A city planning checking method and system based on natural language processing

Abstract

The application relates to a city planning verification method and system based on natural language processing, which comprises the following steps: receiving a user request instruction and performing analysis, obtaining verification intention information and target area information; based on the target area information, corresponding area rule sets are retrieved and matched, and a verification task list and a data requirement list are generated; multi-source heterogeneous data is acquired and preprocessed to obtain standardized verification data and historical fusion features; a preset differentiated verification process is executed according to the task type; a verification report is generated based on the differentiated verification process result; in summary, the application analyzes the user request through natural language processing and constructs a planning knowledge base, determines the certainty and fuzziness rule tasks in combination with the differentiated verification process, realizes the cooperation of rule provisions and case features, solves the problems of low automation degree and insufficient processing capacity of fuzzy clauses in the traditional verification method, and has the effects of improving verification efficiency and enhancing the explainability of decision-making suggestions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban planning, in particular to a city planning verification method and system based on natural language processing. BACKGROUND

[0002] In the field of urban planning management, traditional verification work highly depends on manual interpretation of rules and regulations by planning professionals and manual analysis of spatial data. With the increasing complexity of urban development, planning regulations are numerous and complex, containing not only explicit quantitative indicators such as upper limit of floor area ratio, but also a large number of qualitative descriptions such as "harmonious with the surrounding environment".

[0003] For city planning verification, existing technologies usually adopt two types of methods: one is an automatic verification system based on a rule engine, which can efficiently handle structured rules, but cannot understand natural language texts and lacks the ability to handle ambiguous clauses; the other is an auxiliary decision-making system based on case reasoning, which can provide reference cases, but the matching accuracy of cases is insufficient, and the rule verification process is independent of the case matching process, forming a data island. In summary, the above methods have the following defects: first, there is a huge gap between natural language texts and machine executable logic, and a large number of rules cannot be automatically applied, resulting in a lack of automation capability; second, there is a lack of unified representation method for multi-source heterogeneous data, making it difficult to establish semantic association among rules, data and cases; finally, for flexible clauses requiring professional judgment, existing technologies often rely on simple prompts or complete manual processing, lacking the ability to provide quantitative basis and explainable suggestions.

[0004] The above defects result in limited automation, low processing efficiency and lack of processing capability for complex tasks in current city planning verification work. SUMMARY

[0005] To solve the above defects, the present application provides a city planning verification method and system based on natural language processing.

[0006] The above invention objectives of the present application are achieved by the following technical solutions:

[0007] A city planning verification method based on natural language processing, comprising the steps of:

[0008] receiving a user request instruction and analyzing the user request instruction using a natural language processing algorithm to obtain verification intent information and target area information;

[0009] based on the target area information, retrieving and matching a corresponding area rule set from a pre-constructed planning knowledge base, and generating a verification task list and a data requirement list based on the area rule set and the verification intent information;

[0010] Obtaining multi-source heterogeneous data based on the data requirement list and preprocessing to obtain standardized verification data and historical fusion features, the multi-source heterogeneous data including spatial geographic data, attribute document data and historical case data;

[0011] Traversing the verification task list and executing a preset differentiated verification process according to the task type, the task type including deterministic rule tasks and fuzzy rule tasks;

[0012] Generating a verification report based on the differentiated verification process result and updating the planning knowledge base based on feedback information;

[0013] The differentiated verification process includes:

[0014] When the task type is a deterministic rule task, performing spatial calculation and numerical logic judgment on the standardized verification data based on a preset verification logic rule to generate a verification result containing a numerical comparison process;

[0015] When the task type is a fuzzy rule task, extracting multi-modal features from the standardized verification data, and performing similarity matching between the multi-modal features and the historical fusion features, and matching a similar case set based on the similarity matching result, and generating a verification result containing an auxiliary decision suggestion based on the similar case set.

[0016] In a preferred example, the application can be further configured as: the pre-constructed planning knowledge base, the construction process of which includes the following steps:

[0017] Obtaining planning basis text data and extracting rule elements, the rule elements including core planning indicators, planning constraint conditions and planning target values;

[0018] Classifying the planning provisions in the corresponding planning basis text data into deterministic rules and fuzzy rules based on the rule elements;

[0019] Converting the deterministic rules into machine-readable rules with executable verification logic rules;

[0020] Obtaining historical case data and performing feature extraction and vectorization to form historical fusion features;

[0021] Obtaining a project database and associating the fuzzy rules with the project database and the historical fusion features;

[0022] Storing the machine-readable rules, fuzzy rules and historical fusion features in the form of a knowledge graph to form a planning knowledge base.

[0023] The application can be further configured in a preferred example as follows: the step of obtaining historical case data and performing feature extraction and vectorization representation to form historical fusion features includes the steps of:

[0024] extracting spatial geographic data, document text data and planning drawing images in the historical case data, and performing unified representation learning on them through a deep learning-based multi-modal fusion network to generate a deep feature tensor of each case;

[0025] extracting a set of planning constraints based on the document text data and the planning drawing images, and setting the deep feature tensor and the set of planning constraints of the same case as a condition-feature pair;

[0026] using the condition-feature pair as a training sample to drive the conditional generative adversarial network to perform adversarial training to generate a plurality of virtual case features;

[0027] mapping the virtual case features and the corresponding deep feature tensors to a low-dimensional manifold space and performing reorganization and clustering to form historical fusion features.

[0028] The application can be further configured in a preferred example as follows: the project database includes a spatial database, a rule database and an entity relationship database, and the step of obtaining the project database and associating the fuzziness rule with the project database and the historical fusion features includes the steps of:

[0029] performing semantic encoding on the fuzziness rule and generating a semantic vector;

[0030] extracting geographic entities and rule entities from the spatial database and the rule database respectively, and encoding the geographic entities and the rule entities to generate geographic entity embedding vectors and rule entity embedding vectors;

[0031] based on the association relationship in the entity relationship database, dynamically linking and fusing the geographic entity embedding vectors and the rule entity embedding vectors to form a fusion embedding vector;

[0032] using the semantic vector of the fuzziness rule, the fusion embedding vector and the historical fusion features as a hyperedge to construct a hypergraph model;

[0033] using a meta-learning framework to simulate decision-making tasks in different scenarios in the target region information, dynamically evaluating and adjusting the weights and strengths of various association relationships in the hypergraph model, and feeding back the adjusted association relationships to the hypergraph model.

[0034] The application can be further configured in a preferred example as follows: when the task type is a deterministic rule task, specifically:

[0035] extracting a first rule identifier corresponding to the deterministic rule task from the verification task list;

[0036] identifying, based on the first rule identifier, a corresponding machine-readable rule from the planning knowledge base, the machine-readable rule comprising a predefined verification logic function and a required data pattern;

[0037] matching and extracting, according to the data pattern, data field values required by the machine-readable rule from the standardized verification data, instantiating the verification logic function with the extracted data field values as input parameters, and generating an executable computing object for the current deterministic rule task;

[0038] running the executable computing object and outputting at least one original computing result, reading metadata attributes of the standardized verification data, the metadata attributes comprising data accuracy, error range or confidence level;

[0039] aggregating and calculating the error range of the standardized verification data based on its metadata, obtaining an uncertainty interval of the original computing result;

[0040] comparing the uncertainty interval of the original computing result with a preset target threshold, and generating a verification result comprising the comparison process based on the comparison result.

[0041] In a preferred example, the application can be further configured to, when the task type is a fuzziness rule task, specifically:

[0042] extracting a second rule identifier corresponding to the fuzziness rule task from the verification task list;

[0043] querying, based on the second rule identifier, an associated fuzziness rule from the planning knowledge base, the fuzziness rule comprising a predefined set of natural language descriptions;

[0044] generating corresponding feature extraction instructions according to the set of natural language descriptions, and extracting standardized features from the standardized verification data according to the feature extraction instructions, the standardized features comprising a geospatial feature subset, an attribute document feature subset and an image texture feature subset;

[0045] inputting the standardized feature set into a pre-trained feature fusion model to generate a unified current status feature vector;

[0046] performing similarity calculation on the unified current status feature vector and a case feature vector contained in the historical fusion features, and matching a plurality of similar historical cases to form a similar case set according to the similarity calculation result;

[0047] obtaining, from the planning knowledge base, decision results of all similar historical cases in the similar case set and their decision basis explanation texts;

[0048] Based on the preset attention mechanism, the current status feature vector is taken as a query key to calculate the attention weight of each similar case feature vector, and based on the attention weight, the decision basis description text is weighted and fused to generate an auxiliary decision suggestion and a verification result containing the auxiliary decision suggestion.

[0049] In a preferred example, the application can be further configured to: based on the target area information, retrieve and match the corresponding area rule set from the pre-constructed planning knowledge base, and generate the verification task list and the data requirement list based on the area rule set and the verification intention information, including the steps of:

[0050] Based on the target area information, multi-level spatial retrieval is performed in the planning knowledge base, and the corresponding area rule set is matched based on the retrieval result, and the multi-level spatial retrieval includes area boundary matching, geographic grid matching and spatial semantic matching;

[0051] The area rule set is subjected to rule conflict detection, and the rule conflict includes numerical rule conflict and spatial rule conflict;

[0052] Based on the preset conflict resolution strategy, the rule conflict is resolved to generate a consistent area rule set;

[0053] Based on the verification intention information, the intention-related rule subset is screened from the consistent area rule set, and through semantic correlation calculation, the key rules with a correlation degree higher than a threshold value are retained;

[0054] According to the rule type, timeliness and importance associated with the key rules, the verification priority of the key rules is assigned, and a verification task list with priority marking is generated.

[0055] The above-mentioned second invention purpose of the application is realized by the following technical scheme:

[0056] A city planning verification system based on natural language processing, comprising:

[0057] An instruction analysis module for receiving user request instructions and analyzing the user request instructions using a natural language processing algorithm to obtain verification intention information and target area information;

[0058] A list generation module for retrieving and matching the corresponding area rule set from the pre-constructed planning knowledge base based on the target area information, and generating the verification task list and the data requirement list based on the area rule set and the verification intention information;

[0059] A data processing module for obtaining and preprocessing multi-source heterogeneous data based on the data requirement list to obtain standardized verification data and historical fusion features, wherein the multi-source heterogeneous data includes spatial geographic data, attribute document data and historical case data.

[0060] The verification execution module is configured to traverse the verification task list and execute a preset differentiated verification process according to a task type, the task type including a deterministic rule task and a fuzzy rule task.

[0061] The report generation module is configured to generate a verification report based on the differentiated verification process result and update the planning knowledge base based on feedback information.

[0062] The differentiated verification process includes:

[0063] When the task type is the deterministic rule task, spatial calculation and numerical logic judgment are performed on the standardized verification data based on a preset verification logic rule to generate a verification result including a numerical comparison process.

[0064] When the task type is the fuzzy rule task, multi-modal features are extracted from the standardized verification data, the multi-modal features are matched with historical fusion features in similarity, a similar case set is matched based on a similarity matching result, and a verification result including an auxiliary decision suggestion is generated based on the similar case set.

[0065] The application also relates to a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the urban planning verification method based on natural language processing.

[0066] The application also relates to a computer readable storage medium storing a computer program, wherein the computer program is executable on a processor to implement the steps of the urban planning verification method based on natural language processing.

[0067] To sum up, the urban planning verification method and system based on natural language processing provided by the application can parse a user request and build a planning knowledge base through natural language processing, can process deterministic and fuzzy rule tasks in combination with a differentiated verification process, can realize the collaborative application of rule clauses and case features, can solve the problems of low automation and insufficient fuzzy clause processing capacity of traditional verification methods, and can improve verification efficiency and enhance the explainability of decision suggestions. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is a flowchart of an embodiment of the urban planning verification method based on natural language processing provided by the application;

[0069] Figure 2 is an implementation flowchart of the planning knowledge base construction process in an embodiment of the urban planning verification method based on natural language processing provided by the application;

[0070] Figure 3 is an implementation flowchart of step S20 in an embodiment of the city planning verification method based on natural language processing. DETAILED DESCRIPTION

[0071] The following will be described in detail in combination with the accompanying drawings. Figures 1-3 The present application will be further described in detail.

[0072] In an embodiment, as shown in the figure, the present application discloses a city planning verification method based on natural language processing, specifically comprising the following steps: Figure 1

[0073] S10: receiving a user request instruction and using a natural language processing algorithm to analyze the user request instruction to obtain verification intention information and target area information;

[0074] In this embodiment, the user request instruction is an operation instruction initiated by the user control terminal, which is usually input in the form of language text and may contain one or more verification requirements, such as “check the planning compliance of plot A” or “evaluate the coordination of the newly built school in region B with the surrounding landscape” and the like; the natural language processing algorithm is a technology for converting the input user request instruction into a structured query through a semantic analysis model and the like, which can be specifically implemented by using a sequence labeling model based on Transformer to extract key elements in the verification intention; the verification intention information is the core verification target representation obtained by analyzing the user instruction, which defines the nature of the task, such as “compliance check”, “landscape coordination evaluation” or “land use suitability analysis” and the like, and the verification intention information is used to guide the direction of subsequent knowledge base retrieval and task generation; the target area information is the spatial position or range to be verified parsed from the user instruction, which can be an explicit plot number, geographical name, or a relatively vague spatial description.

[0075] S20: based on the target area information, retrieving and matching the corresponding regional rule set from the pre-constructed planning knowledge base, and generating a verification task list and a data requirement list based on the regional rule set and the verification intention information;

[0076] ​In the embodiment, the pre-constructed planning knowledge base is a knowledge graph storing machine-readable rules and historical case features, which can be implemented by a rule engine and a graph database, and is used to support dynamic rule matching and case retrieval; the regional rule set is a collection of all applicable planning rules associated with the target regional information, as different planning management unit regions and different functional regions (such as core areas and landscape coordination areas) have different planning requirements, the region-specific rule set needs to be located first to ensure the accuracy of the verification; the verification task list is an automatically generated ordered list of specific verification items to be executed, wherein each task in the verification task list corresponds to one or a group of related rules in the regional rule set and has been classified and labeled (such as a deterministic task or a fuzzy task); the data requirement list is a list of required data and metadata requirements derived for completing all items in the verification task list, which clearly indicates which data (such as plot boundaries, building area, and surrounding green space rate) needs to be obtained, the format of the data to be obtained, and the accuracy of the data.

[0077] S30: Obtain multi-source heterogeneous data based on the data requirement list and pre-process the multi-source heterogeneous data to obtain standardized verification data and historical fusion features, wherein the multi-source heterogeneous data includes spatial geographic data, attribute document data, and historical case data;

[0078] In the embodiment, the pre-processing of the multi-source heterogeneous data is a standardization processing of the spatial geographic data, the attribute document data, and the historical case data, which can be implemented by a spatial coordinate system conversion and feature embedding method, and is used to eliminate data format differences; the standardized verification data is obtained after pre-processing of the multi-source heterogeneous data such as cleaning, format conversion, coordinate unification, and structured processing, and can be directly called by subsequent verification algorithms; the historical fusion features are unified numerical representations formed after deep analysis and feature extraction of the multi-source heterogeneous data, and are not simply case archives, but are encoded into high-dimensional vectors by machine learning methods, so as to facilitate efficient operations (such as similarity comparison).

[0079] S40: Traverse the verification task list and execute a preset differentiated verification process according to the task type, wherein the task type includes a deterministic rule task and a fuzzy rule task;

[0080] In the embodiment, the task type is a classification identification of each task in the verification task list, wherein the deterministic rule task refers to a task that requires clear and quantifiable rules, and a conclusion of yes or no can usually be obtained through logical judgment and numerical calculation, for example, "whether the volume rate is ≤2.0", etc., and the fuzzy rule task refers to a task that requires professional experience for interpretation and trade-off, for example, "whether the building style is coordinated with the ancient city style", etc.; the differentiated verification process is a decision mechanism for selecting different processing paths according to the task type, which can be specifically implemented through a rule classifier and a task distribution module to improve the effectiveness of the verification process.

[0081] S50: generating a verification report based on the result of the differentiated verification process, and updating the planning knowledge base based on the feedback information;

[0082] The differentiated verification process comprises:

[0083] When the task type is a deterministic rule task, spatial calculation and numerical logic judgment are performed on the standardized verification data based on a preset verification logic rule to generate a verification result containing a numerical comparison process;

[0084] When the task type is a fuzzy rule task, multi-modal features are extracted from the standardized verification data, and the multi-modal features are matched with historical fusion features for similarity matching, and a similar case set is matched based on the similarity matching result, and a verification result containing an auxiliary decision suggestion is generated based on the similar case set.

[0085] In this embodiment, the verification report is an output result document integrating the verification results of all certainty rule tasks and ambiguity rule tasks, usually including explicit compliance conclusions, numerical comparison details, case reference suggestions, and related data visualization content; the feedback information is the user's modification, confirmation or evaluation information of the output verification report or auxiliary decision suggestion, which is used to optimize and update the planning knowledge base, such as adjusting rule weights and enriching case features; the verification logic rule is a function or algorithm rule predefined for executing certainty rule tasks, which usually takes the form of a conditional judgment statement or a calculation formula; the multi-modal feature is a feature set extracted from standardized verification data, representing the attributes of the current verification object. Since the standardized verification data itself is multi-modal (such as image, text, and numerical value), the extracted multi-modal features also cover visual features (extracted from drawings), text features (extracted from documents), and numerical features (extracted from attribute tables); similarity matching is the process of comparing the distance (such as cosine similarity, Euclidean distance) between multi-modal features and historical fusion features in the numerical space, so as to find out several cases most similar to the current situation from historical cases; the similar case set is a set of historical cases most similar to the current verification project in key features selected through similarity matching; the auxiliary decision suggestion is the output result generated for the ambiguity rule task. Usually, the auxiliary decision suggestion is not a simple "yes / no" conclusion, but a quantitative reference opinion and reason explanation based on the analysis of the "similar case set", aiming to help users make more scientific decisions.

[0086] Specifically, for example, first parse the user's input verification related user request instruction through natural language processing model, such as "check whether the newly built residential area in region A meets the volume rate requirement", extract the target region A and the verification type as volume rate verification; when retrieving the relevant rules of region A in the planning knowledge base, find the explicit volume rate upper limit value and the fuzzy landscape coordination requirement; automatically generate a verification list containing certainty tasks and ambiguity tasks, integrate satellite images, land use approval documents and historical approval cases, etc. Multi-source heterogeneous data through related modules, and convert the volume rate data to a unified unit of measurement, while extracting the spatial layout features in historical cases; for certainty volume rate verification, directly perform numerical comparison and calculate error interval; for landscape coordination requirements, match the feature vectors of similar historical cases, and generate verification results containing spatial layout optimization suggestions; finally, integrate the numerical comparison process and case suggestions in the final report, and feed back the new case features to the knowledge base.

[0087] By the technical solution, the application realizes automatic conversion of natural language instructions to machine executable logic, improves verification task processing efficiency; through differential verification process design, quantitative index verification and qualitative requirement evaluation are effectively distinguished, and complex clause processing capability is enhanced; multi-source data fusion and historical case feature matching provide interpretable decision basis for fuzzy rules, reducing the need for manual intervention; at the same time, the dynamic updating mechanism of the knowledge base continuously optimizes the relevance of rule matching accuracy and case recommendation.

[0088] In an embodiment, as shown in Figure 2 The pre-constructed planning knowledge base includes the following steps in the construction process:

[0089] T10: Obtain planning basis text data and extract rule elements, the rule elements including core planning indicators, planning constraint conditions and planning target values;

[0090] In this embodiment, the planning basis text data is unstructured or semi-structured text material with guiding significance in the field of urban planning, usually including "Urban Planning Management Technical Regulations", control detailed planning text, special planning report, land planning design condition notice and other documents; the rule elements are quantifiable or structured indicators and conditions extracted from the planning text, which can be realized by using named entity recognition and relationship extraction algorithms in natural language processing technology to provide basic semantic units for subsequent rule classification, wherein the rule elements include core planning indicators, planning constraint conditions and planning target values; the core planning indicators are specific objects or attributes constrained by the provisions, such as "volume rate", "green rate", "building height limit", "building density" and the like; the planning constraint conditions are limitations or operations on core indicators, such as "not higher than", "not less than", "prohibited", "preferable" and the like indicating restrictive or suggestive words; the planning target values are specific numerical values, percentages or grades matched with core indicators and constraint conditions, such as "2.0", "35%", "50 meters", "first class" and the like.

[0091] T20: Classify the planning provisions in the corresponding planning basis text data into deterministic rules and fuzzy rules based on the rule elements;

[0092] In this embodiment, the deterministic rules are provisions with explicit numerical values or logical judgment conditions, which can be converted into machine executable code segments containing logical operators through predefined conversion templates to support automated verification; the fuzzy rules are provisions containing subjective descriptions or qualitative requirements, which can be represented by high-dimensional vectors through semantic coding for semantic association with historical cases.

[0093] T30: Convert the deterministic rules into machine readable rules with executable verification logic rules;

[0094] In this embodiment, the machine-readable rule is a rule form generated after the transformation of the deterministic rule and can be directly interpreted and executed by a computer program, which is usually a function with parameters or a piece of logic code.

[0095] T40: Obtain historical case data and perform feature extraction and vectorization representation to form historical fusion features;

[0096] In this embodiment, feature extraction and vectorization representation are technical processes for converting unstructured historical case data into structured numerical features. Feature extraction refers to identifying and extracting representative information (such as building contours and text keywords) from raw data (such as drawings and texts). Vectorization representation refers to converting this information into numerical vectors in a high-dimensional space, so that computers can process and analyze these historical cases through mathematical operations. Historical fusion features are vectorized representations of multi-modal feature fusion of historical cases, which can be generated by jointly encoding spatial geographic data, text data, and image data using deep learning models to improve the accuracy of case matching.

[0097] T50: Obtain the project database and associate the fuzzy rule with the project database and the historical fusion features;

[0098] In this embodiment, the project database is a database collection that stores entities, attributes, and relationships related to urban planning projects. The project database includes a spatial database, a rule database, and an entity relationship database, which can be constructed through data extraction and entity linking techniques to provide context association basis for fuzzy rules. The spatial database stores spatial graphics and attribute information of geographic entities (such as land, roads, mountains, and water systems). The rule database stores rules, regulations, and their metadata. The entity relationship database stores the associations between various entities (such as a land belonging to a certain rule area or a case referencing a certain regulation). Establishing an association means creating clear and machine-understandable connection relationships between the fuzzy rule, entities in the project database, and historical fusion features using graph computing and semantic matching techniques. For example, the fuzzy rule "protect the landscape pattern" is associated with geographic entities such as "mountain A" and "water system B" in the database, as well as case feature vectors that have handled similar scenarios in history.

[0099] T60: Store the machine-readable rule, the fuzzy rule, and the historical fusion features in the form of a knowledge graph to form a planning knowledge base.

[0100] In this embodiment, the knowledge graph is a knowledge base that stores rule and data association relationships in a graph structure, which can be implemented using graph database technology to support multi-dimensional knowledge retrieval and reasoning.

[0101] Specifically, in the construction of the planning knowledge base, first, the core indicators, constraint conditions and other rule elements are extracted from the planning text through natural language processing technology, for example, the BERT model is used to identify the volume rate, green rate and other numerical indicators in the text; then the provisions are divided into two categories of deterministic rules and fuzzy rules according to the definiteness of the rule elements, for example, “the volume rate shall not exceed 2.0” is classified as a deterministic rule, and “the architectural style needs to be coordinated with the surrounding” is classified as a fuzzy rule; for deterministic rules, executable code is generated through a pre-configured conversion rule library, for example, “shall not exceed” is converted to “≤” operator and a corresponding threshold judgment function is generated; historical case data is extracted through a multi-modal fusion network, for example, a convolutional neural network is used to process planning drawing images, and an LSTM model is used to process text descriptions, and finally a unified feature vector is generated; the fuzzy rule is associated with the geographic entities and rule clauses in the project database after the vector is generated through semantic encoding, for example, the “coordination” requirement is associated with the surrounding building height, style and other entity attributes; finally, all rules, cases and associated relationships are stored in the form of a knowledge graph, for example, a Neo4j graph database is used to construct a relationship network of nodes and edges.

[0102] Through the above technical solutions, the present application solves the problem in the prior art that natural language provisions are difficult to convert into executable rules and fuzzy clauses lack quantitative processing basis; through rule classification and machine-readable conversion mechanism, automatic checking of deterministic rules is realized; by constructing a knowledge graph that integrates multi-source data, fuzzy rules can be associated with project data and historical cases, for example, when processing “traffic impact assessment” and other fuzzy requirements, the traffic flow data and approval conclusions of similar cases are automatically matched, providing quantifiable reference for the checking personnel; at the same time, the dynamic association mechanism of the knowledge graph avoids the problem of data silos, for example, when the rule database is updated, the related fuzzy rules can automatically synchronize the latest clauses, ensuring the timeliness of the checking basis.

[0103] In an embodiment, step T40 comprises:

[0104] T41: Extracting spatial geographic data, document text data and planning drawing images from historical case data, and performing unified feature learning through a deep learning-based multi-modal fusion network to generate a deep feature tensor for each case;

[0105] In the embodiment, the multi-modal fusion network is a neural network architecture capable of joint feature learning of images, texts and spatial data, and can specifically be implemented by using a cross-modal encoder based on an attention mechanism. The network realizes unified representation by establishing a correlation mapping relationship between different modal data. Unified representation learning is to map original data of different forms and structures (such as image pixels, text words and spatial coordinates) into a common semantic high-dimensional vector space by using a machine learning model. In the unified high-dimensional vector space, originally heterogeneous data are represented as numerical vectors of the same form, so that the computer can process them by using the same mathematical method (such as calculating the cosine similarity). The deep feature tensor is high-dimensional and structured numerical data capable of deeply representing the essence of a case, which is extracted by the multi-modal fusion network.

[0106] T42: Extract a set of planning constraints based on the document text data and the planning drawing image, and set the deep feature tensor and the set of planning constraints of the same case as a condition-feature pair;

[0107] In the embodiment, the set of planning constraints is a set of planning requirements extracted from the text and drawing of the case and having an impact on the case. Generally, the set of planning constraints is not only simple rules, but also includes specific restrictions followed by the case, such as “height limit of 80 meters”, “in the style protection area”, “10% affordable housing” and the like. The condition-feature pair is a standard data structure for training the condition generation model. The “condition” and “feature” are paired to form a “cause-result” sample, which is used to teach the model to understand how different planning constraints will finally be reflected in the specific spatial and form design (i.e. feature).

[0108] T43: Use the condition-feature pair as a training sample to drive the condition generation adversarial network to perform adversarial training and generate a plurality of virtual case features;

[0109] In the embodiment, the condition generation adversarial network is an adversarial generation model with planning constraint conditions, which can be specifically implemented by using a generator with a conditional encoder and a discriminator structure. The condition generation adversarial network generates virtual case features that meet specific constraint conditions through adversarial training. The virtual case feature is a feature representation of an imaginary case generated by the trained condition generation adversarial network and does not exist in the original historical data. The virtual case feature is reasonable in data distribution and semantics, simulates possible scheme forms under known planning constraints, and is used to expand the sample size and fill in the areas not covered by the historical cases, so as to make the historical fusion features formed subsequently more rich and comprehensive, thereby improving the ability to cope with rare or complex situations.

[0110] T44: The virtual case features are collectively mapped to a low-dimensional manifold space with corresponding deep feature tensors, and are reorganized and clustered to form historical fusion features.

[0111] In this embodiment, the low-dimensional manifold space is an implicit feature space constructed by a nonlinear dimension reduction method, which can be implemented by using a variational autoencoder or a manifold learning algorithm. The low-dimensional manifold space realizes effective aggregation of features by preserving the topological structure of the original data. The mapping to the low-dimensional manifold space is a dimension reduction process, which aims to filter out noise and redundant information, capture the highest distinguishing features of the data, and greatly reduce the computational load. The reorganization is a process of standardizing and rearranging the data after mapping to the low-dimensional manifold, so that the distribution is more regular and convenient for subsequent processing. The clustering is an unsupervised machine learning method, which aims to automatically group case features with similar positions in the manifold space to form different clusters. Each cluster represents a type of plan with similar planning characteristics, such as "high-density commercial development cluster" and "low-density landscape protection cluster".

[0112] Specifically, in the process of historical case data processing, first, the multi-modal fusion network is used to extract convolutional features from planning drawings, encode semantics from document texts, and analyze topologies from spatial geographic data, generating deep feature tensors containing multi-dimensional information. The set of planning constraints is extracted from the document text by natural language processing technology, and the spatial layout features are identified from the planning drawings by image recognition technology. The construction of condition-feature pairs enables subsequent adversarial training to generate cases based on planning constraints. The condition generative adversarial network trains the discriminator to distinguish between real cases and generated cases, forcing the generator to output virtual case features that meet the planning constraints. The mapping process of the low-dimensional manifold space uses a variational autoencoder to reduce the dimension of high-dimensional features, and a clustering algorithm is used to aggregate similar case features into representative historical fusion features.

[0113] Through the above technical solutions, the application can effectively improve the expression ability and generalization performance of historical case features. By generating virtual case features that meet the planning constraints, the application solves the problem of reduced matching accuracy caused by insufficient actual case data. Through deep fusion of multi-modal data, the feature representation ability for complex planning cases is enhanced. The construction of the low-dimensional manifold space makes the similar case matching process more efficient and reliable, providing a more accurate reference for fuzzy rule tasks.

[0114] In an embodiment, the project database includes a spatial database, a rule database, and an entity relationship database, and step T50 includes:

[0115] T51: Semantically encode the fuzzy rule and generate a semantic vector;

[0116] In the embodiment, semantic coding is to convert the ambiguity rules of natural language description into vector representation, which can be implemented by using a BERT pre-training model to extract rule semantic features; and the semantic vector is a high-dimensional numerical vector obtained by processing through the semantic coding technology.

[0117] T52: Extracting geographic entities and rule entities from the spatial database and the rule database respectively, and encoding the geographic entities and the rule entities to generate geographic entity embedding vectors and rule entity embedding vectors;

[0118] In the embodiment, the spatial database is a database for storing geographic spatial data and spatial topological relations, which can be implemented by using a geographic information system (GIS) database to store geographic entity coordinates, spatial boundaries and topological relations; the rule database is a database for storing rule provisions and normative documents, which can be implemented by using a structured document database to store rule entity names, clause contents and effective times; the entity relationship database is a database for storing the association relationship between geographic entities and rule entities, which can be implemented by using a graph database to record the spatial constraint relationship and the rule reference relationship between entities; the geographic entity is a specific object occupying a certain spatial position in the real world and having a clear geographic meaning, such as a land plot, a river, a mountain, a road, a historical building, a protected area boundary, etc.; the rule entity is the smallest unit of information with clear semantics and effectiveness in the rule provision document, which can be a specific provision or a special planning, or a clause or a specific concept in the provision; the geographic entity embedding vector is to map the geographic entity into a low-dimensional vector, which can be implemented by using a graph embedding algorithm to represent the spatial attributes and association features of the geographic entity; and the rule entity embedding vector is to map the rule entity into a low-dimensional vector, which can be implemented by using a word embedding technology to represent the semantic attributes and effectiveness features of the rule entity.

[0119] T53: Based on the association relationship in the entity relationship database, dynamically coupling and fusing the geographic entity embedding vector and the rule entity embedding vector to form a fused embedding vector;

[0120] In the embodiment, the dynamic coupling and fusion is to dynamically adjust the combination mode of the embedding vectors according to the entity relationship, which can be implemented by using an attention mechanism to capture the dynamic association features between entities, wherein the coupling refers to pairing the corresponding geographic entity embedding vector and rule entity embedding vector according to the pre-defined association relationship in the entity relationship database, and the fusion refers to merging the two paired embedding vectors into a fused embedding vector that can comprehensively reflect the joint information of the two vectors through a specific pre-set algorithm (such as vector splicing followed by neural network processing, attention mechanism weighted average).

[0121] T54: The semantic vector of the ambiguity rule, the fusion embedding vector, and the historical fusion feature are jointly used as a super edge to construct a supergraph model;

[0122] In this embodiment, the supergraph model is a graph structure model connecting multi-dimensional data entities by super edges. The supergraph model can be implemented by a supergraph neural network, which is used to establish multi-dimensional associations among rules, data, and cases. The super edge is a set of associations in the supergraph model. In this scheme, a super edge links a semantic vector of an ambiguity rule, a set of fusion embedding vectors of all geographical and rule entities involved in the ambiguity rule, and all historical fusion feature vectors related to the ambiguity rule, forming a semantic association cluster.

[0123] T55: A meta-learning framework is used to simulate decision-making tasks in different scenarios in the target area information, dynamically evaluate and adjust the weights and strengths of various association relationships in the supergraph model, and feed back the adjusted association relationships to the supergraph model.

[0124] In this embodiment, the meta-learning framework is a training framework for optimizing model parameters by simulating multi-scenario tasks. The meta-learning framework can be implemented by a model-agnostic meta-learning (MAML) algorithm, which is used to dynamically adjust the weight distribution strategy of the association relationships in the supergraph model.

[0125] Specifically, geographical entities (such as road networks and land use boundaries) in a spatial database are connected to rule entities (such as volume rate clauses and green rate requirements) in a rule database through association relationships recorded in an entity relationship database (such as a certain region applying a specific rule). The geographical entity embedding vector and the rule entity embedding vector are dynamically fused through an attention mechanism to form a fusion embedding vector containing spatial constraints and rule constraints. The semantic vector of the ambiguity rule and the fusion embedding vector jointly serve as a super edge to establish multi-dimensional associations between historical case features and current project data. The meta-learning framework simulates decision-making tasks in different regional scenarios (such as commercial areas and residential areas) to dynamically optimize the association strength among rules, data, and cases in the supergraph model. For example, in a new development area scenario where historical cases are scarce, the association weight between rule entities and geographical entities is automatically enhanced.

[0126] Through the above technical solutions, the application can enhance the dynamic association capability between ambiguity rules and multi-source data, solve the problem of insufficient decision-making basis caused by the isolation of rules and data in the prior art, and automatically associate building density data in a spatial database, landscape control requirements in a rule database, and coordination judgment basis in historical cases to generate auxiliary decision-making suggestions with quantitative support when evaluating fuzzy clauses such as “coordination with the surrounding environment”. Furthermore, the meta-learning framework dynamically optimizes the association relationships in different scenarios, which can effectively improve the adaptability and explainability of the application of ambiguity rules and avoid decision-making bias caused by rigid data association.

[0127] In an embodiment, when the task type is a deterministic rule task, specifically:

[0128] A10: Extracting a first rule identifier corresponding to the deterministic rule task from the verification task list;

[0129] In this embodiment, the first rule identifier is a coded string used to uniquely identify the deterministic rule, which can specifically be generated by using a hash algorithm to summarize the rule text, for quickly locating the corresponding machine-readable rule in the planning knowledge base.

[0130] A20: Based on the first rule identifier, matching the corresponding machine-readable rule from the planning knowledge base, the machine-readable rule including a predefined verification logic function and required data patterns;

[0131] In this embodiment, the machine-readable rule is a logical expression that converts natural language provisions into programmable execution, which can specifically use semantic parsing technology to convert the constraint conditions in the provisions into a logical judgment function to realize the computability of the rule. The machine-readable rule includes a predefined verification logic function and required data patterns. The verification logic function encapsulates executable code units of specific calculation and judgment logic, which defines how to operate on input parameters and obtain a preliminary result. For example, for the volume rate rule, the verification logic function is to calculate the volume rate. The required data patterns specify which input data are required to execute the above function, as well as the format, type and structure that these data should meet.

[0132] A30: According to the data patterns, matching and extracting data field values required for machine-readable rule calculation from standardized verification data, and instantiating the verification logic function by taking the extracted data field values as input parameters, to generate an executable calculation object for the current deterministic rule task;

[0133] In this embodiment, the data pattern is a structured description that defines the input parameters of the verification logic function, which can specifically be in the form of JSON Schema or XML Schema, used to guide the extraction of matching field values from standardized verification data. Instantiation is a process of creating a specific object from a class or template in programming. In this embodiment, it specifically refers to finding and extracting specific numerical values from standardized verification data according to the data patterns, and then filling these numerical values into the verification logic function template as parameters, to generate a calculation instance that contains specific data for the current task and can be immediately run. The executable calculation object is the specific result produced after the instantiation process is completed, which is a program object that has loaded all necessary input data and is in a ready state, encapsulating complete calculation logic and current data context.

[0134] A40: running the executable computing object and outputting at least one original computing result, reading metadata attributes of the standardized verification data, the metadata attributes including data precision, error range, or confidence level;

[0135] In this embodiment, the original computing result is an initial output value obtained after directly running the executable computing object, for example, a calculated volume rate value or a Boolean judgment; the metadata attribute is a characteristic parameter for describing data quality, which can be specifically defined by a data precision, an error range, and the like attribute fields in the ISO 19115 standard, and is used for quantifying a confidence level of the computing result, wherein the metadata attribute includes data precision, error range, or confidence level, the data precision is an accurate degree of data, for example, whether a coordinate is in meters or in centimeters; the error range is a maximum deviation that can exist in a data value, for example, an area measurement value can have an error of ±5 m 2 ; and the confidence level is a probability evaluation on data accuracy or source reliability.

[0136] A50: aggregating and calculating an error range of the standardized verification data based on metadata to obtain an uncertainty interval of the original computing result;

[0137] In this embodiment, the uncertainty interval is a range value obtained by correcting the original computing result by an error propagation model or the like, rather than a point estimate, and exemplarily, the original volume rate is 2.15, but considering the error of the input data, the true value has a 95% probability of falling within the interval [2.10, 2.20].

[0138] A60: comparing the uncertainty interval of the original computing result with a preset target threshold value, and generating a verification result including a comparison process based on a comparison result.

[0139] In this embodiment, the target threshold value is a reference value for comparison, which is part of the machine-readable rule and is an explicit numerical value; and the comparison process is indispensable interpretability content in the verification result, and in general cases, the comparison process is not only a “pass / fail” conclusion, but also records detailed comparison logic, for example, the interval [2.00, 2.20] of the computing result has an overlapping part with the target threshold value 2.10, and thus it is determined that: “there is an uncertainty risk, which needs manual review”, the comparison process enhances the transparency and credibility of the whole system.

[0140] Specifically, for example, when a deterministic rule task is identified, first, the corresponding verification logic function and data requirement template are quickly retrieved through the rule identifier; for example, for the rule "building density ≤ 30%", the preset density calculation function and the required plot area, building area and other data fields are matched; through field extraction guided by the data mode, it is ensured that the input parameters are strictly matched with the function definition, avoiding execution errors caused by data structure differences; in the calculation result generation stage, the precision parameters of the data source are read synchronously, for example, the measurement error of the plot area data is ± 0.5%, and the possible fluctuation range of the density value is calculated through an error propagation model or the like; finally, the upper and lower limits of the calculation result are compared with the target threshold value, and when the threshold value of 30% is completely higher than the upper limit of the calculation result, it is determined to be compliant, and if the threshold value falls within the result interval, data review prompts are triggered.

[0141] Through the above technical solutions, the present application realizes the fine processing of the deterministic rule verification process, while maintaining the efficiency of automatic processing, and significantly improves the credibility of the verification conclusion by introducing a data quality evaluation mechanism.

[0142] In an embodiment, when the task type is a fuzzy rule task, specifically:

[0143] B10: extracting a second rule identifier corresponding to the fuzzy rule task from the verification task list;

[0144] In this embodiment, the second rule identifier is a coding sequence for uniquely identifying the fuzzy rule task, which can be implemented by using a hash algorithm to calculate the summary of the rule text, which functions to quickly locate the corresponding fuzzy rule item in the planning knowledge base.

[0145] B20: querying the associated fuzzy rule from the planning knowledge base based on the second rule identifier, wherein the fuzzy rule includes a set of natural language descriptions;

[0146] In this embodiment, the fuzzy rule is a planning constraint clause containing natural language descriptions, which can be implemented by using semantic parsing technology to convert the clause into a structured rule label, which functions to guide the generation of feature extraction instructions; the set of natural language descriptions is a set of keywords or phrases extracted from the fuzzy rule for guiding feature extraction, and the set of natural language descriptions is a specific interpretation of the abstract rule, for example, for the rule "coordinate with the surrounding environment", the set of natural language descriptions may include "architectural style", "color", "material", "volume", "skyline", etc.

[0147] B30: generating corresponding feature extraction instructions according to the natural language description set, and extracting the standardized checking data according to the feature extraction instructions to obtain a standardized feature set, the standardized feature set including a geospatial feature subset, an attribute document feature subset, and an image texture feature subset;

[0148] In the embodiment, the feature extraction instructions are a set of data extraction instructions generated according to rule semantics, and specifically, a template-based instruction generation engine can be used to implement the feature extraction instructions, which functions to filter out feature data related to the ambiguity rule from the multi-source heterogeneous data; the standardized feature set is a set of feature data converted to a unified format, and specifically, a data cleaning tool can be used to normalize the geospatial coordinates, document keywords, and image pixel matrix, which functions to eliminate the representation differences between the multi-source data, wherein the standardized feature set includes a geospatial feature subset, an attribute document feature subset, and an image texture feature subset, the geospatial feature subset is a feature describing the spatial form, distribution, and relationship, such as building density, spatial sequence, and contour complexity, etc.; the attribute document feature subset is a semantic feature extracted from the text, such as keyword frequency, theme distribution, and sentiment tendency, etc.; and the image texture feature subset is a visual feature extracted from the drawing or image, such as color distribution, texture feature, and edge information, etc.

[0149] B40: inputting the standardized feature set into a pre-trained feature fusion model to generate a unified status feature vector;

[0150] In the embodiment, the feature fusion model is a neural network model for integrating multi-modal features, and specifically, a multi-modal fusion network based on graph convolution can be used to implement the feature fusion model, which functions to generate a status feature vector with semantic consistency; and the unified status feature vector is an output result of the feature fusion model, which is a feature vector formed by compressing all multi-modal information (spatial, text, and visual) of the current checking project.

[0151] B50: performing similarity calculation on the unified status feature vector and a case feature vector included in the historical fusion feature, and matching a plurality of similar historical cases according to the similarity calculation result to form a similar case set;

[0152] In the embodiment, the similarity calculation is a mathematical operation for measuring the closeness of two vectors in a numerical space, and commonly used methods include cosine similarity and Euclidean distance, and in the embodiment, the similarity score between the unified status feature vector and each historical fusion feature vector in the knowledge base is calculated to quantify the similarity between the current project and each historical case; and the similar case set is a set of Top-K most relevant historical cases selected after sorting according to the similarity calculation result.

[0153] B60: Obtain the decision results and their decision basis texts of all similar historical cases in the similar case set from the planning knowledge base;

[0154] In the embodiment, the decision basis text is a literal record of the expert decision reason corresponding to each historical case stored in the knowledge base, which usually contains the main consideration factors, trade-off process and final decision reason of the expert review, for example: "because scheme A retreats in volume, reducing the oppression of the mountain, it is passed", which is the key text information for converting case experience into operational suggestions.

[0155] B70: Based on the preset attention mechanism, the uniform current situation feature vector is taken as the query key to calculate the attention weight of each similar case feature vector, and based on the attention weight, the decision basis texts are weighted and fused to generate an auxiliary decision suggestion and a verification result containing the auxiliary decision suggestion.

[0156] In the embodiment, the attention mechanism is an algorithm technology for calculating feature correlation, which can be implemented by combining multi-head self-attention mechanism with cosine similarity calculation, which dynamically evaluates the relevance of each historical case to the current task; in the attention mechanism, "query" is a vector representing the current needs or problems, in the embodiment, the "uniform current situation feature vector" is taken as the query key to "ask" the historical case library (key-value library), for example: "which case has the most similar features?"; the attention weight is the probability distribution value obtained by calculating the query key with all historical case features, wherein each value represents the importance or relevance of the corresponding historical case to solving the current problem, the higher the weight, the greater the reference value of the case; the auxiliary decision suggestion is the text information generated after weighted fusion and finally output, usually, the auxiliary decision suggestion is a natural language summary, which may be in the format: "according to the experience of [case A] [case B], the current project is similar to them in [certain aspect], the historical decision mainly considers [factor X] [factor Y], among which [case A] is passed due to [reason], [case B] needs to be modified. It is recommended to focus on reviewing [certain aspect] and referring to [certain] handling method.", the auxiliary decision suggestion provides quantitative basis and operational insight, rather than simple yes or no judgment.

[0157] Specifically, when the ambiguity rule task is identified, firstly, the corresponding natural language description rule in the knowledge base is quickly located through the second rule identifier; based on the rule semantics, feature extraction instructions are generated, for example, for the "coordination with the surrounding environment" clause, instructions containing spatial buffer analysis, building style feature extraction, etc. can be generated; after the standardized check data is subjected to multi-modal feature extraction, a unified present situation feature vector is generated through a feature fusion model, for example, the road network topology features and building facade texture features are jointly coded; the unified present situation feature vector is matched with the historical case features, and when the similarity is matched, the approximate nearest neighbor algorithm is used to retrieve the historical cases with a spatial layout similarity reaching a threshold; finally, through the preset attention mechanism, the correlation weight of the present situation features and each case feature is calculated, for example, the case with similar plot ratio adjustment amplitude is given a higher weight, and finally the weighted fusion auxiliary decision suggestion is generated.

[0158] Through the above technical solutions, the application realizes the automatic processing capability of the ambiguity planning clause, solves the problem of single case matching dimension and fragmented decision basis in the traditional method; through multi-modal feature fusion and dynamic weight distribution, the historical cases with substantial relevance to the current task can be accurately identified, and the generated auxiliary decision suggestion contains quantitative similarity indicators and weighted basis explanation, providing the planning review party user with decision support with traceability.

[0159] In an embodiment, as shown in FIG. 2, step S20 includes: Figure 3

[0160] S21: based on the target area information, multi-level spatial retrieval is performed in the planning knowledge base, and the corresponding regional rule set is matched based on the retrieval result, the multi-level spatial retrieval includes regional boundary matching, geographic grid matching and spatial semantic matching;

[0161] ​In the embodiment, the multi-level spatial retrieval is a joint retrieval of the planning knowledge base through three ways of superimposing regional boundaries, geographical grid division and spatial semantic analysis. The multi-level spatial retrieval can be implemented by using a spatial overlay analysis algorithm combined with a natural language processing technology, and is used to solve the rule omission problem caused by a single retrieval mode. The multi-level spatial retrieval includes regional boundary matching, geographical grid matching and spatial semantic matching. The regional boundary matching is the first level screening of the multi-level spatial retrieval, and is used to match rules according to a planning management unit division (such as a city, a district or a street) in which the target region information is located. The geographical network matching is the second level screening of the multi-level spatial retrieval, and is used to divide the urban space into regular or irregular geographical grids (such as a latitude and longitude grid or an H3 grid). Each grid has a corresponding planning attribute. By judging in which grid or grids the target region falls, more detailed rules across the regional boundaries (such as a certain style control area that may cross two planning management unit areas) are matched. The spatial semantic matching is the third level screening of the multi-level spatial retrieval, which goes beyond the simple geographical position to understand the function, attribute and context of the space. By analyzing the surrounding environment, self attribute and spatial relationship of the target region, rules related to the semantics are matched. For example, the analysis of the surrounding environment includes whether it is close to a water source or a transportation hub, and the analysis of the self attribute includes the analysis of the land use property.

[0162] S22: performing rule conflict detection on the regional rule set, the rule conflict including a numerical rule conflict and a spatial rule conflict;

[0163] In the embodiment, the rule conflict detection is an automatic logical checking process, which is used to identify contradictory or inconsistent provisions in the regional rule set, so as to avoid the situation that no decision can be made or an incorrect conclusion is output due to the contradiction of the rules. The numerical rule conflict refers to that two or more rules propose different numerical requirements for the same index, which cannot be satisfied at the same time. For example, one rule requires that the green space ratio is greater than or equal to 30%, and another rule of the same level requires that the green space ratio is greater than or equal to 35%, but the current land condition cannot satisfy both requirements. The spatial rule conflict refers to that two or more rules propose mutually exclusive functional or morphological requirements for the same space. For example, on the same land, one rule requires that the existing ancient trees be protected, and another rule requires that an underground garage be constructed according to the planning, and the two requirements may directly conflict in the spatial layout.

[0164] S23: resolving the rule conflict based on a preset conflict resolution strategy, to generate a consistent regional rule set;

[0165] In the embodiment, the conflict resolution strategy is to resolve the conflict according to the preset priority rule or the dynamically adjusted weight parameter, and specifically, a dynamic decision model based on reinforcement learning can be used to ensure the consistency of the rule set; the consistency area rule set is a rule set obtained after conflict detection and conflict resolution processing, and the internal contradictions are eliminated.

[0166] S24: Screen an intent-related rule subset from the consistency area rule set based on the verification intent information, and retain key rules with a semantic correlation degree higher than a threshold value through semantic correlation degree calculation;

[0167] In the embodiment, the intent-related rule subset is a rule set selected from the consistency area rule set and related to the verification intent information in semantics; the semantic correlation degree calculation is to evaluate the correlation degree of the rule and the verification intent through a vector space model or a semantic similarity algorithm, and specifically, a pre-trained language model can be used to calculate the text similarity, which is used to screen key rules highly related to the current task.

[0168] S25: Assign a verification priority to the key rule according to the rule type, timeliness and importance associated with the key rule, and generate a verification task list with a priority mark.

[0169] In the embodiment, the verification priority assignment is to dynamically sort according to the rule type, timeliness and importance indicators, and specifically, a multi-attribute decision analysis method can be used to optimize the execution order of the verification task, wherein the assignment basis includes: based on the rule type, the deterministic rule is usually prior to the fuzzy rule to quickly obtain a clear conclusion; based on the timeliness, the newly released and the rules about to expire may obtain a higher priority; based on the importance, the rules related to safety, ecological red line and other core control elements have the highest priority.

[0170] Specifically, when receiving the target area information, first, the jurisdiction range is determined through the area boundary matching, then the spatial unit is refined through the geographic grid matching, and finally the area characteristic description is analyzed through the spatial semantic matching; for example, when the target area is a new district of a city, the planning management unit division to which the new district belongs is matched first, and then the specific plot is located according to the latitude and longitude grid, and finally the relevant rules are screened in combination with the semantic description such as “waterfront area”; then the conflict detection is performed on the retrieved rule set, for example, the numerical conflict of “volume rate ≤ 3.0” and “volume rate ≥ 3.5” or the spatial contradiction of “building setback ≥ 10 meters” and “river view corridor” is found; at this time, the conflict resolution strategy is called, for example, the latest revised rule is preferentially adopted, or the weight is automatically adjusted according to the regional development positioning; after the conflict resolution is completed, the relevant rules are screened based on the checking intention, for example, when the checking intention is “commercial land compliance review”, the key rules related to commercial development are reserved through semantic calculation; finally, the task list with priority is generated according to the rule type (such as priority of mandatory indicators), timeliness (such as priority of new rules) and importance (such as priority of safety indicators).

[0171] Through the above technical solution, the application effectively solves the problems of incomplete planning rule retrieval and inefficient rule conflict processing, significantly improves the accuracy and integrity of the checking task generation; through the dynamic priority allocation mechanism, it ensures that the key rules are preferentially executed, reduces the situation that important indicators are missed, and optimizes the configuration efficiency of the checking resources.

[0172] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the application.

[0173] In an embodiment, a city planning checking system based on natural language processing is provided, which corresponds one-to-one to the city planning checking method based on natural language processing in the above embodiment. The city planning checking system based on natural language processing comprises:

[0174] An instruction analysis module is configured to receive a user request instruction and analyze the user request instruction by using a natural language processing algorithm to obtain checking intention information and target area information;

[0175] A list generation module is configured to retrieve and match a corresponding area rule set from a pre-constructed planning knowledge base based on the target area information, and generate a checking task list and a data requirement list based on the area rule set and the checking intention information;

[0176] The data processing module is configured to obtain multi-source heterogeneous data and perform preprocessing based on a data requirement list, to obtain standardized verification data and historical fusion features, and the multi-source heterogeneous data includes spatial geographic data, attribute document data, and historical case data.

[0177] The verification execution module is configured to traverse a verification task list and perform a preset differentiated verification process according to a task type, and the task type includes a deterministic rule task and a fuzzy rule task.

[0178] The report generation module is configured to generate a verification report based on a differentiated verification process result and update a planning knowledge base based on feedback information.

[0179] The differentiated verification process includes:

[0180] When the task type is the deterministic rule task, spatial calculation and numerical logic judgment are performed on the standardized verification data based on a preset verification logic rule, and a verification result including a numerical comparison process is generated.

[0181] When the task type is the fuzzy rule task, multi-modal features are extracted from the standardized verification data, the multi-modal features are matched with the historical fusion features in terms of similarity, a similar case set is matched based on a similarity matching result, and a verification result including an auxiliary decision suggestion is generated based on the similar case set.

[0182] The specific limitation of the city planning verification system based on natural language processing can refer to the limitation of the city planning verification method based on natural language processing in the above, and will not be repeated here. Each module in the city planning verification system based on natural language processing can be realized by software, hardware, and a combination thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0183] In an embodiment, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements a city planning verification method based on natural language processing when executing the computer program.

[0184] In an embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement a city planning verification method based on natural language processing.

[0185] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A city planning verification method based on natural language processing, characterized in that, The method comprises the steps of: receiving a user request instruction and parsing the user request instruction using a natural language processing algorithm to obtain verification intention information and target area information; based on the target area information, retrieving and matching a corresponding area rule set from a pre-constructed planning knowledge base, and generating a verification task list and a data requirement list based on the area rule set and the verification intention information; based on the data requirement list, obtaining multi-source heterogeneous data and preprocessing to obtain standardized verification data and historical fusion features, the multi-source heterogeneous data including spatial geographic data, attribute document data, and historical case data; traversing the verification task list and executing a preset differentiated verification process according to the task type, the task type including deterministic rule tasks and fuzzy rule tasks; generating a verification report based on the differentiated verification process result and updating the planning knowledge base based on the feedback information; wherein the differentiated verification process comprises: when the task type is a deterministic rule task, performing spatial calculation and numerical logic judgment on the standardized verification data based on a preset verification logic rule to generate a verification result containing a numerical comparison process; when the task type is a fuzzy rule task, extracting multi-modal features from the standardized verification data, and performing similarity matching between the multi-modal features and the historical fusion features, and matching a similar case set based on the similarity matching result, and generating a verification result containing an auxiliary decision suggestion based on the similar case set; wherein the step of retrieving and matching a corresponding area rule set from a pre-constructed planning knowledge base based on the target area information, and generating a verification task list and a data requirement list based on the area rule set and the verification intention information, comprises the steps of: based on the target area information, performing multi-level spatial retrieval in the planning knowledge base, and matching a corresponding area rule set based on the retrieval result, the multi-level spatial retrieval including area boundary matching, geographic grid matching, and spatial semantic matching; detecting rule conflicts in the area rule set, the rule conflicts including numerical rule conflicts and spatial rule conflicts; resolving the rule conflicts based on a preset conflict resolution strategy to generate a consistent area rule set; filtering an intention-related rule subset from the consistent area rule set based on the verification intention information, and retaining key rules with a correlation degree higher than a threshold value through semantic correlation degree calculation; assigning verification priorities to the key rules according to the rule types, time effectiveness, and importance associated with the key rules, and generating a verification task list with priority labels.

2. The urban planning verification method based on natural language processing according to claim 1, characterized in that: The construction process of the pre-constructed planning knowledge base comprises the following steps: obtaining planning basis text data and extracting rule elements, the rule elements including core planning indicators, planning constraint conditions, and planning target values; classifying the planning provisions in the corresponding planning basis text data into deterministic rules and fuzzy rules based on the rule elements; converting the deterministic rules into machine-readable rules with executable verification logic rules; obtaining historical case data and performing feature extraction and vectorization to form historical fusion features; obtaining a project database and associating the fuzzy rules with the project database and the historical fusion features; The machine-readable rules, the fuzziness rules and the historical fusion features are stored in the form of a knowledge graph to form a planning knowledge base. 3.The urban planning verification method based on natural language processing of claim 2, wherein: The step of obtaining historical case data and performing feature extraction and vectorization to form historical fusion features comprises the steps of: extracting spatial geographic data, document text data and planning drawing images in the historical case data, and performing unified representation learning on the data through a multi-modal fusion network based on deep learning to generate a deep feature tensor of each case; extracting a set of planning constraints based on the document text data and the planning drawing images, and setting the deep feature tensor and the set of planning constraints of the same case as a condition-feature pair; using the condition-feature pair as a training sample to drive the conditional generative adversarial network to perform adversarial training to generate a plurality of virtual case features; mapping the virtual case features and the corresponding deep feature tensors to a low-dimensional manifold space and performing reorganization and clustering to form historical fusion features.

4. The urban planning verification method based on natural language processing according to claim 2, characterized in that: The project database comprises a spatial database, a rule database and an entity relationship database, and the step of obtaining the project database and associating the fuzziness rules with the project database and the historical fusion features comprises the steps of: performing semantic coding on the fuzziness rules to generate a semantic vector; extracting geographic entities and rule entities from the spatial database and the rule database respectively, and coding the geographic entities and the rule entities to generate geographic entity embedding vectors and rule entity embedding vectors; based on the association relationship in the entity relationship database, dynamically linking and fusing the geographic entity embedding vectors and the rule entity embedding vectors to form a fusion embedding vector; using the semantic vector of the fuzziness rule, the fusion embedding vector and the historical fusion feature as a hyper-edge to construct a hypergraph model; using a meta-learning framework to simulate decision-making tasks in different scenarios in the target area information, dynamically evaluating and adjusting the weights and strengths of various association relationships in the hypergraph model, and feeding back the adjusted association relationships to the hypergraph model.

5. The urban planning verification method based on natural language processing according to claim 1, characterized in that: When the task type is a deterministic rule task, specifically: extracting a first rule identifier corresponding to the deterministic rule task from the verification task list; based on the first rule identifier, matching the corresponding machine-readable rule from the planning knowledge base, wherein the machine-readable rule comprises a predefined verification logic function and a required data pattern; extracting data field values required for machine-readable rule calculation from standardized verification data according to the data pattern, and instantiating the verification logic function by taking the extracted data field values as input parameters to generate an executable calculation object for the current deterministic rule task; running the executable calculation object and outputting at least one original calculation result, reading the metadata attributes of the standardized verification data, wherein the metadata attributes include data accuracy, error range or confidence; based on the metadata of the standardized verification data, aggregating the error range to obtain the uncertainty interval of the original calculation result; comparing the uncertainty interval of the original calculation result with a preset target threshold, and generating a verification result including the comparison process based on the comparison result.

6. The urban planning verification method based on natural language processing according to claim 1, characterized in that: When the task type is a fuzzy rule task, specifically: extracting a second rule identifier corresponding to the fuzzy rule task from the verification task list; querying an associated fuzzy rule from the planning knowledge base based on the second rule identifier, the fuzzy rule including a predefined natural language description set; generating corresponding feature extraction instructions according to the natural language description set, and extracting standardized features from the standardized verification data according to the feature extraction instructions, the standardized features including a geospatial feature subset, an attribute document feature subset, and an image texture feature subset; inputting the standardized features into a pre-trained feature fusion model to generate a unified current status feature vector; calculating the similarity between the unified current status feature vector and a case feature vector included in the historical fusion features, and matching a number of similar historical cases to form a similar case set according to the similarity calculation result; obtaining decision results and decision basis description texts of all similar historical cases in the similar case set from the planning knowledge base; based on a preset attention mechanism, taking the unified current status feature vector as a query key, calculating the attention weight of each similar case feature vector, and generating an auxiliary decision suggestion by weighting and fusing each decision basis description text based on the attention weight.

7. A natural language processing based urban planning verification system for performing the steps of the natural language processing based urban planning verification method according to any one of claims 1-6. comprise: an instruction analysis module configured to receive a user request instruction and analyze the user request instruction using a natural language processing algorithm to obtain verification intent information and target area information; a list generation module configured to retrieve and match a corresponding area rule set from a pre-constructed planning knowledge base based on the target area information, and generate a verification task list and a data requirement list based on the area rule set and the verification intent information; a data processing module configured to obtain and preprocess multi-source heterogeneous data based on the data requirement list to obtain standardized verification data and historical fusion features, the multi-source heterogeneous data including spatial geographic data, attribute document data, and historical case data; a verification execution module configured to traverse the verification task list and execute a preset differentiated verification process according to the task type, the task type including deterministic rule tasks and fuzzy rule tasks; a report generation module configured to generate a verification report based on the differentiated verification process result and update the planning knowledge base based on feedback information; wherein the differentiated verification process comprises: when the task type is a deterministic rule task, performing spatial calculation and numerical logic judgment on the standardized verification data based on a preset verification logic rule to generate a verification result including a numerical comparison process; when the task type is a fuzzy rule task, extracting multi-modal features from the standardized verification data, and matching the multi-modal features with the historical fusion features for similarity matching, and matching a similar case set based on the similarity matching result, and generating a verification result including an auxiliary decision suggestion based on the similar case set.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the urban planning verification method based on natural language processing according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by a processor to implement the steps of the city planning checking method based on natural language processing according to any one of claims 1-6.

Images