Construction field long text intelligent analysis and plan generation method based on semantic lens

Abstract

This application provides a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, belonging to the field of construction plan generation technology. The method includes: S1: using the fast layer and fine layer of semantic lenses to extract semantic features from the construction text, perform coarse screening, fine screening, and sealing, and obtain multiple semantic units after screening and semantic merging; S2: using semantic lattices to logically associate the scattered semantic units, and integrating key constraints, process information, resource and environmental conditions from different semantic units into a unified structure; S3: using a multi-dimensional constraint fusion algorithm to transform the nodes of the semantic lattice and their logical associations into an executable construction plan. This application, through the use of semantic lens model, lattice structure, and constraint fusion algorithm, eliminates irrelevant information, avoids computational overload caused by full input, and improves the model inference speed; in addition, it automatically associates and comprehensively analyzes scattered constraints to improve the logical consistency of the construction plan.

Description

Technical Field

[0001] This application relates to the field of construction plan generation technology, and in particular to a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lens. Background Technology

[0002] A construction plan is a systematic arrangement and planning of the entire construction process of a project, covering multiple aspects such as schedule, resource allocation, quality control, and safety management. The quality of the construction plan plays a crucial role in the orderly implementation of the project.

[0003] With the advancement of digitalization and intelligentization in the construction industry, construction project management is increasingly relying on big data and artificial intelligence technologies, especially large language models, which demonstrate powerful information processing capabilities in scenarios such as construction planning, contract parsing, and resource scheduling. In addition, knowledge graphs and multi-dimensional constraint solving algorithms are also commonly used technologies in construction project management. However, in practice, some application pain points still exist with these intelligent technologies: Traditional large language models typically employ a full input approach when processing long texts, using internal attention mechanisms for filtering and summarizing. However, documents in the construction field often exceed hundreds of pages, and full input can easily overload the model's computational resources and prolong response time. Furthermore, due to the presence of a large amount of non-core information in the text, the accuracy of extracting key parameters is not high.

[0004] Existing intelligent construction methods mostly rely on knowledge graphs to construct relationship networks of elements such as processes, resources, time, and environment for constraint reasoning and plan generation. However, knowledge graph construction is costly, requires extensive annotation and rule maintenance, and has limited ability to automatically associate scattered information in unstructured text, making it difficult to efficiently capture cross-segment logical relationships in long text parsing.

[0005] Some studies use constraint-solving algorithms (such as linear programming and scheduling optimization) combined with text parsing to generate construction plans, but these usually rely on structured input and are not capable of processing long, scattered text information. They also lack the ability to dynamically associate different constraints, which can easily lead to logically inconsistent or practically unexecutable construction plans.

[0006] In summary, there are still significant challenges in the efficient parsing of long texts, the extraction of key constraints, and the generation of logically consistent construction plans in the construction field. There is an urgent need to propose a new method that can reduce computational power consumption while ensuring information accuracy and logical consistency. Summary of the Invention

[0007] In view of this, the purpose of this application is to provide a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lens. It can eliminate irrelevant information by using semantic lens model, lattice structure and constraint fusion algorithm, avoid computational overload caused by full input, improve the model inference speed and improve the accuracy of feature extraction. In addition, it can automatically associate and comprehensively analyze scattered constraints to improve the logical consistency of construction plan.

[0008] Firstly, this application provides a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, the method comprising: S1: Semantic lens is used for hierarchical text filtering, including: S11: Preprocessing the construction document to extract multiple text blocks; S12: Encoding the text blocks and outputting their semantic vectors; S13: Calculating information entropy based on the frequency of word segmentation and character occurrence in the text blocks, and filtering the first candidate text blocks according to the first quantile; S14: Constructing the retention score of the first candidate text blocks and filtering the second candidate text blocks according to the second quantile; S15: Applying term weighting to the retention scores and setting a mandatory retention mechanism to update the second candidate text blocks; S16: Merging similar second candidate text blocks that meet the merging rules to obtain multiple semantic units. S2: Semantic units are associated and integrated using a semantic lattice, including: S21: Semantic units are set as nodes, and the structured attributes of the nodes are extracted from the semantic vectors; S22: A candidate edge set is generated using heuristic rules; S23: The semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge are calculated to obtain the feature vector of the candidate edge; S24: The dependency strength value and corresponding confidence score of the candidate edge are calculated based on the feature vector, and the candidate edge is filtered based on the dependency strength value; S25: A priority determiner checks for priority conflicts of nodes; S26: Resource conflicts and quantity / specification conflicts of nodes are checked; S27: Highly interconnected node clusters are locally merged. S3: Through a multi-dimensional constraint fusion algorithm, the nodes of the semantic lattice and their logical associations are transformed into an executable construction plan.

[0009] In one possible implementation, S13: calculating information entropy based on the frequency of word segmentation and character occurrence in the text block, and filtering the first candidate text block according to the first quantile, includes: The fast layer computation of the semantic lens model is based on the entropy of the word segmentation probability distribution:

[0010] Where, n i Indicates the number of words in text block i; p ikThis represents the probability that the k-th word in text block i appears in text block i; Calculate the entropy based on the character frequency distribution:

[0011] Among them, C i This represents the character set in text block i; q ic This indicates the frequency of character 'c' in text block i; Pick and Weighted combination:

[0012] in, ∈[0,1], where ∈ is the weight coefficient; Information entropy of text block i Based on the first quantile, the first candidate text block is coarsely screened out from multiple text blocks.

[0013] In one possible implementation, S14: constructing the retention score of the first candidate text block and filtering the second candidate text block according to the second quantile includes: The fine-layer construction of the semantic lens model retains the following scores:

[0014] Here, 'u' refers to the first candidate text block, and 'q' is generated by encoding core terms and task objective keywords in the construction field, representing the core theme of the task. The cosine similarity between the semantic vector of u and q is represented. The topic relevance of u and q to the task is represented by the output of a small classifier. Represents the entropy value of u. Normalization process, Cosine similarity weights Weighting based on relevance to the theme. Weights for normalized entropy values; Retention score of the first candidate text block u The second candidate text block is selected from multiple first candidate text blocks according to the second quantile.

[0015] In one possible implementation, S15: weighting the retention scores with terms and setting a mandatory retention mechanism, and updating the second candidate text block, includes: A weighted scoring formula is used to weight the retention scores of the first candidate text block containing construction-related terms. The weighted scoring formula is as follows:

[0016] The semantic similarity output by the embedding model is equal to the cosine similarity. ; This indicates the proportion of terms that are hit in u out of the total number of terms in the terminology database; Indicates the weighting coefficient of the term; The first candidate text block that meets the mandatory retention rule in terms of table type, keywords, and file area is determined as the second candidate text block.

[0017] In one possible implementation, S16: merging similar second candidate text blocks that satisfy the merging rules to obtain multiple semantic units, including: Two second candidate text blocks with the same process name, a page number distance less than the first threshold, and a semantic vector cosine similarity greater than the second threshold are identified as text pairs to be merged. The text content of the text pairs to be merged is concatenated in the original document order, while retaining source references, to achieve the merging of text content; If the relative deviation coefficient of the numerical fields of the text pair to be merged is less than the third threshold, the average value is taken as the merged value to achieve the merging of the numerical fields; otherwise, it is marked for manual review. The formula for calculating the relative deviation coefficient is:

[0018] σ(x) is the average of the numerical fields of the text pairs to be merged, and σ(x) is the standard deviation.

[0019] In one possible implementation, S23: calculating the semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge to obtain the feature vector of the candidate edge, including: use Calculate the semantic similarity. and They represent and semantic vector; use Calculate the page number spacing. The page spacing scale constant, and They represent and Page number of the page; use Calculate the numerical ratio. and They represent and Numerical attributes; use Calculate the resource overlap. and They represent and A collection of resources; use Calculate the overlap of the time windows. and They represent and The time window interval; Obtain candidate edges Feature vectors:

[0020] in, and Representing candidate edges respectively Two nodes connected.

[0021] In one possible implementation, step S24: calculating the dependency strength value and corresponding confidence score of candidate edges based on the feature vector, and filtering candidate edges based on the dependency strength value, includes: eigenvectors Mapping to dependency strength value :

[0022] σ is the sigmoid function. It is the model's weight vector. It is the bias term of the model; Calculate the credibility score :

[0023]

[0024] yes and The probability that a dependency exists is equal to ; Candidate edges whose dependency strength value is greater than the fourth threshold are retained.

[0025] In one possible implementation, S25: the priority determiner checks for priority conflicts among nodes, including: If node and nodes There is a priority conflict; calculate the nodes separately. and nodes Weighted sum of indicators :

[0026] For edge weight, i.e. edge The dependence strength value, For credibility score, For node importance, , , They represent , , Weighting coefficients; node and nodes If the weighted sum of the indicators of any node is greater than the fifth threshold, the node with the larger weighted sum will be retained; otherwise, it will be marked for manual review.

[0027] In one possible implementation, S27: performing local merging of highly interconnected clusters of nodes includes: use Calculate the attribute vectors of f nodes a within the node cluster. Comprehensive attribute vector , As weight, Reuse node importance of node a ; If there are conflicts among nodes within a node cluster, Conversion: Define the degree of conflict , This represents the value of node a. This represents the average value of the f nodes within a node cluster; Define the conversion factor , It is the sensitivity constant; Weights after conversion ; Converted composite attribute vector .

[0028] Secondly, this application provides a device for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, the device comprising: The first module is used to perform hierarchical filtering of text using semantic lenses; The second module is used to associate and integrate semantic units using a semantic lattice; The third module is used to transform the nodes of the semantic lattice and their logical associations into an executable construction plan through a multi-dimensional constraint fusion algorithm.

[0029] The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lens provided in this application adopts a semantic lens model, lattice structure, and constraint fusion algorithm to efficiently process long texts, avoid computational overload caused by full input, and improve the model inference speed. This solution can accurately extract key constraint information, automatically remove irrelevant information, and improve the accuracy of parameter extraction. This solution can ensure logical consistency, automatically associate and comprehensively analyze scattered constraints, and generate executable and reasonable construction plans.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The flowchart illustrates a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, as provided in an embodiment of this application. Figure 2 This document illustrates a flowchart of a text hierarchical filtering method using semantic lenses, as provided in an embodiment of this application. Figure 3 This document illustrates a flowchart illustrating the association and integration of semantic units using a semantic lattice, as provided in an embodiment of this application. Figure 4 This illustration shows a structural schematic diagram of a long text intelligent parsing and plan generation device based on semantic lens in the construction field, provided in an embodiment of this application. Figure 5 A schematic diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] Existing long text processing technologies in the construction field generally suffer from the following problems: 1. Computational Overload Issue: Current large language models typically use a "full input - internal model filtering" approach when processing hundreds of pages of documents such as construction contracts, specifications, and drawing attachments. This method leads to excessive consumption of model computational resources, a significant decrease in inference speed, and affects the efficiency of construction plan generation.

[0035] 2. Information Interference Issue: Construction documents contain a large amount of non-target information, such as general legal clauses, background information, or descriptions of non-critical procedures. This redundant information interferes with parameter extraction and decision-making in the model, resulting in low accuracy in extracting critical construction constraints and incomplete constraint information.

[0036] 3. Lack of Logical Coherence: Key constraints in construction plans, such as the sequence of procedures, resource allocation, environmental conditions, and schedule requirements, are often scattered across different chapters or documents. Existing methods lack effective mechanisms for automatically correlating and logically integrating this scattered information, leading to logical contradictions or unenforceable situations in the generated construction plans, such as failing to consider the extended concrete curing period caused by winter construction.

[0037] To effectively address the aforementioned issues, this application provides a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses. This method can eliminate irrelevant information while maintaining information integrity, avoid computational overload caused by full input, improve model inference speed, and enhance the accuracy of feature extraction. Furthermore, it automatically associates and comprehensively analyzes dispersed constraints to improve the logical consistency of construction plans.

[0038] Example 1: See Figure 1 As shown, Figure 1 A flowchart of a method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, provided in this application embodiment, is included in the following steps: S1: Use semantic lenses to filter text in layers; S2: Use a semantic lattice to associate and integrate semantic units; S3: Through a multi-dimensional constraint fusion algorithm, the nodes of the semantic lattice and their logical associations are transformed into an executable construction plan.

[0039] See Figure 2 As shown, Figure 2 A flowchart for hierarchical text filtering using semantic lenses, provided in this application embodiment, includes the following steps: S11: Perform text preprocessing on the construction documents and extract multiple text blocks; S12: Encode the text block and output the semantic vector of the text block; S13: Calculate the information entropy based on the frequency of word segmentation and character occurrence in the text block, and filter the first candidate text block according to the first quantile; S14: Construct the retention score of the first candidate text block, and filter the second candidate text block according to the second quantile; S15: Apply terminological weighting to the retention scores and set a mandatory retention mechanism, then update the second candidate text block; S16: Merge similar second candidate text blocks that satisfy the merging rules to obtain multiple semantic units.

[0040] See Figure 3 As shown, Figure 3 A flowchart for associating and integrating semantic units using a semantic lattice, provided in an embodiment of this application, includes the following steps: S21: Set the semantic unit as a node and extract the structured attributes of the node from the semantic vector; S22: Generate a set of candidate edges using heuristic rules; S23: Calculate the semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge to obtain the feature vector of the candidate edge; S24: Calculate the dependency strength value and corresponding confidence score of the candidate edges based on the feature vector, and filter the candidate edges based on the dependency strength value; S25: The priority determiner checks for priority conflicts between nodes; S26: Investigate resource conflicts and quantity / specification conflicts of nodes; S27: Perform local merging of highly interconnected clusters of nodes.

[0041] The following section details the specific processing steps of the long text intelligent parsing and plan generation method based on semantic lenses in the construction field.

[0042] S1: Use semantic lenses to filter text in layers.

[0043] 1. Objectives and Principles The purpose of step S1 is to quickly identify semantic information that is highly relevant to the construction task in long texts, remove redundant or interfering content, reduce the input scale of large models, and at the same time ensure the integrity of key constraint information.

[0044] The concept of "semantic lens" refers to filtering and amplifying textual information at different scales through hierarchical focusing and semantic enhancement, so that the model only processes the most relevant and compact content.

[0045] 2. Semantic Lens Model Construction "Semantic lens" is a multi-level filtering mechanism that includes steps such as semantic focusing (embedding calculation), multi-scale filtering, semantic enhancement and encapsulation.

[0046] S11: Perform text preprocessing on the construction documents and extract multiple text blocks.

[0047] Text preprocessing: Text extraction is performed on construction documents (contracts, construction requirements, technical specifications, and drawing attachments). Optical Character Recognition (OCR) is used to recognize non-text image content, converting tables and diagrams into text descriptions. Text standardization is then performed: headers, footers, tables of contents, standardized units, normalized process and material terminology, etc., are removed. Finally, multiple text blocks (paragraphs, table descriptions, and drawing annotations) are extracted from the construction documents.

[0048] Fast layer (for coarse screening): Lazy loading of lightweight embeddings (such as Tiny Sentence Transformer), high throughput, low latency, used for initial screening of thousands of text blocks. This layer outputs semantic vectors and information entropy values ​​(based on word segmentation diversity or character distribution entropy).

[0049] S12: Encodes the text block and outputs the semantic vector of the text block.

[0050] Each text block (paragraph, table description, or drawing annotation) is encoded by the embedding model (the feature extraction layer in the semantic lens model), and outputs a fixed-length semantic vector, denoted as: , where v i Let d represent the semantic vector of the i-th text block, where d represents the vector dimension (usually 384, 768, or 1024, depending on the model structure), and the vector component vij reflects the representation strength of the text block on the j-th semantic feature dimension.

[0051] semantic vector v i It has the following information structure: Thematic dimension information: reflects the thematic semantics of construction tasks, resources, constraints, etc.; Relationship dimension information: reflects the similarity or logical dependency with adjacent semantic blocks; Register characteristics: Reflecting vocabulary specific to the construction field (such as concrete grade, curing cycle, etc.).

[0052] S13: Calculate the information entropy based on the frequency of word segmentation and character occurrence in the text block, and filter the first candidate text block according to the first quantile.

[0053] Information entropy is used to measure the information density and content complexity of a text block.

[0054] ① Entropy based on word segmentation probability distribution:

[0055] n i Indicates the number of words in text block i; p ik This represents the probability that the k-th word in text block i appears in text block i; Entropy The larger the entropy value, the more dispersed and semantically diverse the information contained in text block i; The lower the value, the more focused and less redundant the text block i is, making it suitable for priority entry into the lens screening stage.

[0056] ② Entropy based on character frequency distribution:

[0057] C i This represents the character set in text block i; q ic This indicates the frequency of character 'c' in text block i; It can detect the "noise level" of text extracted by OCR and filter out formatted, repetitive, or garbled content. The smaller the value, the higher the frequency of a single character. For example, if it consists entirely of repeated characters or garbled text, the character distribution is extremely uneven, the noise level is high, and it is not suitable for fine screening. The larger the size, the more scattered the character distribution, potentially resulting in messy and meaningless text, making it unsuitable for fine-tuning. Therefore, When the frequency of characters is within a reasonable range, there is no large amount of repetition or garbled text, and the text content is relatively concentrated, indicating high text quality.

[0058] In implementation, the system can take a weighted combination of two entropies:

[0059] in, ∈[0,1] is the weighting coefficient, which is usually taken as 0.7.

[0060] Weighted entropy Judging the merits of different sizes: There is no absolute "bigger is better" or "smaller is better." From the perspective of thematic focus, The size should not be too large, as an excessively large size indicates that the word segmentation / character distribution is too scattered, resulting in cluttered text information, unclear themes, and excessive redundant or distracting content; from a text quality perspective: The value should not be too small. If it is too small, it means that the word / character distribution is too concentrated, which is likely to be duplicate content, formatted text, or OCR garbled characters.

[0061] The actual filtering logic is: entropy value Its core function is to help determine the quality and priority of text blocks, and ultimately whether they should proceed to fine screening. The text needs to be within a moderate range. The retention score is calculated by combining semantic similarity, term hit rate, etc. Finally, the first candidate text block is selected according to the quantile. The selection ratio is determined by the quantile, rather than a fixed threshold.

[0062] Fine-grained layer (for secondary screening and sealing): The first candidate text blocks retained by the fast layer are processed using high-precision embedding (such as Siamese-BERT-large or industry fine-tuned models) and combined with term weighting (multiplying the vector by the term weight or weighting the score).

[0063] S14: Construct the retention score of the first candidate text block, and filter the second candidate text block according to the second quantile.

[0064] Constructing the retention fraction:

[0065] Here, 'u' refers to the first candidate text block retained by the fast layer, which is the paragraph that enters the fine layer after being coarsely screened.

[0066] 'q' is generated using a high-precision embedding model, encoding core construction terminology and task objective keywords (such as "construction of the main structure of a 10-story frame residential building," "concrete pouring," and "formwork erection"). It represents the core theme that this analysis task needs to focus on. 'q' serves as the "anchor point" for the entire selection process, used to measure whether candidate paragraph 'u' is relevant to the core task. The higher the similarity between 'u' and 'q', the closer the semantic content of candidate paragraph 'u' is to the core task theme, resulting in a higher score and making it more likely to be retained.

[0067] The cosine similarity between the semantic vector of u and the vector representation of q is expressed as follows.

[0068] This represents the task-based topic relevance between paragraph u and q, and can be output by a small classifier. Specifically, the semantic vector of paragraph u and the term hit information are input into the small classifier, and the small classifier outputs... The TS value is a continuous score between 0 and 1. The higher the TS value, the stronger the adaptability of paragraph u in the construction field and the more valuable the information it contains.

[0069] : Entropy value of paragraph u Normalization processing (when i=u), right Subtraction was performed. The smaller the value, the more focused the text's theme and the higher its quality, which is more conducive to the final result. The greater the positive contribution to the score.

[0070] for , , It is necessary to consider the characteristics of the construction field, and the recommended reference values ​​are as follows: Cosine similarity weight The value range is 0.4~0.5, which is the most critical indicator and prioritizes semantic relevance. Theme relevance weight The value range is 0.3~0.4, which enhances domain adaptability and avoids paragraphs that are semantically similar but have no practical value; Normalized entropy weights Values ​​range from 0.1 to 0.2, used to adjust the impact of text quality, with a relatively low weight.

[0071] The retention ratio is determined by quantiles (p) rather than a fixed threshold: coarse screening retains the upper p1 quantile (e.g., 0.55), and fine screening retains the upper p2 quantile (e.g., 0.9). Quantiles can be optimized through small-sample cross-validation.

[0072] S15: Term-weighted evaluation of the retention scores and mandatory retention mechanism are set, and the second candidate text block is updated.

[0073] S151: Terminology weighting is a dynamic scoring weighting mechanism in the semantic lensing fine-tuning stage, used to highlight the importance of construction-related technical terms. Its essence is to weight similarity scores... Domain-weighted adjustments are performed. The goal of the keyword / term enhancement mechanism is to increase the score weight of paragraphs containing key terms in semantic similarity calculation.

[0074] Establish a terminology database for the construction field (including specifications, material grades, process names, and legal / standard keywords). Additional weighting is applied to prevent paragraphs with low similarity but containing key terms from being mistakenly deleted.

[0075] Weighted score formula:

[0076] This represents semantic similarity calculated by the embedding model, such as cosine similarity. .

[0077] This indicates the proportion of terms that are hit in paragraph u out of the total number of terms in the terminology database.

[0078] This represents the term weighting coefficient, typically ranging from 0.2 to 0.5, used to adjust the degree of influence of construction-related terms on the similarity score.

[0079] If paragraph u contains segmented words from a construction-related terminology database, in the calculation... When Replace with In this way, even if the semantic similarity is low, the paragraph can be retained as long as it contains key terms (such as "concrete strength C30" and "formwork support").

[0080] S152: Establish a "whitelist" for explicitly defined tables (bill of materials, quantity tables) to enforce their retention. The "whitelist" is a mandatory retention mechanism used to prevent algorithms from mistakenly deleting critical data structures. Specifically, it includes: Table type whitelist: Tables containing structured data such as "Bill of Materials" and "Summary Table of Quantities"; Keyword whitelist: such as "budget sheet", "acceptance record", "safety inspection form"; Document area whitelist: Special areas detected by OCR or layout analysis.

[0081] Regarding "special areas," for example: ① The marked area in the drawing attachments.

[0082] Scenario: In the corner or side of the construction drawings, there are technical parameters marked such as "Elevation ±0.000" and "Seismic resistance level II".

[0083] Reason: The text in these areas is often fragmented and lacks context, but it contains core constraints for construction and must be preserved.

[0084] ② Supplementary Agreement section at the end of the contract.

[0085] Scenario: The "Supplementary Clauses" or "Special Agreements between the Parties" section on the last page of the construction contract, which includes key constraints such as "Winter Construction Anti-freezing Requirements".

[0086] Reason: The text format in these areas is special, and the semantic similarity may not be high, but it directly affects the compliance of the construction plan.

[0087] ③ The notes area around the table.

[0088] Scenario: Small print notes below the "Bill of Materials" table, such as "The steel bars must meet the HRB400E seismic requirements".

[0089] Reason: The text in the remarks area is attached to the table. Although it is not the main body of the table, it supplements the key technical requirements.

[0090] Once the first candidate text block is identified as a whitelist item, it is directly retained into the parsing result set even if its semantic score is low or it lacks contextual matching.

[0091] Steps S13 and S14 are used to calculate the dynamic retention score of the first candidate text block, and the second candidate text block is selected by using terminological weighting of the dynamic retention score and forced retention by the "whitelist".

[0092] S16: Merge similar second candidate text blocks that satisfy the merging rules to obtain multiple semantic units.

[0093] Sealing layer screening is a second-stage candidate segment retained after the fine screening stage, merging similar segments into unified semantic units. It does not merge all similar segments, but only those that meet specific conditions, aiming to provide more structured and logically consistent input for subsequent construction plan generation.

[0094] Semantic unit merging rules: simultaneously satisfy ① same process name; ② page number distance less than the threshold; ③ high topic similarity.

[0095] After the second candidate text blocks are merged into semantic units, the source citation list of the semantic units (i.e. the sources of the multiple second candidate text blocks that make up the semantic units) is retained, and the credibility of the merged values ​​is verified (if the numerical conflict is >10%, it is marked for manual verification).

[0096] A multi-label classifier is used to label each finely screened paragraph with a process label, such as "concrete pouring" or "formwork erection", and then it is determined whether any two second candidate text blocks have the same process name.

[0097] Topic similarity is used to determine whether two paragraphs or tables belong to the same semantic topic. It can be achieved using vector cosine similarity or KL divergence of topic distributions. (Paragraph) With paragraph The cosine similarity of the semantic vectors is:

[0098] like If θ > 0.85, then the paragraph is considered to be... With paragraph The themes are similar and can be merged.

[0099] Specific methods of merging: Text content: Concatenate paragraphs within the same group in the original document order, retaining source citations (e.g., "from page 5, paragraph 2, page 6, paragraph 1"). Numerical fields: Extract all values ​​within the same group (such as concrete volume and steel reinforcement weight), and proceed to the numerical verification stage. The merged numerical fields need to be verified for their rationality and consistency.

[0100] Let the set of values ​​for the same field before merging be . ; The average value of all elements in the set is ; The credibility index can be defined as the relative deviation coefficient: σ(x) is the standard deviation.

[0101] Using the relative deviation coefficient To indicate the degree of conflict, if the relative coefficient of deviation exceeds 10%, a value is not directly taken, but marked as "awaiting manual verification," and all original values ​​are retained for manual confirmation. If the relative coefficient of deviation is ≤10%, the average value is used as the merged value, or the value with the highest frequency is retained.

[0102] The processing flow of steps S11-S15 is illustrated with an example, taking the "construction of the main structure of a 10-story frame residential building" as an example: Original document: Contract + Construction Requirements, 500 pages.

[0103] Coarse lens screening: Remove 200 pages of paragraphs that are not related to construction procedures (such as acceptance standards and final settlement clauses).

[0104] Fine-tuning: From the remaining 300 pages, 80 more pages were selected that relate to the construction of the main structure.

[0105] Sealing lens treatment: "Total concrete pouring volume 1550m³" 3 ", ..." and "80 tons of steel bars, ..." are encapsulated as independent semantic units that can be directly used in the model.

[0106] S2: Use semantic lattice to associate and integrate semantic units.

[0107] The purpose of step S2 is to logically associate the scattered semantic units selected by the semantic lens in step S1, and to integrate the key constraints, process information, resources and environmental conditions of different paragraphs or files into a unified structure, so as to provide a logically consistent and reasonable basis for the generation of construction plans.

[0108] The concept of "semantic lattice" refers to establishing a hierarchical and reasonable network of associations for dispersed semantic information through a node-edge structure, enabling the automatic detection and aggregation of hidden logical constraints.

[0109] S21: Set the semantic unit as a node and extract the structured attributes of the node from the semantic vector.

[0110] Node attribute engineering: Treat each semantic unit as a node, extract structured attributes for each semantic unit (paragraph / merged block), and extract structured attributes from the semantic vector of the semantic unit, including but not limited to: ① Process label (output of multi-label classifier): The multi-label classifier will label each semantic unit after fine screening with a process label, such as "concrete pouring" or "formwork erection".

[0111] ② Numerical attributes (duration, quantity, level).

[0112] ③ Resource demand vector ({people:2, pump truck:1,...}).

[0113] ④ Time window (optional; if the document does not specify a time, set it to an open window and represent it with a confidence interval). The time window refers to the time range or time constraint that a certain process can be executed. It is the core basis for time scheduling when generating the construction plan.

[0114] ⑤ Environmental constraint markings (such as "No pouring at low temperatures").

[0115] S22: Generate a set of candidate edges using heuristic rules.

[0116] The purpose of candidate edge generation is to reduce the computational cost of full comparison between nodes, generating only edges with potential logical relationships, thus significantly reducing the complexity of full comparison.

[0117] Available heuristic rules include those shown in Table 1: Table 1

[0118] ① Terminology overlap rules: for example, nodes Semantic unit extraction terms: {C30 concrete, formwork support, curing}, nodes The semantic unit extracted terms are: {C30 concrete, rebar tying, curing}, intersection: {C30 concrete, curing}, overlap size is 2, smaller set size is 3, overlap rate = 2 / 3 ≈ 67% ≥ 30%, satisfying the term overlap rule, node and nodes There exist candidate edges that satisfy the term overlap rule.

[0119] ② Numerical Proportioning Rules: In the construction field, there are usually stable engineering proportions between processes, resources, and costs, such as the ratio of steel reinforcement usage to concrete volume, the ratio of formwork area to pouring volume, and the ratio of construction period to project quantity. When the numerical attributes (such as material usage and construction period in days) of two nodes fall within a preset range of conventional engineering proportions, it can be quickly determined that they are likely to belong to the same construction process or have upstream and downstream dependencies, thereby generating candidate edges between these two nodes and avoiding the computational overhead of comparing all nodes.

[0120] ③ Resource sharing rule: The cosine similarity of resource vectors is not based on the complete semantic vector in step S1, but on the resource demand vector extracted separately from the node attributes. The resource demand of each node is an independent vector, such as {workers: 2, pump trucks: 1, tower cranes: 1}, and the similarity of the resource vectors is compared.

[0121] S23: Calculate the semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge to obtain the feature vector of the candidate edge.

[0122] Feature vectors: {semantic similarity, page number spacing, numerical ratio, resource overlap, time window overlap}.

[0123] Each candidate edge Corresponding to a feature vector:

[0124] and Representing candidate edges respectively Two nodes connected.

[0125] Each feature has been normalized to the [0,1] interval for use in the subsequent dependency strength function.

[0126] ① Semantic similarity: Semantic lens model from step S1 (cosine similarity of semantic vectors). and They represent and The semantic vector.

[0127] ② Page number spacing: This indicates the degree of physical proximity; the closer the value is to 1, the closer the two are. This is the page spacing scale constant, typically 2. It is a node Page number of the page in question. It is a node The page number of the page it is on. If the node and nodes When documents belong to different documents, their page number spacing is usually considered to be infinite, indicating that they are not physically adjacent.

[0128] ③ Numerical ratio: Nodes connected to candidate edges and nodes Numerical attributes, such as construction period in days and steel reinforcement weight, indicate data consistency when the ratio is close to 1. When nodes... and nodes When there are multiple numerical attributes, a weighted average is used, and the weight is determined by the importance of the attribute in the construction field (e.g., construction work weight 0.6, material weight 0.4).

[0129] ④ Resource overlap: This represents the set of resources or sparse vectors corresponding to a node, such as {worker, crane}, and measures the similarity of resource vectors. When a node... and nodes When multiple resources overlap, a weighted average is used.

[0130] ⑤ Time window overlap: The time window interval representing the node is the time range within which the process can be executed. If it is an open window, it is calculated using the overlap ratio of the confidence intervals. , For the estimated average time point, The time uncertainty is estimated by combining the context and historical project data. and When node and nodes When multiple time windows overlap, a weighted average is used.

[0131] S24: Calculate the dependency strength value and corresponding confidence score of the candidate edges based on the feature vector, and filter the candidate edges based on the dependency strength value.

[0132] During the weakly supervised training phase, if the project has historical manual operation logs or change orders, these manually confirmed changes such as process dependencies and resource adjustments are first accurately mapped to the corresponding candidate edges in the semantic lattice as positive samples. Negative samples are constructed by selecting logically unrelated node pairs or manually marked invalid edges to ensure sample balance. If there is no historical data, a cold start process is initiated, randomly selecting 30 to 100 edges covering different heuristic rules from the candidate edges of the current project. These edges are then labeled as positive or negative samples by personnel with construction domain knowledge based on the standard of "whether they meet the construction logical dependency". Enhancement can be achieved through feature perturbation and sample replacement. Subsequently, new manual change data will be continuously collected as the project progresses to iteratively optimize the sample set, thereby providing a high-quality training foundation for the edge weight prediction model.

[0133] A lightweight model (small Transformer) is used to learn an edge weight prediction function to predict candidate edges. eigenvectors Map the dependency strength values ​​to [0,1] and output the confidence score.

[0134] The edge weight prediction function will use the feature vector Mapping to dependency strength value :

[0135] σ is the sigmoid function. and These are the training weights. It is the model's weight vector, whose dimension is the same as the feature vector. Totally consistent. These are the parameters learned by the model during training, used to assign different importance weights to different dimensions of the feature vector. It is the bias term of the model, a scalar used to adjust the baseline value of linear calculations and improve the model's fitting ability. and Initial values ​​can be set randomly, but only as the starting point for training. In subsequent weakly supervised training, the model will be continuously updated iteratively using algorithms such as gradient descent. and Ultimately, the optimal weights and biases are obtained.

[0136] Dependency strength value Represents a node and nodes The degree of logical or semantic dependency between them can be interpreted as candidate edges. The probability of its existence will Edges smaller than a preset threshold are deleted directly to reduce redundancy and improve the efficiency and accuracy of the semantic lattice.

[0137] The confidence score measures how confident the model is in the prediction. Common methods include using the entropy of the output distribution or the variance of Dropout sampling.

[0138]

[0139]

[0140] It is a node and nodes The probability that a dependency exists, i.e. .

[0141] When output When the entropy is close to 0 or 1 (extreme confidence), the confidence level is high.

[0142] Border rights This is the edge dependency strength value, representing the probability that there is a logical dependency between nodes, while The model is for this The confidence level of the prediction results The larger the value (the closer to 1), the stronger the model's dependence on that factor. The more reliable the prediction, the better.

[0143] S25: Priority determiner checks for priority conflicts in nodes.

[0144] The priority determiner focuses only on the order conflicts of edges in the semantic lattice (i.e., contradictory dependencies such as A→B and B→A). Its core task is to quantify which dependency edge is more credible by calculating the retain suggestion score (RScore) of the conflicting edge, thereby providing a decision basis for which side to retain.

[0145] This part plays a crucial role in the semantic lattice construction process. When contradictory dependencies arise, it automatically selects the more reliable side to retain, ensuring the logical consistency of the entire network and providing a conflict-free foundation for subsequent construction plan generation. Simply put, when the system encounters contradictory rules such as "A must follow B" and "B must follow A," it automatically determines which rule is more reliable and retains the correct one, avoiding logical conflicts.

[0146] The weighted sum of the three key indicators:

[0147] (Edge Weight): Edge The dependency strength value represents the probability that the dependency relationship exists.

[0148] (Confidence Score): Model Pair The confidence level of the prediction result is such that the closer the value is to 1, the more reliable the model's judgment on the dependency.

[0149] (Node Importance): Conflicting Related Nodes The importance is determined by factors such as the number of references to standards and key terms a node contains. For example, this value could be equal to the sum of the number of referenced standards and the number of key construction terms.

[0150] , , Weighting coefficients, set based on experience in the construction field, are used to balance the contributions of the three indicators. They are typically set based on experience in the construction field (e.g., =0.4、 =0.3、 =0.3), used to balance the contributions of the three indicators.

[0151] The conflict referred to here is a contradiction between the semantic lattice and the construction logic and data consistency. The RS score corresponding to conflicting nodes is uncertain. The RS score depends on the dependency strength of the conflicting edges, the model's credibility, and the importance of the node itself. It reflects the credibility of the corresponding dependency relationship, not whether it is "in conflict." In a conflict, one party may have a high RS score and the other a low RS score, or both parties may have high or low RS scores. If the RS scores corresponding to conflicting nodes are both low, it means that the credibility of these two dependencies is insufficient. The system will not force a choice between the two, but will instead place the conflict in a manual review queue for engineers to judge and handle based on the actual construction situation.

[0152] S26: Investigate resource conflicts and quantity / specification conflicts of nodes.

[0153] Breakthrough detection and interpretive output constitute a more comprehensive conflict investigation and handling process, corresponding to two independent conflict types: resource conflicts (the total resource demand of all nodes exceeds the available quantity within the same time period); and quantity / specification conflicts (differences in records of the same material or parameter at different nodes exceed engineering tolerances). This part requires manual review to ultimately ensure the logical consistency of the semantic lattice and the reliability of the data.

[0154] Resource conflict: The total resource demand exceeds the available resources within the same time period.

[0155] Quantity / specification conflict: The discrepancy between two records of the same material exceeds the tolerance.

[0156] Automatic conflict type determination: The system verifies conflicts based on information such as the dependencies, resource requirements, time windows, and numerical attributes of nodes in the semantic lattice, using preset rule logic. For resource conflicts, the system first extracts the resource requirement type and quantity of each node within the same time window, summarizes them, and compares them with the preset upper limit of available resources. If the total requirement of a certain type of resource exceeds the available quantity, it is determined to be a resource conflict. For quantity / specification conflicts, the system locates different nodes describing the same material or parameter, extracts the corresponding values, and calculates the difference ratio. If the difference exceeds the preset engineering tolerance threshold, it is determined to be a quantity / specification conflict.

[0157] Each conflict generates an evidence package: {Conflict type, involved node ID, supported text fragments, numerical differences, suggested actions (by priority)}.

[0158] S27: Perform local merging of highly interconnected clusters of nodes.

[0159] Highly interconnected node clusters are locally merged using a "weighted summation + conflict reduction" strategy, and resource / time attributes are updated (e.g., by taking the union of time windows and adding an uncertainty coefficient). After lattice formation, multiple similar nodes are locally aggregated to form enhanced nodes, improving overall consistency.

[0160] Highly interconnected clusters of nodes are generally identified through two dimensions: Connection density: The ratio of the actual number of edges in a node cluster to the theoretical maximum number of edges. When the ratio exceeds a preset threshold (such as 0.6), it is determined to be highly interconnected.

[0161] Semantic similarity: Cosine similarity is calculated based on the attribute vectors of nodes (such as process name, resource requirements, and time window). When the average similarity of nodes within a cluster exceeds a threshold (such as 0.7), they are considered highly interconnected.

[0162] S271: Aggregation Method: Assume the cluster of nodes to be aggregated is There are a total of f nodes.

[0163] The attribute vector of each node a is These represent information related to time window, resource requirements, duration, and environmental constraints, respectively.

[0164] The attribute vector here It is a structured vector output from the previous "node attribute engineering" step, which includes quantified engineering attributes such as time window, resource requirements, duration, and environmental constraints.

[0165] The comprehensive attribute vector of each node in the aggregate node cluster is: , The weight represents the degree of credibility. The importance of node 'a' can be reused by using node importance. ( =a), because the higher the importance of a node, the more reliable its attribute information usually is.

[0166] If nodes a1 and a2 represent the same process "concrete pouring", the duration of a1 is 2 days and the weight is 0.8; the duration of a2 is 3 days and the weight is 0.6.

[0167] Then the duration after aggregation: sky.

[0168] S272: Conflict Reduction: Numerical or semantic conflicts often occur in aggregations (e.g., duration 2 days vs. 5 days). "Conflict reduction" is used to reduce the interference of inconsistent information on the aggregation results.

[0169] If the descriptions of the same process are contradictory (e.g., "concrete strength C30" and "concrete strength C40"), it is considered a semantic conflict. Determining whether it's a conflict or within the normal error range requires setting a tolerance threshold based on engineering experience. For numerical attributes such as time and quantity, thresholds can be set with reference to industry standards or project requirements (e.g., duration error ≤ 5%, material usage error ≤ 8%). For semantic attributes, preset keyword matching rules can be used. For example, if "the strength grade and construction method of the same process are inconsistent", it will be judged as a conflict; otherwise, it will be regarded as a normal error.

[0170] Define the degree of conflict: ; This represents the value of node a. This represents the average value of f nodes within a node cluster.

[0171] Define the conversion factor: , It is the sensitivity constant (generally taken between 2 and 5). Used to control the intensity of conflict reduction. The larger the value, the more pronounced the weight penalty for nodes with high conflicts.

[0172] The converted weights are: ; Recalculate the weighted sum: ; If a node differs greatly from the group (significant conflict), then its It will approach 0 and be automatically "downgraded" to avoid polluting the overall result.

[0173] S273: Calculation of Uncertainty Coefficient: The larger the difference in weights before and after the conversion, the more internal conflicts there are and the higher the uncertainty. The uncertainty coefficient can be used to adjust the time fluctuation range or generate manual verification marks in the subsequent construction plan generation.

[0174] Uncertainty coefficient needs to be set with a threshold for decision-making in subsequent processes: when the uncertainty coefficient exceeds the threshold (e.g., 0.3), it indicates that there are many conflicts within the node cluster, and the credibility of the generated enhanced node is insufficient. The system will automatically mark it and trigger manual review; when the uncertainty coefficient is below the threshold, the enhanced node can be directly used for subsequent construction plan generation.

[0175] S3: Through a multi-dimensional constraint fusion algorithm, the nodes of the semantic lattice and their logical associations are transformed into an executable construction plan.

[0176] Step S3 aims to transform the semantic lattice nodes and their logical associations constructed in Step 2 into an executable construction plan. The core idea is to integrate constraints such as process sequence, resource allocation, time arrangement, and environmental conditions through a multi-dimensional constraint fusion algorithm to achieve a logically consistent and resource-feasible construction plan output.

[0177] Each node from the semantic lattice is treated as a task (or task set), and constraints such as "time, resources, sequence, environment, and quantified values" are treated as constraints of different dimensions. Joint reasoning is performed through a unified mathematical model so that the final plan satisfies both key (hard) constraints and soft constraints and efficiency goals as much as possible.

[0178] ① Constraint classification and mathematical expression: Constraints are categorized as follows: Hard constraints: contract delivery date, critical structure acceptance sequence, regulatory restrictions, and total amount of critical materials; Soft constraints: equipment continuous operation preference, peak and trough of manpower, and noise avoidance time window.

[0179] ② Hybrid solution framework: The first stage (ILP solver) – ensuring hard constraints: Modeling the hard constraints and critical tasks as integer linear programming (or mixed integer linear programming, MILP), and solving for a feasible region or boundary solution that guarantees the satisfaction of the critical constraints (usually a macroscopic piecewise arrangement of the task). This can be done using open-source solvers (e.g., CBC / CP-SAT) or commercial solutions (CPLEX / Gurobi).

[0180] The second stage (parallel heuristic refinement) – optimizing soft constraints: Within the macro-time window and resource boundaries given by ILP, parallel heuristic algorithms (greedy scheduling + local search / taboo / simulated annealing) are used to perform fine-grained optimization of task sequences and resource allocation to improve indicators such as equipment utilization, shorten idle time, and reduce shift change costs.

[0181] Interaction strategy: If the heuristic cannot find a feasible solution within the boundary, return to the ILP solution space and relax some of the secondary hard constraints (which constraints can be fine-tuned is indicated by the priority learner, by introducing slack variables or expanding the time window / reducing resource requirements; the relaxation objects are selected according to priority, and the penalty is included in the objective to minimize the relaxation cost), and solve again; the number of loops is limited to ensure response time.

[0182] Relax some secondary hard constraints: relax by introducing slack variables or expanding the time window / reducing resource requirements; select relaxation targets by priority and include penalties in the target to minimize the cost of relaxation.

[0183] ③ Robustness and uncertainty modeling: For key parameters (probability of material delivery delay and fluctuation of daily pump truck capacity), interval / scenario modeling is adopted. ILP can choose robust optimization or scenario-based constraint weighted solution to ensure that the solution is still feasible under common disturbance conditions.

[0184] ④ Formal output and review of the plan: The output files include: Gantt chart (daily / hourly granularity), resource usage table, critical path report, conflict and change proposal list, and evidence chain for each project item (source text fragments and confidence levels).

[0185] Provides an API interface for BIM / ERP system integration (input: semantic lattice JSON; output: plan JSON, GANTT, CSV).

[0186] ⑤ Manual looping and online learning: Manual modifications are automatically recorded and written back to the training database, prioritizing updates to the edge-weight model, priority model, and terminology database, forming a closed-loop improvement. The write-back format includes the plans before and after the change, the reasons for the manual modification, and a timestamp.

[0187] Examples are provided to illustrate steps S1, S2, and S3 above: Construction project: Main structure of a 10-story frame residential building.

[0188] Scenario input: The contract and technical specifications total 500 pages, including a list of materials, pump truck specifications, sectional pouring diagrams, and seasonal temperature descriptions.

[0189] Step S1 output: After preprocessing, the text was split into 4000 blocks; after rapid layer filtering, 2400 blocks were retained; after fine filtering and term enhancement, they were merged into 120 semantic units (the semantic units include "total concrete pouring volume 1550m"). 3 "Total steel reinforcement: 80 tons", "Daily pump truck capacity: 200m³" 3 Each semantic unit includes the source page number and confidence level.

[0190] Step S2 output: Based on candidate edge rules and a learner, a lattice G is constructed with 120 initial nodes and approximately 4000 candidate edges. The learner filters out 420 significant edges. The following conflicts are automatically detected: There is resource overlap between A (pouring section 1) and B (pouring section 2), resulting in an over-allocation of pump trucks; The curing period for concrete in a certain location needs to be extended due to seasonal constraints.

[0191] The system generates an evidence package (referencing the original contract page numbers) and makes recommendations for adjustments based on priority (e.g., extending curing time and pouring in stages).

[0192] Step S3 output: The ILP phase solves for key constraints (contract completion date, minimum appointment window for pump truck, and order of acceptance of key nodes) to obtain the macro schedule; The heuristic phase will use 1550m within the macroscopic window. 3 The project is allocated to 8 pouring days (approximately 194m per day). 3 Considering the 200m pump truck 3 (with an upper limit and a buffer), and strictly control the order of steel bar binding and pouring.

[0193] Final output: Segmented Gantt chart (3 construction segments, each ≤52 days), pump truck reservation form, critical path and risk warnings (e.g., an additional 7 days are required for extended winter maintenance). All plan items include source citations and confidence levels; low-confidence items are added to the manual review checklist.

[0194] Example 2: Based on the same inventive concept, this application also provides a semantic lens-based long text intelligent parsing and plan generation device for construction, which corresponds to the semantic lens-based long text intelligent parsing and plan generation method for construction. Since the principle of the device in this application is similar to the semantic lens-based long text intelligent parsing and plan generation method for construction described above, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0195] See Figure 4 As shown, Figure 4 This application provides a schematic diagram of a device for intelligent parsing and planning of long texts in the construction field based on semantic lenses. The device includes: The first module 401 is used to perform hierarchical filtering of text using semantic lenses; The second module 402 is used to associate and integrate semantic units using a semantic lattice; The third module 403 is used to transform the nodes of the semantic lattice and their logical associations into an executable construction plan through a multi-dimensional constraint fusion algorithm.

[0196] In one possible implementation, the first module 401, when employing semantic lenses to perform hierarchical text filtering, includes: Perform text preprocessing on the construction documents to extract multiple text blocks; Encode text blocks and output semantic vectors for the text blocks; The information entropy is calculated based on the frequency of word segmentation and character occurrence in the text block, and the first candidate text block is selected according to the first quantile. Construct the retention score of the first candidate text block, and filter the second candidate text block according to the second quantile; The retention scores are then weighted by terms and a mandatory retention mechanism is set up to update the second candidate text block. Similar second candidate text blocks that meet the merging rules are merged to obtain multiple semantic units.

[0197] In one possible implementation, the second module 402, when associating and integrating semantic units using a semantic lattice, includes: Semantic units are set as nodes, and the structured attributes of the nodes are extracted from the semantic vectors; A set of candidate edges is generated using heuristic rules; Calculate the semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge to obtain the feature vector of the candidate edge; The dependency strength value and corresponding confidence score of the candidate edges are calculated based on the feature vectors, and the candidate edges are filtered based on the dependency strength value. The priority determiner checks for priority conflicts between nodes; Investigate resource conflicts and quantity / specification conflicts of nodes; Perform local merging of highly interconnected node clusters.

[0198] The intelligent parsing and plan generation device for long texts in the construction field based on semantic lens provided in this application can eliminate irrelevant information by using semantic lens model, lattice structure and constraint fusion algorithm, avoid computational overload caused by full input, improve model inference speed and improve the accuracy of feature extraction; in addition, it can automatically associate and comprehensively analyze scattered constraints to improve the logical consistency of construction plan.

[0199] See Figure 5 As shown, Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 500 includes a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 communicates with the memory 502 through the bus 503. The processor 501 executes the machine-readable instructions to perform the steps of the above-described method for intelligent parsing and planning of long texts in the construction field based on semantic lenses.

[0200] Specifically, the memory 502 and processor 501 mentioned above can be general-purpose memory and processor, without any specific limitations. When the processor 501 runs the computer program stored in the memory 502, it can execute the above-mentioned method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses.

[0201] Corresponding to the above-mentioned method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, this application embodiment also provides a computer-readable storage medium storing a computer program, which, when run by a processor, executes the steps of the above-mentioned method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses.

[0202] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, characterized in that, The method includes: S1: Semantic lens is used for hierarchical text filtering, including: S11: Preprocessing the construction document to extract multiple text blocks; S12: Encoding the text blocks and outputting their semantic vectors; S13: Calculating information entropy based on the frequency of word segmentation and character occurrence in the text blocks, and filtering the first candidate text blocks according to the first quantile; S14: Constructing the retention score of the first candidate text blocks and filtering the second candidate text blocks according to the second quantile; S15: Applying term weighting to the retention scores and setting a mandatory retention mechanism to update the second candidate text blocks; S16: Merging similar second candidate text blocks that meet the merging rules to obtain multiple semantic units. S2: Semantic units are associated and integrated using a semantic lattice, including: S21: Semantic units are set as nodes, and the structured attributes of the nodes are extracted from the semantic vectors; S22: A candidate edge set is generated using heuristic rules; S23: The semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge are calculated to obtain the feature vector of the candidate edge; S24: The dependency strength value and corresponding confidence score of the candidate edge are calculated based on the feature vector, and the candidate edge is filtered based on the dependency strength value; S25: A priority determiner checks for priority conflicts of nodes; S26: Resource conflicts and quantity / specification conflicts of nodes are checked; S27: Highly interconnected node clusters are locally merged. S3: Through a multi-dimensional constraint fusion algorithm, the nodes of the semantic lattice and their logical associations are transformed into an executable construction plan.

2. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 1, characterized in that, S13: Calculate information entropy based on the frequency of word segmentation and character occurrence in the text block, and filter the first candidate text block according to the first quantile, including: The fast layer computation of the semantic lens model is based on the entropy of the word segmentation probability distribution: Where, n i Indicates the number of words in text block i; p ik This represents the probability that the k-th word in text block i appears in text block i; Calculate the entropy based on the character frequency distribution: Among them, C i This represents the character set in text block i; q ic This indicates the frequency of character 'c' in text block i; Pick and Weighted combination: in, ∈[0,1], where ∈ is the weight coefficient; Information entropy of text block i Based on the first quantile, the first candidate text block is coarsely screened out from multiple text blocks.

3. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 1, characterized in that, S14: Constructing the retention score of the first candidate text block and filtering the second candidate text block according to the second quantile includes: The fine-layer construction of the semantic lens model retains the following scores: Here, 'u' refers to the first candidate text block, and 'q' is generated by encoding core terms and task objective keywords in the construction field, representing the core theme of the task. The cosine similarity between the semantic vector of u and q is represented. The topic relevance of u and q to the task is represented by the output of a small classifier. Represents the entropy value of u. Normalization process, Cosine similarity weights Weighting based on relevance to the theme. Weights for normalized entropy values; Retention score of the first candidate text block u The second candidate text block is selected from multiple first candidate text blocks according to the second quantile.

4. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 3, characterized in that, S15: Weight the retention scores using terms and set a mandatory retention mechanism to update the second candidate text block, including: A weighted scoring formula is used to weight the retention scores of the first candidate text block containing construction-related terms. The weighted scoring formula is as follows: The semantic similarity output by the embedding model is equal to the cosine similarity. ; This indicates the proportion of terms that are hit in u out of the total number of terms in the terminology database; Indicates the weighting coefficient of the term; The first candidate text block that meets the mandatory retention rule in terms of table type, keywords, and file area is determined as the second candidate text block.

5. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 1, characterized in that, S16: Merge similar second candidate text blocks that satisfy the merging rules to obtain multiple semantic units, including: Two second candidate text blocks with the same process name, a page number distance less than the first threshold, and a semantic vector cosine similarity greater than the second threshold are identified as text pairs to be merged. The text content of the text pairs to be merged is concatenated in the original document order, while retaining source references, to achieve the merging of text content; If the relative deviation coefficient of the numerical fields of the text pair to be merged is less than the third threshold, the average value is taken as the merged value to achieve the merging of the numerical fields; otherwise, it is marked for manual review. The formula for calculating the relative deviation coefficient is: σ(x) is the average of the numerical fields of the text pairs to be merged, and σ(x) is the standard deviation.

6. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 1, characterized in that, S23: Calculate the semantic similarity, page number spacing, numerical ratio, resource overlap, and time window overlap between the two nodes connected by the candidate edge to obtain the feature vector of the candidate edge, including: use Calculate the semantic similarity. and They represent and semantic vector; use Calculate the page number spacing. The page spacing scale constant, and They represent and Page number of the page; use Calculate the numerical ratio. and They represent and Numerical attributes; use Calculate the resource overlap. and They represent and A collection of resources; use Calculate the overlap of the time windows. and They represent and The time window interval; Obtain candidate edges Feature vectors: in, and Representing candidate edges respectively Two nodes are connected.

7. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 6, characterized in that, S24: Calculate the dependency strength value and corresponding confidence score of candidate edges based on the feature vector, and filter candidate edges based on the dependency strength value, including: eigenvectors Mapping to dependency strength value : σ is the sigmoid function. It is the model's weight vector. It is the bias term of the model; Calculate the credibility score : yes and The probability that a dependency exists is equal to ; Candidate edges whose dependency strength value is greater than the fourth threshold are retained.

8. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 7, characterized in that, S25: The priority determiner checks for priority conflicts among nodes, including: If node and nodes There is a priority conflict; calculate the nodes separately. and nodes Weighted sum of indicators : For edge weight, i.e. edge The dependence strength value, For credibility score, For node importance, , , They represent , , Weighting coefficients; node and nodes If the weighted sum of the indicators of any node is greater than the fifth threshold, then the node with the larger weighted sum of indicators is retained; otherwise, it is marked for manual review.

9. The method for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses according to claim 8, characterized in that, S27: Local merging of highly interconnected node clusters includes: use Calculate the attribute vectors of f nodes a within the node cluster. Comprehensive attribute vector , As weight, Reuse node importance of node a ; If there are conflicts among nodes within a node cluster, Conversion: Define the degree of conflict , This represents the value of node a. This represents the average value of the f nodes within a node cluster; Define the conversion factor , It is the sensitivity constant; Weights after conversion ; Converted composite attribute vector .

10. A device for intelligent parsing and plan generation of long texts in the construction field based on semantic lenses, characterized in that, The device includes: The first module is used to perform hierarchical filtering of text using semantic lenses; The second module is used to associate and integrate semantic units using a semantic lattice; The third module is used to transform the nodes of the semantic lattice and their logical associations into an executable construction plan through a multi-dimensional constraint fusion algorithm.

Images