Figure 2 This is a structural diagram of an information accuracy comparison system based on intelligent verification of project documents, provided as an embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an information accuracy comparison method based on intelligent verification of project documents proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details a specific scheme for an information accuracy comparison method based on intelligent verification of project documents provided by this invention.

[0022] This invention provides a method and system for accurate information comparison based on intelligent verification of project documents. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of an embodiment of the present invention providing a method for accurate information comparison based on intelligent verification of project documents. The method includes the following steps: S101. Obtain the design documents to be verified and the specification documents used as the basis for verification.

[0023] For example, in this embodiment, obtaining the design documents to be verified and the specification documents used as the basis for verification specifically includes: identifying the design documents and specification documents as the verification object and the verification basis, respectively; parsing the design documents to construct structured paragraph units; and parsing the specification documents to extract indicator names, numerical requirements, modifiers, and relationships to form a structured verification task list.

[0024] The specific implementation method can be as follows: read the engineering design drawings to be verified and their supporting design specification documents from the project document management system as the design documents to be verified; read the national, industry or local standards and specifications corresponding to the engineering type to which the design documents belong from the standard and specification database as the standard documents on which the verification is based.

[0025] For example, taking the energy efficiency verification of the HVAC system of a large public building as an example, in the pre-processing stage, it is necessary to first clarify the positioning of two types of documents: design documents are the technical documents of the project to be verified, which contain specific design schemes, parameters and implementation details; and specification documents are the standard basis for verification, which contain various technical clauses and key indicators.

[0026] For design documents, this embodiment reads HVAC system construction drawings, design specifications, equipment lists, etc., from the project document management system. By parsing their chapter structure and text content, structured paragraph units are constructed. For example, the design specifications are divided into paragraph units such as "General Principles," "Air Conditioning Cold and Heat Source System," "Air Conditioning Water System," and "Air Conditioning Air System," and the equipment list is divided into item units such as "Chiller Units," "Water Pumps," and "Air Conditioning Units" according to equipment type. Each unit corresponds to an independent paragraph, which facilitates subsequent retrieval and matching.

[0027] For regulatory documents, this embodiment retrieves national standards such as the "Design Standard for Energy Conservation of Public Buildings" and the "Design Code for Heating, Ventilation and Air Conditioning of Civil Buildings" from the standard and specification database. By parsing the clauses of the regulatory documents, key indicator names (such as "COP," "energy efficiency rating," and "cooling load"), numerical requirements (such as "not less than 6.0" and "≤0.5"), modifiers (such as "Level 1" and "nominal operating conditions"), and the relationships between indicators (e.g., "COP" and "energy efficiency rating" appearing simultaneously in the same clause constitute a co-occurrence relationship). This information is integrated to form a structured verification task list, clearly defining the indicator name, regulatory requirements, and related indicators for each verification task.

[0028] It should be noted that the design documents mentioned in this embodiment include, but are not limited to, various project technical documents such as engineering design drawings, design specifications, calculation sheets, equipment lists, and construction plans. Their file formats can be Word documents, PDF documents, CAD drawings, structured documents, or any other electronic document format. The specification documents include, but are not limited to, various technical specification documents such as national standards, industry standards, local standards, and enterprise standards. Their file formats can also be various electronic document formats. Those skilled in the art can use appropriate parsing tools to process files of different formats according to actual needs; no specific limitations are made here.

[0029] S102. Extract key indicators from the standard documents, and construct a list of related indicators for each key indicator based on the co-occurrence relationship of each key indicator in the standard clauses.

[0030] In this embodiment, key indicators are extracted from the specification documents, and based on the co-occurrence relationship of each key indicator in the specification clauses, a list of associated indicators for each key indicator is constructed, specifically including: The standard document is divided into clauses, with each clause treated as an independent unit; Perform natural language processing on each independent unit to identify key metrics; Calculate the frequency of each key indicator in all independent units, and the frequency of any two key indicators appearing together in the same independent unit; For any two key indicators, the ratio of the frequency of their co-occurrence to the total frequency of at least one of them is determined as the co-occurrence strength between the two key indicators. For any key indicator, other key indicators whose co-occurrence intensity exceeds a preset co-occurrence threshold are used as associated indicators of that key indicator, and a list of associated indicators of that key indicator is constructed.

[0031] For example, this embodiment is mainly used to extract key indicators that need to be verified from specification documents, and to construct a list of related indicators for each indicator by analyzing the co-occurrence relationship of the indicators in the clauses, providing a basis for subsequently locating relevant paragraphs in the design documents. The specific implementation process is as follows: First, the acquired specification documents are structured. Since specification documents are typically organized in clause form, with each clause containing independent technical requirements, the document needs to be segmented by clause, treating each clause as an independent processing unit. For example, a building electrical design specification can be divided into independent clause units such as "General Provisions," "Power Supply and Distribution System," "Lighting System," and "Lightning Protection and Grounding." Each clause unit is assigned a unique identifier, and its original chapter information is retained for later traceability. In practice, this can be achieved using methods such as document parsing libraries, regular expression matching, or directed acyclic graph algorithms.

[0032] Each individual clause unit undergoes natural language processing to identify key metrics. The specific implementation is as follows: First, the clause text is preprocessed, including word segmentation, part-of-speech tagging, named entity recognition, etc. The processing methods used in the preprocessing are all existing technologies, and any technology can be selected to achieve the desired effect in the preprocessing. We will not go into too much detail here.

[0033] Then, based on a pre-defined domain dictionary and grammatical rules, noun phrases and technical terms are extracted as candidate key indicators. For example, from the clause "the continuous power supply time of emergency lighting fixtures shall not be less than 90 minutes", the key indicator "continuous power supply time" can be identified; from the clause "the protection level of the distribution box shall not be lower than IP54", the key indicator "protection level" can be identified.

[0034] For each identified key indicator, additional information must be recorded, including: numerical requirements (e.g., "90", "54"), units (e.g., "minutes", "IP"), and modifiers (e.g., "should not be less than" or "should not be lower than"). This information will be used for subsequent numerical comparison and verification.

[0035] After identifying the key indicators for all clauses, the frequency of each key indicator appearing in all clause units is calculated. Simultaneously, the frequency of any two key indicators co-occurring in the same clause unit is also calculated. For any two key indicators, the co-occurrence strength is calculated based on the above statistical results. The co-occurrence strength is defined as: the frequency of the two indicators co-occurring, divided by the total frequency of at least one of them appearing (i.e., the union of the clauses containing the two indicators). This ratio reflects the degree of correlation between the two indicators. The closer the ratio is to 1, the higher the probability of the two indicators appearing simultaneously in the clauses, and the stronger the correlation; the closer the ratio is to 0, the less likely the two indicators appear simultaneously, and the weaker the correlation. Existing text analysis tools can be selectively used to assist in calculating co-occurrence relationships.

[0036] A predefined co-occurrence threshold (e.g., 0.3) is used to determine whether a correlation exists between two indicators. For any target key indicator, all other key indicators are iterated over, and their co-occurrence strength with the target indicator is calculated. Indicators with co-occurrence strengths exceeding the predefined threshold are considered as associated indicators of the target indicator, thus constructing a list of associated indicators for that target indicator. Similarly, a list of associated indicators can be constructed for each key indicator. Social network analysis methods can be used to visualize the co-occurrence relationship as a co-occurrence network graph, intuitively displaying the relationship structure between indicators.

[0037] Each key indicator and its related indicator list are stored to form structured relational data.

[0038] In subsequent steps, when a key indicator is not found or lacks a value in the design file, its related indicator list can be used to locate the design file paragraph that may contain information related to that indicator, thereby narrowing the search scope and improving the efficiency of verification.

[0039] Through the above steps, this embodiment realizes the process of extracting key indicators from the specification documents and constructing a list of related indicators based on the co-occurrence relationship between the indicators, laying a data foundation for subsequent intelligent verification.

[0040] S103. Divide the design file into paragraph units and extract the keyword set for each paragraph unit.

[0041] For example, this step aims to structure the design document, dividing it into independent paragraph units, and extracting keywords from each paragraph that represent its content, laying the foundation for subsequent metric matching and paragraph location. The specific implementation process is as follows: First, the acquired design files are parsed and preprocessed. Design files typically include documents in various formats such as design specifications, design drawings, and calculation sheets, requiring the use of appropriate parsing methods based on the file type. For Word documents (.doc / .docx): Use the python-docx library to read the document content and extract paragraph text and its style information (such as heading level, font, bold, etc.).

[0042] For PDF documents: Use PyPDF2 or pdfplumber libraries to extract text content and combine it with OCR technology to process text recognition in scanned documents.

[0043] For CAD drawings: Extract text annotations and explanatory text from the drawings using CAD parsing tools.

[0044] For structured documents, such as XML / HTML documents, parsing libraries like BeautifulSoup can be used to extract the text content.

[0045] After parsing, the extracted text is cleaned, including removing extra spaces, special characters, line breaks, and other noise information to ensure text quality.

[0046] Design documents typically consist of paragraphs at different levels, including chapter headings, body paragraphs, and chart examples. They need to be divided into independent paragraph units according to the document structure. The specific division method is as follows: The design file is divided into paragraphs, with each paragraph serving as an independent unit. Paragraphs are typically separated by line breaks, periods, or other similar delimiters.

[0047] Based on the document's chapter structure, consecutive content under the same heading is treated as a single paragraph unit.

[0048] For complex documents containing multi-level headings, a hybrid strategy can be adopted: first, identify all headings in the document (based on features such as font size, bolding, and numbering patterns) and construct a tree structure for the document; then, segment the document into paragraphs according to a preset granularity (such as the smallest independent unit being a natural paragraph, or by dividing it into three levels of headings).

[0049] Each paragraph unit is assigned a unique identifier (such as P001, P002), and its chapter and paragraph type are recorded to facilitate subsequent retrieval and analysis.

[0050] Keyword extraction is performed on each paragraph unit to obtain a set of keywords that represent the content of that paragraph. The specific implementation process is as follows: The text of each paragraph unit is preprocessed, including word segmentation, part-of-speech tagging, and stop word filtering.

[0051] From the preprocessed word sequence, all nouns, noun phrases, and technical terms are extracted as candidate keywords.

[0052] Dependency parsing-based methods can be used to extract noun phrases with modifying relationships, thereby improving the quality of keywords.

[0053] The extracted candidate words are compared with the key indicators in the specification document. If a candidate word matches the name of a key indicator in the specification document (including exact match or fuzzy match based on a thesaurus), the word is marked as a "specification keyword". These keywords are directly related to the verification indicators and have a higher priority.

[0054] For candidate words not marked as canonical keywords, the term frequency-inverse document frequency (TF-IDF) method is used to calculate the weight of each word. The TF-IDF value comprehensively reflects the importance of a word in the current paragraph and its distinctiveness in the entire document; the higher the value, the more representative the word is of the paragraph's theme.

[0055] After calculating the TF-IDF value of each candidate word, select the top 5-10 words with the highest weights as "paragraph topic words".

[0056] The obtained "standard keywords" and "paragraph topic words" are combined to form the keyword set for that paragraph unit. Duplicate words are then removed.

[0057] The identifier information of each paragraph unit and its corresponding set of keywords are stored to form a structured paragraph-keyword mapping table.

[0058] S104. For any paragraph unit, match its keyword set with each key indicator, determine the indicators to be verified in the design document from the key indicators, and locate the candidate paragraphs corresponding to the indicators to be verified.

[0059] In this embodiment, for any paragraph unit, its keyword set is matched with each key indicator to determine the indicator to be verified in the design file from the key indicators, and the candidate paragraph corresponding to the indicator to be verified is located, specifically including: For any given paragraph unit, match its keyword set with each key metric; If the keyword set includes a key indicator, and the paragraph unit includes verification information for that key indicator, then the key indicator is determined to be a non-verification indicator. If the keyword set does not include a key indicator, or if the keyword set includes a key indicator but the paragraph unit does not include verification information for that key indicator, then the key indicator is identified as an indicator to be verified. For any indicator to be verified, each associated indicator in its associated indicator list is matched with the keyword set of each paragraph unit. The paragraph unit containing the successfully matched associated indicator is determined as the candidate paragraph for that indicator to be verified.

[0060] For example, this step aims to dynamically determine which indicators need further verification by traversing each paragraph unit, matching its keyword set with the key indicators in the specification document, and using the list of associated indicators to locate candidate paragraphs that may contain this indicator information. The specific implementation process is as follows: For each segmented paragraph unit, the process is performed sequentially. For the current paragraph unit, its keyword set is matched one by one with each extracted key indicator.

[0061] Matching methods include: exact match: the keyword and the key indicator name are exactly the same; synonym match: based on the domain thesaurus, it is determined whether the keyword and the key indicator are synonymous; fuzzy match: based on edit distance or semantic similarity, it is determined how close the keyword and the key indicator are.

[0062] Based on the matching results and whether the paragraph contains verification information for the indicators, we dynamically determine which indicators are "indicators to be verified" and which are "indicators not to be verified." The specific judgment logic is as follows: If the keyword set of the current paragraph unit contains a key indicator, and the verification information (such as numerical value, unit, performance parameter, etc.) of that key indicator can be extracted from the paragraph unit, it means that the indicator has been clearly described in the current paragraph and meets the verification requirements, requiring no further processing. At this point, the key indicator is identified as a "non-verification indicator," and the next indicator is processed.

[0063] Suppose a key indicator in the specification document is "rated voltage," which is required to be "10kV." The keyword "voltage" is matched to the keyword set of a certain paragraph, and the value "10" and the unit "kV" are extracted from the paragraph's content "powered by two independent 10kV power supplies." After comparison, this value meets the specification requirement. Therefore, "rated voltage" is determined to be a non-verifiable indicator and requires no further processing.

[0064] If a key indicator is not included in the keyword set of the current paragraph unit, or if the indicator is included in the keyword set but the corresponding verification information cannot be extracted from the paragraph (e.g., missing values, unclear units), it indicates that the indicator is not fully described in the current paragraph and requires further verification. In this case, the key indicator is designated as an "indicator to be verified".

[0065] Suppose a key indicator in the specification document is "Protection Level," which requires "IP54." If no match is found for "Protection Level" or its synonyms in the keyword set of a certain paragraph, then "Protection Level" is identified as the indicator to be verified.

[0066] For example, suppose a paragraph contains the keyword "protection level," but the paragraph content only states "the equipment protection level should meet the specifications," without mentioning specific values. In this case, although the indicator exists, there is a lack of verification information, so "protection level" is also identified as an indicator to be verified.

[0067] For each paragraph unit, record all the indicators to be verified corresponding to that unit, forming a "paragraph-indicator to be verified" mapping relationship. These indicators to be verified will be the focus of verification in subsequent steps.

[0068] For each indicator to be verified, it is necessary to locate the design document paragraphs that may contain relevant information about that indicator, as candidates for further in-depth analysis. The location method is based on the constructed list of related indicators: For any indicator to be verified, obtain its list of associated indicators (i.e., other indicators whose co-occurrence intensity with this indicator exceeds a preset threshold). Match each associated indicator in the list with the keyword set of all paragraph units.

[0069] For each related indicator, if the keyword set of a certain paragraph unit contains the related indicator, then the paragraph unit is determined as a "candidate paragraph" for the indicator to be verified.

[0070] For certain indicators to be verified, there may be situations where none of their associated indicators match any paragraph unit. In this case, a supplementary strategy is adopted: all paragraph units in the entire design file containing any specification keyword are considered as candidate paragraphs for that indicator to be verified.

[0071] For each indicator to be verified, all its corresponding candidate paragraphs are deduplicated to form a set of candidate paragraphs for that indicator.

[0072] S105. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph, and construct a hierarchical tree to represent the semantic hierarchical structure of the paragraph based on the semantic level value.

[0073] In this embodiment, for any candidate paragraph corresponding to any indicator to be verified, the semantic level value of each keyword in the paragraph is obtained, and a hierarchical tree representing the semantic hierarchy of the paragraph is constructed based on the semantic level value, specifically including: Obtain a pre-built vocabulary library, which includes at least three words and the semantic level value corresponding to each word. The semantic level value is used to quantify the semantic generalization degree of the words. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph from the thesaurus; Based on the semantic level values ​​of each keyword, the keyword set of the candidate paragraph is clustered from bottom to top to construct the hierarchical tree of the candidate paragraph.

[0074] Obtain a pre-built vocabulary, specifically including: The text in standard documents, qualified design documents, and related field documents is cleaned, segmented into words, and divided into sentences to obtain a corpus; Each corpus in the corpus is processed using a word embedding model to obtain word vectors for each corpus; Obtain at least three clustering density parameters. For each clustering density parameter, perform density clustering on all word vectors based on that clustering density parameter to obtain the clusters under that clustering density parameter. For each cluster under each cluster density parameter, calculate the average of all word vectors in the cluster to obtain the cluster center vector of that cluster; For each corpus under each cluster density parameter, calculate the distance between the word vector of the corpus and the cluster center vectors of all clusters under that density, convert the distance into similarity, and normalize the similarity to obtain the membership degree of the corpus to each cluster. For each corpus under each density parameter, record its maximum membership degree and second maximum membership degree, and calculate the difference between the two; For each corpus, the difference sequence under different density parameters is obtained to form the difference change curve of the corpus; The difference change curve for each corpus is fitted, and the slope of the trend representing the decrease of the difference as the density parameter increases is calculated. Sort the absolute values ​​of the trend slopes of all corpora to determine the maximum and minimum trend slopes; For each corpus, the difference between the absolute value of its trend slope and the minimum value of its trend slope is divided by the difference between the maximum value of its trend slope and the minimum value of its trend slope to calculate the normalized semantic level value. Each corpus and its corresponding semantic level value are stored to construct a vocabulary database.

[0075] For example, this step aims to obtain the semantic hierarchy value of each keyword using a pre-built vocabulary, and construct a paragraph hierarchy tree based on these values ​​to structurally represent the semantic hierarchy of paragraph content. The specific implementation process is as follows: First, a vocabulary database containing words and their semantic hierarchy values ​​needs to be pre-built. The semantic hierarchy value quantifies the degree of semantic generalization of a word; a higher value indicates a more general and higher-level expression among words with similar meanings, while a lower value indicates a more specific and lower-level expression. The vocabulary database construction process is as follows: We collect specification documents, qualified design documents, and relevant domain documents to form a domain corpus. We then clean the text in the corpus, removing irrelevant characters and noise; finally, we perform word segmentation and sentence splitting to create a training corpus.

[0076] For example, in the field of building electrical engineering, standard documents, qualified design documents, and related domain documents can be collected to form a domain corpus. These documents include, but are not limited to, technical documents, standard documents, and design drawings in various formats, such as text documents, PDF documents, and structured documents. The text in the corpus is cleaned to remove irrelevant characters and noise; then, word segmentation and sentence segmentation are performed to form a corpus that can be used for training.

[0077] The preprocessed corpus is trained using a word embedding model to obtain the word vector for each word. Specific parameter settings are as follows: Model selection: Choose the Skip-gram model or CBOW model from Word2Vec, or the GloVe model; Vector dimension: set to 200 dimensions to fully express the semantic features of words; Context window: Set to display 5 words before and after to capture local context information; Minimum word frequency threshold: Set to 5 to ensure that low-frequency technical terms can also be effectively represented; The preprocessed corpus is input into the model, which learns the co-occurrence patterns of words through a neural network, iteratively updates the word vector parameters until convergence, and finally outputs the vector representation of each word.

[0078] The cluster density parameter can be obtained in ways including but not limited to the following: Calculate the Euclidean distance (or cosine distance) between all pairwise word vectors to obtain the distance matrix D. Find the minimum and maximum values ​​of all non-zero distances. Generate K cluster density parameters (e.g., K=20) from the minimum to the maximum values, using logarithmic intervals. Optionally, the cluster density parameters can also be manually determined based on the minimum and maximum values ​​of the non-zero distances, selecting appropriate parameters according to the actual parameter distribution. These parameters are used to control the clustering results at different granularities.

[0079] For each clustering density parameter (from smallest to largest), the DBSCAN algorithm is used to perform density clustering on the word vectors. Parameter settings: neighborhood radius equals the clustering density parameter, and the minimum number of neighborhood points is 2. The clustering result is several clusters, some of which may be labeled as noise. To ensure uniform processing, each noise point is treated as an independent cluster (i.e., each noise point forms its own cluster with a cluster size of 1). The final result is the set of clusters at this clustering density.

[0080] For example, when the clustering density parameter is small, the clustering results are fine, and "transformer", "interchange transformer" and "reactor" may be divided into different clusters; as the clustering density parameter gradually increases, these terms may be clustered into the same "electrical equipment" cluster.

[0081] For each cluster at each clustering density, calculate the average of all word vectors within the cluster to obtain the cluster center vector. If the cluster size is 1, then the cluster center is the point itself.

[0082] For each word at each cluster density, calculate the Euclidean distance *d* between the word vector of that word and the cluster center vectors of all words at that density, and convert the distance *d* into a similarity score. The calculation formula can be: ; Then, softmax normalization is applied to the similarity of all clusters to obtain the membership degree of each word to each cluster. For each word, its maximum and second-largest membership degrees are obtained. Membership degree reflects the degree to which a word belongs to each semantic cluster. The difference between the maximum and second-largest membership degrees reflects the distinguishability of the word relative to other clusters. The larger the difference, the more clearly the word belongs to a cluster and the simpler its semantics; the smaller the difference, the more ambiguous the word is between two clusters and the more broad its semantics may be.

[0083] For each word, record its difference across all cluster density parameters to form a difference sequence. Observe the trend of the difference for each word as the cluster density increases (i.e., the cluster granularity becomes coarser). If the difference remains relatively large and essentially constant, it indicates that the word's semantics are fixed; if the difference decreases significantly with increasing cluster density, it indicates that the word's semantic range expands, meaning it can cover a wider range of concepts.

[0084] For each word, a linear regression is performed on its difference sequence to obtain the slope of the relationship between the cluster density parameter and the difference. The larger the absolute value of the slope, the more obvious the downward trend and the wider the semantic range.

[0085] Sort all words by the absolute value of their slopes and find the maximum and minimum values. Then, use the maximum and minimum values ​​to normalize the slope of each word's relation to the word, obtaining its semantic level value. The normalized semantic level value is distributed between 0 and 1, with a larger value indicating a higher degree of semantic generalization (higher level) and a smaller value indicating a more specific word (lower level).

[0086] Each word and its corresponding semantic level value are stored to construct a word library.

[0087] For any candidate paragraph corresponding to any indicator to be verified, the semantic level value of each keyword in the paragraph is obtained from the pre-constructed thesaurus.

[0088] For example, assuming the keyword set for the candidate paragraph is {equipment, protection level, enclosure, seal, IP65}, the query results from the thesaurus are: equipment: 0.92; protection level: 0.58; enclosure: 0.45; seal: 0.38; IP65: 0.15.

[0089] Based on the semantic level values ​​of each keyword, a bottom-up hierarchical clustering process is performed on the keyword set of the candidate paragraph to construct a hierarchical tree for the candidate paragraph. The specific implementation process is as follows: Each keyword is treated as a leaf node, and the value of the node is the semantic hierarchy value of that keyword. Initially, each node belongs to its own category.

[0090] Repeat the following steps until all nodes are merged into a single root node: Calculate the distance between all current classes, where the distance is defined as the difference in the semantic hierarchy values ​​of nodes in the two classes (such as the difference in minimum, maximum, or average values); select the two classes with the closest distance to merge them to generate a new parent node; the semantic hierarchy value of the new parent node is the average (or maximum or minimum) of the semantic hierarchy values ​​of the two child nodes; record the merging order and hierarchical relationship.

[0091] Based on the clustering process, a hierarchical tree is generated. The leaf nodes represent specific keywords, the internal nodes represent abstract semantic categories, and the root node represents the overall semantics of the paragraph. The depth of the tree reflects the number of abstract levels within the paragraph content.

[0092] Store the hierarchical tree of each candidate paragraph, including the tree structure, the semantic hierarchy value of each node, and the parent-child relationship between nodes.

[0093] Through the above steps, this embodiment obtains semantic level values ​​for the keywords of each candidate paragraph and constructs a hierarchical tree representing the semantic hierarchy of the paragraph based on these values. The hierarchical tree not only reflects the composition of the paragraph content but also quantifies the degree of abstraction of each part through the semantic level values ​​of the nodes, laying the foundation for subsequent paragraph structure similarity comparison.

[0094] S106. Calculate the structural similarity between the hierarchical trees of different candidate paragraphs and select paragraph pairs with similar structures.

[0095] In this embodiment, the structural similarity between the hierarchical trees of different candidate paragraphs is calculated, and paragraph pairs with similar structures are selected, specifically including: For any two candidate paragraphs, calculate the average semantic hierarchy value of their hierarchical trees; Compare the average semantic level values ​​of the two, and identify the candidate paragraph with the lower average semantic level value as the summary paragraph, and identify the candidate paragraph with the higher average semantic level value as the detailed paragraph. The hierarchical tree of the summary paragraph is compared with the hierarchical tree of the detailed paragraph layer by layer, and the distribution similarity of the semantic hierarchy values ​​corresponding to each node is calculated. If the number of levels with distribution similarity exceeding the preset structural similarity threshold accounts for more than the preset proportion threshold of the total number of levels in the hierarchical tree, then the two candidate paragraphs are determined to be a pair of paragraphs with similar structure.

[0096] For example, this step aims to determine whether two candidate paragraphs are structurally similar by comparing their hierarchical trees; that is, whether the level of detail in one paragraph matches the level of generalization in the other. Paragraphs with high similarity provide a foundation for subsequent keyword matching by offering different levels of granularity descriptions that may correspond to the same topic. The specific implementation process is as follows: For each candidate paragraph's hierarchical tree, its average semantic level value is first calculated. The average semantic level value reflects the overall level of abstraction of the paragraph; the smaller the value, the more general (higher level) the paragraph content is, and the larger the value, the more specific (lower level) the paragraph content is.

[0097] The average semantic hierarchy value is calculated by averaging the semantic hierarchy values ​​of all leaf nodes (i.e., specific keywords) in the hierarchy tree. Since leaf nodes represent the most specific conceptual expressions in a paragraph, their average value can effectively reflect the level of detail in the paragraph.

[0098] For any two candidate paragraphs, compare their average semantic level values. The paragraph with the lower average semantic level value is designated as the "summary paragraph," and the paragraph with the higher average semantic level value is designated as the "detailed paragraph." The content of a summary paragraph is relatively abstract, describing higher-level concepts; the content of a detailed paragraph is relatively concrete, describing lower-level concepts or specific implementations.

[0099] The hierarchical tree of the summary paragraphs is compared layer by layer with the hierarchical tree of the detailed paragraphs. Starting from the root node, the distribution similarity of the semantic hierarchy values ​​corresponding to each node is calculated. The specific implementation process is as follows: Since the two paragraphs may have different hierarchical tree depths, hierarchical alignment is necessary. Typically, the shallower tree is used as a reference, and the first few levels of the deeper tree are aligned with it. Alternatively, dynamic programming can be used to find the optimal matching level between the two trees.

[0100] For each level, extract the semantic level values ​​of all nodes in the summary paragraph hierarchy tree to form a distribution vector; similarly, extract the semantic level values ​​of all nodes in the detailed paragraph hierarchy tree to form a distribution vector.

[0101] The similarity between two distribution vectors can be calculated using one of the following methods: cosine similarity, Jaccard coefficient, Pearson correlation coefficient, etc.

[0102] Two preset thresholds are used: a structural similarity threshold (e.g., 0.8) and a proportion threshold (e.g., 0.6). The number of levels with a similarity exceeding the structural similarity threshold is counted, and their proportion of the total number of levels in the hierarchical tree is calculated. If this proportion exceeds the preset proportion threshold, the two candidate paragraphs are considered a pair of paragraphs with similar structure.

[0103] Optionally, the specific values ​​of the structural similarity threshold and the proportion threshold can be adjusted according to the actual application scenario: If strict hierarchical matching is required, a higher threshold (such as 0.9) can be set to ensure that only highly similar layers are counted. If you want to increase the recall rate of similar paragraph pairs, you can set a lower threshold (such as 0.7) to allow for a certain degree of difference; If you want two paragraphs to be similar at almost all levels, you can set a higher ratio threshold (such as 0.8). If only the main levels are required to be similar, a lower ratio threshold can be set (e.g., 0.5).

[0104] For all candidate paragraphs, iterate through any two paragraphs and repeat the above steps to find all paragraph pairs that satisfy the structural similarity condition. Record each similar paragraph pair and its related information, including: summary paragraph ID, detailed paragraph ID, similarity distribution at each level, and the proportion of levels that meet the condition.

[0105] The selected structurally similar paragraph pairs are stored to form structured data for use in subsequent steps.

[0106] This step of structural similarity calculation effectively identifies paragraph pairs in design documents that describe the same topic but at different levels of abstraction. For example, one paragraph may provide a general description of "power supply system requirements," while another may describe "specific power supply parameters" in detail. Their structural similarity suggests that they may address different aspects of the same verification indicator. This structural similarity analysis overcomes the limitations of traditional keyword matching, laying the foundation for subsequent cross-paragraph contextual understanding and capturing the expression of indicators at different levels.

[0107] Through the above steps, this embodiment realizes the structural similarity calculation of the hierarchical tree of different candidate paragraphs, and filters out paragraph pairs with similar structures based on the proportion of levels with distribution similarity exceeding the threshold, providing an accurate candidate range for subsequent keyword matching.

[0108] S107. Match keywords in any pair of paragraphs with similar structures, and select candidate keywords that correspond to the indicators to be verified based on the matching results.

[0109] In this embodiment, keywords in any pair of paragraphs with similar structures are matched, and candidate keywords corresponding to the indicators to be verified are selected based on the matching results. Specifically, this includes: For any pair of paragraphs, obtain the keyword set of the summary paragraph and the keyword set of the detailed paragraph in that pair; A bipartite graph is constructed using the keyword set of the summary paragraph and the keyword set of the detailed paragraph as two vertex sets. The weight of each edge is calculated based on the semantic similarity and semantic level value difference between the two keywords. The KM algorithm is used to solve the maximum weight matching of the bipartite graph, and the matching pairs and the matching degree corresponding to each matching pair are obtained. Based on the matching degree from smallest to largest, a preset number of matching pairs are selected, and the keywords belonging to the detailed paragraphs in each selected matching pair are used as candidate keywords corresponding to the indicators to be verified.

[0110] For example, this step aims to perform precise keyword matching on the selected structurally similar paragraph pairs to identify the words in the detailed paragraphs that least correspond to the keywords in the summary paragraphs. These words with low matching scores are likely another expression of the metric in the design document, i.e., candidate keywords corresponding to the metric to be verified. The specific implementation process is as follows: For any structurally similar paragraph pair determined in step S106, obtain the keyword set of the summary paragraph and the keyword set of the detailed paragraph in the paragraph pair.

[0111] Using the keyword set A of the summary paragraph and the keyword set B of the detailed paragraph as two vertex sets, a bipartite graph G=(A, B, E) is constructed. Each keyword in A is connected to each keyword in B by an edge, and the weight E of each edge is calculated based on the semantic similarity and semantic level difference between the two keywords.

[0112] Semantic similarity calculation is as follows: For any two keywords, calculate their semantic similarity. One of the following methods can be used: word vector cosine similarity, thesaurus matching, or a combination thereof.

[0113] The semantic hierarchy value difference calculation is as follows: Obtain the semantic level value of each keyword from the constructed vocabulary, and calculate the difference in semantic level values ​​between two keywords. The smaller the difference, the closer the two words are in terms of abstraction; the larger the difference, the more general one word is and the more specific the other.

[0114] The weight of each edge is calculated by combining semantic similarity and semantic hierarchy differences. A higher weight indicates a better match between the two keywords; a lower weight indicates a less good match. Common calculation methods include weighted combination, product form, and harmonic average.

[0115] After constructing the complete bipartite graph, the KM algorithm (Kuhn-Munkres algorithm, also known as an extension of the Hungarian algorithm) is used to solve for the maximum weight matching in the bipartite graph. The KM algorithm can find the matching scheme that maximizes the total weight in polynomial time, that is, assigning the keyword of each summary paragraph to the most matching detailed paragraph keyword (or vice versa), and each keyword can only be matched once. The KM algorithm is a well-known existing technique, and the matching process will not be described in detail.

[0116] Sort all matching pairs by matching degree from least to most, i.e., from least matching to most matching. The lower the matching degree, the less similar the keyword in the detailed paragraph is to the corresponding keyword in the summary paragraph. It is more likely that there is no direct corresponding new word in the summary paragraph, and it is very likely that it is a synonym of the specification indicator in the design document.

[0117] Sort the matching results by matching degree from lowest to highest; pre-determine a number Z (e.g., Z=2), and select the top Z matching pairs (i.e., the lowest matching degree) from the sorted pairs. These matching pairs represent the keywords in the detailed paragraphs that least match the summary paragraphs. From each selected matching pair, extract the keywords belonging to the detailed paragraphs as candidate keywords corresponding to the indicators to be verified. If there are multiple paragraph pairs with similar structures, repeat the above steps for each paragraph pair to obtain multiple sets of candidate keywords. Merge these sets and remove duplicates to obtain the final list of candidate keywords.

[0118] Optionally, the value of the preset quantity Z can be adjusted according to the actual application scenario: Fixed value method: Set a fixed value, such as Z=3 or Z=5. This method is simple and easy to implement, and is suitable for most scenarios.

[0119] Proportional method: Selecting keywords based on a percentage of the total number of matching pairs, such as the top 10% of matching pairs. This method can adapt to keyword sets of different sizes.

[0120] Thresholding method: Select all matching pairs with a matching score below a certain threshold. For example, select all matching pairs with a matching score less than 0.2.

[0121] Inflection point method: Observe the inflection point of the matching degree ranking curve and select matching pairs before the inflection point. When the matching degree drops significantly, the point before the drop may be more meaningful.

[0122] By using bipartite graph matching and candidate keyword filtering in this step, we can accurately identify alternative expressions that may represent specification indicators from structurally similar paragraph pairs. The core principle is that in structurally similar paragraphs, the summary paragraph describes higher-level concepts (such as specification indicators), while the detailed paragraph describes lower-level concepts (such as specific implementations). Through maximum weight matching, the keywords in the summary paragraph are aligned with the keywords in the detailed paragraph. Keywords with low matching scores are precisely the specific expressions unique to the detailed paragraphs and not covered by the summary paragraphs. These expressions are likely synonyms or specific parameter forms of the specification indicators in the design documents.

[0123] For example, in the above example, the indicator to be verified might be "protection level," and the keywords such as "outer shell" and "sealing" selected from the detailed paragraphs are specific implementation elements closely related to "protection level." These keywords will be used as candidates in subsequent knowledge base verification steps.

[0124] Through the above steps, this embodiment achieves accurate matching of keywords in structurally similar paragraph pairs, and filters out candidate keywords corresponding to the indicators to be verified based on the matching degree, providing accurate input for subsequent external knowledge base verification.

[0125] S108. Compare and verify the candidate keywords with the preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified.

[0126] In this embodiment, candidate keywords are compared and verified with a preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified, specifically including: Obtain a preset knowledge base, which includes standard names, synonyms, abbreviations, and definitions of domain terms; For each candidate keyword, retrieve the corresponding standard term from the preset knowledge base; If the retrieved standard terminology matches the name of the indicator to be verified, it is determined that there is a semantic relationship between the candidate keyword and the indicator to be verified. If the standard term corresponding to the candidate keyword is not found in the preset knowledge base, the semantic similarity between the candidate keyword and the name of the indicator to be verified is calculated. If the semantic similarity exceeds the preset semantic similarity threshold, it is determined that there is a semantic association between the candidate keyword and the indicator to be verified.

[0127] For example, this step aims to compare and verify the selected candidate keywords with an external authoritative knowledge base to confirm whether they have the same meaning as the indicator to be verified, thereby avoiding misjudgment and ensuring the accuracy of the verification results. The specific implementation process is as follows: First, a pre-defined external knowledge base is acquired. This knowledge base is an authoritative data source containing standard names, synonyms, abbreviations, and definitions of domain terminology, and can be built based on resources such as national standards, industry specifications, and domain dictionaries. Each term record in the pre-defined knowledge base contains the following fields: Standard Name: The standardized terminology expression within the domain, such as "protection level"; Synonyms: Other words with the same meaning as the standard name, such as "protection level" or "IP level"; Abbreviation: The abbreviated form of the standard name, such as "IP" (Ingress Protection); Definition: The detailed definition and explanation of the term; Related Terms: Other terms related to this term and their relationship types. The pre-defined knowledge base can be constructed in the following ways: National Standards Library: Integrating terminology definitions from standards such as GB, IEC, and ISO; Domain Dictionary: Including terminology explanations from professional dictionaries and encyclopedias; Expert Knowledge: A network of terminology relationships compiled by domain experts; Existing Resources: Utilizing domain subsets of existing terminology databases such as WordNet and HowNet.

[0128] For each selected candidate keyword, the corresponding standard term is retrieved from the pre-defined knowledge base. The retrieval can employ several methods: exact match: the candidate keyword perfectly matches the standard name, synonym, or abbreviation in the knowledge base; fuzzy match: based on edit distance or string similarity, the closest term is found; indexed retrieval: using search engines such as Elasticsearch to achieve efficient full-text retrieval.

[0129] If the retrieved standard term matches the name of the indicator to be verified, it is directly determined that there is a semantic association between the candidate keyword and the indicator to be verified. When the standard term corresponding to the candidate keyword in the knowledge base is completely consistent with the name of the indicator to be verified, it indicates that the candidate keyword has been explicitly identified by the knowledge base as a synonym, abbreviation, or equivalent expression of the indicator to be verified, and therefore its semantic association can be directly confirmed. If the standard term corresponding to the candidate keyword is not found in the preset knowledge base, or the retrieved standard term does not match the name of the indicator to be verified, it is necessary to calculate the semantic similarity between the candidate keyword and the name of the indicator to be verified, and determine whether there is a semantic association by using a similarity threshold. One of the following methods can be used to calculate semantic similarity: Knowledge base relationship network: Based on the relationship between terms in the knowledge base (such as hyponymy / hypernymy, correlation), calculate the distance between the two terms in the relationship network; Word vector similarity: Use the domain word vectors trained in step S105 to calculate the cosine similarity between the candidate keyword and the name of the indicator to be verified; Hybrid method: Combine multiple similarity indicators and take the weighted average.

[0130] A preset semantic similarity threshold (e.g., 0.6) is used to determine whether there is a semantic connection between two terms. The similarity calculation result is compared with the threshold to arrive at a judgment. Finally, by combining the results of direct matching and semantic similarity calculation, all candidate keywords are judged, and relevant information for each valid match is recorded.

[0131] Optionally, special cases may be handled as follows: Handling of specific numerical keywords: For keywords containing specific values, such as "IP65" and "10kV", special handling is required: If the knowledge base clearly associates the value with the indicator to be verified (e.g., "IP65" is a specific level of "protection level"), it can be determined as valid; otherwise, it is necessary to proceed to the next step S109 to extract and compare the values, and indirectly verify them by whether the values ​​meet the specification requirements.

[0132] Disambiguation of polysemous words: If a candidate keyword has multiple meanings, it needs to be disambiguated in conjunction with the context. For example, "sealing" may refer to "sealing components" in the mechanical field and "sealing adhesive" in the construction field. In this case, other keywords in the paragraph containing the keyword can be consulted to determine its domain affiliation, and then matched with the corresponding meaning in the knowledge base.

[0133] The verified candidate keywords are marked as "valid matches" and their relevant information is recorded, including: the candidate keyword itself, the corresponding indicator to be verified, the verification method (direct matching / semantic similarity), the basis for association (synonyms / abbreviations / similarity values, etc.), and the paragraph position.

[0134] This step of external knowledge base verification effectively confirms the semantic relevance between candidate keywords and the indicators to be verified, avoiding potential misjudgments that might arise from relying solely on semantic matching. Specifically: direct matching utilizes explicit relationships such as synonyms and abbreviations in authoritative knowledge bases to ensure the accuracy of verification; semantic similarity calculation, as a supplement, covers expressions not included in the knowledge base but actually related, improving recall; the dual verification mechanism achieves a good balance between accuracy and recall, providing a reliable basis for subsequent numerical extraction and verification result generation.

[0135] Through the above steps, this embodiment achieves accurate verification of candidate keywords, ensuring that only words that are truly semantically related to the indicators to be verified will enter the final verification stage, thereby improving the accuracy and reliability of the entire intelligent verification method.

[0136] S109. Extract verification information from the paragraphs containing candidate keywords whose semantic relevance has been verified, compare the verification information with the verification requirements of the indicators to be verified in the specification document, and generate verification results.

[0137] In this embodiment, verification information is extracted from the paragraphs containing candidate keywords whose semantic relevance has been verified. This verification information is then compared with the verification requirements of the indicators to be verified in the specification document to generate verification results. Specifically, this includes: For each candidate keyword that passes the semantic relevance verification, locate the sentence containing the candidate keyword and at least one sentence before and after it, as the scope of information extraction; Extract numerical values ​​and unit combinations from the information extraction range to serve as the verification information corresponding to the candidate keyword; The extracted verification information was converted to units according to the unit requirements of the indicators to be verified in the standard documents; Obtain the numerical requirements and modifiers of the indicator to be verified in the specification document, and compare the converted verification information with the numerical requirements and modifiers; Based on the comparison results, the verification result of the indicator to be verified is determined, including whether the verification is passed, failed, or the information is missing.

[0138] For example, this step aims to extract key information from the paragraphs containing candidate keywords that have passed the semantic relevance verification, obtain specific numerical parameters, compare them with the verification requirements in the specification document, and finally draw verification conclusions for each indicator to be verified. The specific implementation process is as follows: For each candidate keyword that passes the semantic relevance verification, first locate the specific position of the candidate keyword in the paragraph and determine the scope of information extraction.

[0139] Locate the candidate keywords within the paragraph text. Since the same keyword may appear multiple times in a paragraph, it is necessary to record all occurrences or select the location most likely to contain numerical information (such as the first occurrence, the most recent occurrence of a numerical value, etc.).

[0140] Centered on the sentence containing the candidate keyword, extend at least one sentence forward and one sentence backward to define the information extraction scope. The aim is to capture numerical information that may appear in the keyword's context, as numerical values ​​sometimes appear in the sentence preceding or following the keyword.

[0141] From the defined information extraction range, extract the numerical values ​​and unit combinations as the verification information corresponding to the candidate keyword.

[0142] Use regular expressions or rule-based extraction methods to identify numerical patterns in the text. Common numerical patterns include: integers (e.g., "65", "30", "100"); decimals (e.g., "5.8", "0.5", "3.14"); fractions (e.g., "1 / 2", "3 / 4"); and ranges (e.g., "10-20", "5~8", "not less than 15"). Identify unit information near the numerical values, including: standard units (e.g., "kV", "A", "mm", "℃"); unit symbols (e.g., "%", "°", "'"); full unit names (e.g., "kilovolt", "ampere", "millimeter"); and compound units (e.g., "kW·h", "m / s²"). Combine the identified numerical values ​​with the most likely associated units to form "numerical value + unit" verification information. If multiple numerical-unit combinations exist within the same range, determine which one is most relevant to the candidate keywords based on the context.

[0143] The extracted verification information is converted to the appropriate units according to the unit requirements of the indicators to be verified in the standard documents, ensuring that the numerical comparisons are conducted under the same dimensions. A unit conversion rule library is established, containing conversion relationships between various units. Based on the unit requirements in the standard documents, the extracted verification information is converted to values ​​in the target units.

[0144] Obtain the numerical requirements and modifiers for the indicator to be verified in the specification document. Compare the converted verification information with the numerical requirements and modifiers to determine whether it meets the requirements. From the key indicator information, obtain the following for the indicator to be verified: numerical requirements (e.g., "54" in "10", "90", or "IP54"); unit requirements (e.g., "kV", "minutes", or "IP"); modifiers (e.g., "should not be less than", "should not be greater than", "should be", "should not be less than", etc.). Different comparison rules are used depending on the type of modifier. Compare the converted verification information with the specification requirements to draw a preliminary conclusion on whether it meets the requirements.

[0145] Based on the comparison results, the final verification result for each indicator to be verified is determined. The verification results are divided into three categories: The verification is considered passed when one of the following conditions is met: valid verification information can be extracted, and after unit conversion, the information fully meets the numerical requirements and modifiers in the specification document; the verification information corresponding to multiple candidate keywords meets the requirements (the best or comprehensive judgment is taken).

[0146] The verification is deemed unsuccessful if any of the following conditions are met: Valid verification information can be extracted, but after unit conversion, the information does not meet the numerical requirements or modifiers in the specification document; or the verification information corresponding to multiple candidate keywords is inconsistent, with some meeting the requirements and others not. The worst-case scenario will be used for judgment.

[0147] Information is considered missing when any of the following conditions are met: no numerical information can be extracted from the paragraph containing the candidate keyword; the extracted numerical information is incomplete (e.g., there is a value but no unit, or there is a unit but no value); or the extracted numerical information cannot be reliably associated with the candidate keyword.

[0148] When a single indicator to be verified corresponds to multiple verified candidate keywords, the verification information of all candidate keywords needs to be combined to arrive at the final verification conclusion.

[0149] Summarize the verification information of all valid candidate keywords under the indicator to be verified. The following rules are used for comprehensive judgment: Priority rule: Numerical information takes precedence over information without numerical information. As long as any candidate keyword yields a valid numerical value and meets the requirements, the overall result is deemed passable; Worst-case scenario principle: For security-related indicators, if there is any non-compliant numerical information, even if other compliant values ​​exist, the result is deemed failable; Majority voting: For indicators with high consistency requirements, a majority voting method can be used.

[0150] Summarize the verification results of all indicators to be verified and generate a complete verification result report.

[0151] Through the information extraction and comparison in this step, accurate quantitative judgment of the indicators to be verified is achieved: the information extraction range positioning ensures the accuracy of numerical extraction and avoids omitting key information in the context; unit conversion eliminates misjudgments caused by inconsistent units, ensuring that the comparison is conducted under the same dimension; multi-dimensional comparison comprehensively considers numerical requirements and modifiers to fully evaluate the compliance of indicators; the result classification clearly defines three states: pass, fail, and missing information, providing clear guidance for subsequent processing; the report generation fully records the verification process and results, facilitating review and traceability; through the above steps, this embodiment completes the entire process of extracting verification information from design documents, comparing it with specification requirements, and generating verification results, realizing intelligent and accurate verification of project documents.

[0152] In summary, this invention, through the construction of a list of associated indicators for key indicators, achieves indicator association analysis based on co-occurrence relationships, accurately identifying the indicators to be verified and locating candidate paragraphs. By using a pre-constructed vocabulary and semantic level value calculation, it achieves a quantitative assessment of the semantic generalization degree of keywords, and effectively identifies similar paragraph pairs at different levels of abstraction based on hierarchical tree construction and structural similarity calculation. Through bipartite graph matching and the KM algorithm, candidate keywords corresponding to the indicators to be verified are selected from structurally similar paragraph pairs. Combined with external knowledge base verification and numerical extraction comparison, accurate verification results are finally generated. Compared to existing technologies, this method overcomes the semantic understanding limitations of traditional keyword matching, enabling cross-paragraph context understanding, identifying the expression of indicators at different levels, and avoiding misjudgments through external knowledge base verification, significantly improving the accuracy, intelligence, and automation level of project document verification.

[0153] This invention also proposes an information accuracy comparison system based on intelligent verification of project documents. Please refer to [link / reference]. Figure 2 The diagram shows a structural diagram of an information accuracy comparison system based on intelligent verification of project documents provided by an embodiment of the present invention. The system includes: a data acquisition module 101, a data processing module 102, and a comparison and verification module 103.

[0154] The data acquisition module 101 is used to acquire the design documents to be verified and the specification documents used as the basis for verification; The data processing module 102 is used to extract key indicators from the specification documents and construct a list of related indicators for each key indicator based on the co-occurrence relationship of each key indicator in the specification clauses. Divide the design file into paragraph units and extract the keyword set for each paragraph unit; For any paragraph unit, its keyword set is matched with each key indicator, the indicators to be verified in the design document are determined from the key indicators, and the candidate paragraphs corresponding to the indicators to be verified are located. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph, and construct a hierarchical tree to represent the semantic hierarchical structure of the paragraph based on the semantic level value; Calculate the structural similarity between the hierarchical trees of different candidate paragraphs, and filter out paragraph pairs with similar structures; Match keywords in any pair of paragraphs with similar structures, and filter out candidate keywords that correspond to the indicators to be verified based on the matching results; The candidate keywords are compared and verified with the preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified. The comparison and verification module 103 is used to extract verification information from the paragraphs containing candidate keywords whose semantic relevance has been verified, and compare the verification information with the verification requirements of the indicators to be verified in the specification document to generate verification results.

[0155] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the information accuracy comparison system based on intelligent verification of project documents and the information accuracy comparison method based on intelligent verification of project documents provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0156] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0157] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for accurate information comparison based on intelligent verification of project documents, characterized in that, include: Obtain the design documents to be verified and the specification documents used as the basis for verification; Key indicators are extracted from the specification documents, and based on the co-occurrence relationships of each key indicator in the specification clauses, a list of associated indicators for each key indicator is constructed, specifically including: The standard document is divided into clauses, with each clause treated as an independent unit; Perform natural language processing on each independent unit to identify key metrics; Calculate the frequency of each key indicator in all independent units, and the frequency of any two key indicators appearing together in the same independent unit; For any two key indicators, the ratio of the frequency of their co-occurrence to the total frequency of at least one of them is determined as the co-occurrence strength between the two key indicators. For any key indicator, other key indicators whose co-occurrence intensity exceeds a preset co-occurrence intensity threshold are used as related indicators of the key indicator, and a list of related indicators of the key indicator is constructed. Divide the design file into paragraph units and extract the keyword set for each paragraph unit; For any paragraph unit, its keyword set is matched with each key indicator, the indicators to be checked in the design file are determined from the key indicators, and the candidate paragraphs corresponding to the indicators to be checked are located using the list of related indicators. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph, and construct a hierarchical tree to represent the semantic hierarchy of the paragraph based on the semantic level value, specifically including: Obtain a pre-built vocabulary library, which includes at least three words and the semantic level value corresponding to each word. The semantic level value is used to quantify the semantic generalization degree of the words. For any candidate paragraph corresponding to any indicator to be verified, obtain the semantic level value of each keyword in the paragraph from the thesaurus; Based on the semantic level values ​​of each keyword, the keyword set of the candidate paragraph is clustered from bottom to top to construct the hierarchical tree of the candidate paragraph. Calculate the structural similarity between the hierarchical trees of different candidate paragraphs, and filter out paragraph pairs with similar structures; Match keywords in any pair of paragraphs with similar structures, and filter out candidate keywords that correspond to the indicators to be verified based on the matching results; The candidate keywords are compared and verified with the preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified. The verification information is extracted from the paragraphs containing candidate keywords whose semantic relevance has been verified, and the verification information is compared with the verification requirements of the indicators to be verified in the specification document to generate verification results.

2. The method for accurate information comparison based on intelligent verification of project documents according to claim 1, characterized in that, For any paragraph unit, its keyword set is matched with each key indicator to determine the indicators to be verified in the design document from the key indicators, and the candidate paragraphs corresponding to the indicators to be verified are located using the list of associated indicators. Specifically, this includes: For any given paragraph unit, match its keyword set with each key metric; If the keyword set includes a key indicator, and the paragraph unit includes verification information for that key indicator, then the key indicator is determined to be a non-verification indicator. If the keyword set does not include a key indicator, or if the keyword set includes a key indicator but the paragraph unit does not include verification information for that key indicator, then the key indicator is identified as an indicator to be verified. For any indicator to be verified, each associated indicator in its associated indicator list is matched with the keyword set of each paragraph unit. The paragraph unit containing the successfully matched associated indicator is determined as the candidate paragraph for that indicator to be verified.

3. The method for accurate information comparison based on intelligent verification of project documents according to claim 1, characterized in that, The acquisition of the pre-built vocabulary specifically includes: The text in standard documents, qualified design documents, and related field documents is cleaned, segmented into words, and divided into sentences to obtain a corpus; Each corpus in the corpus is processed using a word embedding model to obtain word vectors for each corpus; Obtain at least three clustering density parameters. For each clustering density parameter, perform density clustering on all word vectors based on that clustering density parameter to obtain the clusters under that clustering density parameter. For each cluster under each cluster density parameter, calculate the average of all word vectors in the cluster to obtain the cluster center vector of that cluster; For each corpus under each cluster density parameter, calculate the distance between the word vector of the corpus and the cluster center vectors of all clusters under that density, convert the distance into similarity, and normalize the similarity to obtain the membership degree of the corpus to each cluster. For each corpus under each density parameter, record its maximum membership degree and second maximum membership degree, and calculate the difference between the two; For each corpus, the difference sequence under different density parameters is obtained to form the difference change curve of the corpus; The difference change curve for each corpus is fitted, and the slope of the trend representing the decrease of the difference as the density parameter increases is calculated. Sort the absolute values ​​of the trend slopes of all corpora to determine the maximum and minimum trend slopes; For each corpus, the difference between the absolute value of its trend slope and the minimum value of its trend slope is divided by the difference between the maximum value of its trend slope and the minimum value of its trend slope to calculate the normalized semantic level value. Each corpus and its corresponding semantic level value are stored to construct a vocabulary database.

4. The method for accurate information comparison based on intelligent verification of project documents according to claim 1, characterized in that, The step of comparing and verifying candidate keywords with a preset knowledge base to determine the semantic correlation between the candidate keywords and the indicators to be verified specifically includes: Obtain a preset knowledge base, which includes standard names, synonyms, abbreviations, and definitions of domain terms; For each candidate keyword, retrieve the corresponding standard term from the preset knowledge base; If the retrieved standard terminology matches the name of the indicator to be verified, it is determined that there is a semantic relationship between the candidate keyword and the indicator to be verified. If the standard term corresponding to the candidate keyword is not found in the preset knowledge base, the semantic similarity between the candidate keyword and the name of the indicator to be verified is calculated. If the semantic similarity exceeds the preset semantic similarity threshold, it is determined that there is a semantic association between the candidate keyword and the indicator to be verified.

5. The method for accurate information comparison based on intelligent verification of project documents according to claim 1, characterized in that, The process involves extracting verification information from the paragraphs containing candidate keywords whose semantic relevance has been verified, comparing this verification information with the verification requirements of the indicators to be verified in the specification document, and generating verification results. Specifically, this includes: For each candidate keyword that passes the semantic relevance verification, locate the sentence containing the candidate keyword and at least one sentence before and after it, as the scope of information extraction; Extract numerical values ​​and unit combinations from the information extraction range to serve as the verification information corresponding to the candidate keyword; The extracted verification information was converted to units according to the unit requirements of the indicators to be verified in the standard documents; Obtain the numerical requirements and modifiers of the indicator to be verified in the specification document, and compare the converted verification information with the numerical requirements and modifiers; Based on the comparison results, the verification result of the indicator to be verified is determined, including whether the verification is passed, failed, or the information is missing.

6. A precise information comparison system based on intelligent verification of project documents, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for accurate information comparison based on intelligent verification of project files as described in any one of claims 1-5.

Images