Spatial semantic enhanced real estate registration knowledge retrieval enhanced generation method

Abstract

The invention discloses a spatial semantic enhanced real estate registration knowledge retrieval enhancement generation method, and belongs to the cross technical field of geographic information technology, natural language processing and retrieval enhancement generation. Extracting spatial geometry and topology constraint features, ownership business features and text semantic features, and generating a joint query representation in a unified vector space; and performing multi-level mixed retrieval in the real estate registration multi-mode knowledge base, fusing retrieval results to construct context enhanced representation, inputting the context enhanced representation into the large language model, generating a structured response, and completing output. According to the retrieval reasoning technology based on spatial semantic enhancement, an efficient, accurate and reliable intelligent processing scheme is provided for real estate registration service query, the service query quality and efficiency are remarkably improved, and good application and popularization value is achieved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of geographic information technology, natural language processing, and enhanced retrieval generation, specifically to a spatial semantic enhancement method and system for enhancing real estate registration knowledge retrieval generation. Background Technology

[0002] Real estate registration knowledge exhibits typical multimodal characteristics, encompassing both spatial geometric data based on coordinate strings and unstructured text data such as registers and ownership certificates. Achieving a unified representation that combines spatial accuracy with semantic richness is the core technological challenge currently facing intelligent question answering for real estate registration.

[0003] In current real estate registration practices, intelligent question answering is mainly implemented based on a single text knowledge base of vectors. For cross-modal needs such as "finding all commercial real estate units within a 200-meter buffer zone of the target parcel and without any mortgage restrictions," it is necessary to break it down into independent spatial queries, text retrieval, and result intersection operations. Existing intelligent question answering solutions struggle to address this problem. The existing technical solutions mainly have the following limitations:

[0004] First, natural language understanding capabilities are limited. Traditional semantic retrieval methods based on vector similarity cannot understand complex spatial business intents such as "adjacent," "containment," and "seizure status association," nor can they handle open queries like "all unmortgaged properties under Zhang San's name in Zhonglou District," which require multi-hop reasoning and logical integration. Insufficient semantic parsing depth leads to low accuracy in query intent recognition, often requiring users to interact multiple times to obtain the desired results.

[0005] Secondly, there is a separation between spatial and textual modalities. Existing systems typically process spatial queries and attribute queries separately, lacking a joint embedding mechanism that maps spatial geometric features and textual semantic features to a unified vector space. This separation leads to inefficient cross-modal retrieval, failing to fully utilize the correlation between spatial relationships and business semantics. For complex queries involving multiple constraints such as spatial extent, topological relationships, and ownership status, the system response is slow and the results are poorly relevant.

[0006] Third, there is a lack of information retrieval regarding ownership. Real estate registration involves complex legal relationships related to ownership, such as mortgages, seizure registrations, and inheritance chains. Existing technologies have not constructed a coupled representation of the ownership timeline and spatial relationship network, making it difficult to support composite temporal and spatial queries. Furthermore, there is a lack of effective integration with business rules such as the "Interim Regulations on Real Estate Registration," making it impossible to automatically detect business conflicts such as "processing transfer registration while the property is already mortgaged." Summary of the Invention

[0007] Purpose of the invention: To address the shortcomings of existing technologies, this invention provides a spatial semantic enhancement method for real estate registration knowledge retrieval and generation, which can provide a real estate registration intelligent question-answering solution that deeply integrates spatial and textual semantic understanding and hybrid retrieval-generation of the entire chain.

[0008] Technical Solution: The present invention provides a spatial semantic enhancement method for real estate registration knowledge retrieval and generation, the method comprising the following steps:

[0009] S1. Obtain a user query request, wherein the query request contains real estate registration business and spatial information described in natural language;

[0010] S2. Perform spatial semantic parsing on the query request, extract spatial geometric and topological constraint features, ownership business features and text semantic features, and generate a joint query representation in a unified vector space;

[0011] S3. Based on the joint query representation, perform multi-level hybrid retrieval in the real estate registration multimodal knowledge base. The multi-level hybrid retrieval includes structured field matching, spatial range query and topological relationship matching based on spatial index, spatial semantic fuzzy matching based on vector similarity, and entity and association information retrieval based on graph.

[0012] S4. Construct a context-enhanced representation by fusing search results, wherein the context-enhanced representation includes ownership time-series information, spatial topological associations, and business rule constraints;

[0013] S5. The context-enhanced representation is used to input the large language model, generate a structured response, and complete the output.

[0014] Furthermore, in step S2, the spatial semantic parsing specifically includes:

[0015] Named entity recognition is used to extract real estate unit identifiers, rights holder information, and spatial relationship descriptions.

[0016] Identify implicit business constraints in query intent based on knowledge graphs in the real estate sector;

[0017] The joint query representation is generated by cross-modal alignment of textual semantic features and spatial geometric features.

[0018] Furthermore, in step S3, the multi-level hybrid retrieval specifically includes:

[0019] Perform precise field matching based on structured indexes;

[0020] Spatial range queries and topological relationship matching are performed based on spatial indexes, including R-tree indexes, quadtree indexes, or grid indexes, which are used to retrieve and query real estate units whose spatial geometry satisfies specific topological relationships.

[0021] Perform spatial semantic fuzzy matching based on vector similarity calculation;

[0022] Entity and related information retrieval is performed based on real estate knowledge graph.

[0023] Furthermore, the specific steps of performing spatial range queries and topological relationship matching based on spatial indexes include:

[0024] The spatial geometric features in the joint query representation are analyzed, including: geometric type features (point, line, surface, multiple points, multiple lines, multiple surfaces), spatial coordinate features (latitude and longitude coordinate strings or planar coordinate strings), geometric metric features (area, length, perimeter, buffer radius), and spatial reference system identifier (CGCS2000 National Geodetic Coordinate System or WGS84 coordinate system); the geometric boundaries and topological constraint types of the query range are extracted, and the geometric boundaries include the minimum bounding rectangle (MBR) quadruple (x... min , y min , x max , y max The topological constraint types include at least one of the nine basic topological relationships defined based on the 9-Intersection Model (9IM): disjoint, connected, intersecting, contained, contained, overlapping, equal, covering, and covered, and are mapped to corresponding spatial predicate operators to perform matching.

[0025] Spatial objects are quickly filtered using spatial indexes. A nine-intersection model topological relationship matrix between the spatial objects and the query geometry is calculated to verify whether the specified topological constraints are met. The nine-intersection model topological relationship matrix is ​​defined as follows:

[0026] in, Represents the interior of a spatial object. Represents the boundary of a spatial object. Represents the exterior of a spatial object. Representation of spatial objects internal, Representation of spatial objects boundary, Representation of spatial objects Externally, the value of each matrix element indicates whether the corresponding intersection is empty, with 0 indicating emptiness and 1 indicating non-emptiness.

[0027] Furthermore, the vector similarity calculation employs the Hierarchical Navigable Small World Graph (HNSW) algorithm, and its vector similarity calculation formula is as follows:

[0028]

[0029] in, For query vector, For candidate space vectors, (·) is the spatial geometric similarity function. To query the spatial geometric feature components in vector q, Candidate space vector The spatial geometric feature components, where λ is the spatial-semantic fusion weight coefficient and 0 ≤ λ ≤ 1; the spatial geometric similarity function (·) is defined as:

[0030]

[0031] in Geometric distance This is the maximum allowed distance;

[0032] The configuration parameters for the HNSW algorithm include: maximum number of connections M=16, and the size of the dynamic candidate set built during the indexing phase ef. Construction =200, dynamic candidate set size during the query phase (ef) Search =50, level decay coefficient levelMult=1 / ln(M), maximum number of levels in the graph structure maxLevel=5;

[0033] The method further includes: performing initial spatial range screening based on R-tree index, performing HNSW vector similarity fine ranking on the retained candidate set, and returning Top-K matching results; dynamically adjusting the λ value when the query contains spatial semantic keywords.

[0034] Furthermore, the entity and related information retrieval based on the real estate knowledge graph can be implemented by encoding the topological and business relationships between real estate units through graph neural networks, or it can be implemented in a graph database. It adopts a graph embedding-based traversal algorithm and supports path reasoning with more than 5 hops.

[0035] This invention also discloses a spatial semantic enhancement system for real estate registration knowledge retrieval and generation, comprising:

[0036] The query parsing module is used to obtain user query requests, perform spatial semantic parsing on the query requests, extract spatial geometric and topological constraint features, ownership business features and text semantic features, and generate a joint query representation;

[0037] The hybrid retrieval module is used to perform structured field matching, spatial range and topological relationship retrieval based on spatial index, text retrieval based on vector similarity, and entity and association information retrieval based on graph in the real estate registration multimodal knowledge base based on the joint query representation.

[0038] The context enhancement module is used to integrate search results to construct a context enhancement representation, which includes ownership time sequence information, spatial topology association and business rule constraints.

[0039] The generation management module is used to take the context-enhanced representation of the input language model, generate structured responses, and encapsulate and output them in a standardized manner.

[0040] Furthermore, the hybrid retrieval module includes:

[0041] The precise retrieval submodule is used to perform structured field matching;

[0042] The spatial retrieval submodule is used to retrieve and query real estate units whose spatial geometry satisfies specific topological relationships based on spatial indexes.

[0043] The semantic retrieval submodule is used to perform spatial semantic similarity calculation;

[0044] The graph retrieval submodule is used to perform cross-entity relationship reasoning.

[0045] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0046] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0047] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0048] This invention presents a spatial semantic enhancement method for knowledge retrieval and generation in real estate registration. This method addresses the problems of limited natural language understanding, spatial-text cross-modal fragmentation, and missing ownership information in the field of intelligent real estate registration. Based on spatial semantic enhancement retrieval and reasoning technology, it provides an efficient, accurate, and reliable intelligent processing solution for real estate registration business queries, significantly improving the quality and efficiency of business queries and possessing good application value. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention;

[0050] Figure 2 This is a schematic diagram of multi-level hybrid retrieval in an embodiment of the present invention;

[0051] Figure 3 This is a flowchart illustrating the construction of the ownership timeline in an embodiment of the present invention;

[0052] Figure 4A hierarchical architecture diagram for constructing the spatial relationship network in this embodiment of the invention;

[0053] Figure 5 This is a system structure diagram of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example

[0056] like Figure 1 As shown, this invention provides a spatial semantic enhancement method and system for generating enhanced real estate registration knowledge retrieval, comprising the following steps:

[0057] Step S1: Obtain the user's query request, which includes a natural language description and a real estate business type identifier.

[0058] Furthermore, the process of obtaining a user's query request includes the following steps:

[0059] Step S1-1: The system receives multimodal user input, including plain text queries, hand-drawn spatial sketches, and voice queries. For plain text queries, information such as the location of real estate, owner, and business type is described in natural language. For hand-drawn spatial sketches, users can draw points, lines, and polygons on the web interface, and the system automatically recognizes the spatial topological meaning of the sketch (such as "within the drawing area" or "adjacent to the drawn line"). For voice queries, the input is converted into text through an integrated speech recognition engine.

[0060] Step S1-2: The system automatically identifies the real estate business type identifier, including six categories: initial registration, transfer registration, change registration, mortgage registration, seizure registration, cancellation registration, and query. Business type identification uses a BERT text classification model. When a user enters "process transfer procedures," the system automatically maps it to the "transfer registration" type; when the user enters "check if this property is mortgaged," it is identified as the "query" type with the added "mortgage status" search intent.

[0061] Steps S1-3: Preprocess the query request, including text normalization (unifying place name spelling and standard terminology conversion), spatial reference unification (converting different coordinate systems to the CGCS2000 national geodetic coordinate system), and speech noise reduction and accent adaptation. The preprocessed query request then enters the spatial semantic parsing module.

[0062] Step S2: Perform spatial semantic parsing on the query request, extract spatial constraint features, ownership business features and text semantic features, and generate a joint query representation in a unified vector space.

[0063] Furthermore, the specific implementation process of spatial semantic parsing includes the following steps:

[0064] S2-1: Extracting real estate unit identifiers, rights holder information, and spatial relationship descriptions through named entity recognition. The BERT-BiLSTM-CRF model is used to process the query text, identifying key entities such as real estate unit number, location, rights holder name, ID number, and real estate right type. For spatial relationship descriptions, a specially trained spatial relationship extraction model is used to recognize spatial constraint expressions such as "located to the east of...", "within 500 meters of...", and "contained within...".

[0065] S2-2: Identifying Implicit Business Constraints in Query Intents Based on Real Estate Knowledge Graphs. The knowledge graph contains six core entities: real estate units, rights holders, rights certificates, mortgage registrations, seizure registrations, and registration agencies, as well as various relationship types such as "ownership," "mortgage," "seizure," "inheritance," and "located." By transforming business rules into vector representations through graph embedding technology, it can automatically identify implicit business constraints in queries, such as rights holder qualifications and rights restriction status.

[0066] S2-3: Cross-modal alignment of textual semantic features and spatial geometric features is performed. An adversarial training strategy is used to construct a cross-modal alignment discriminator, mapping the semantic vectors output by the text encoder and the geometric vectors output by the spatial encoder to a unified high-dimensional vector space. The training objective is to minimize the modality discrimination accuracy, making the discriminator unable to distinguish the source of the vectors, thereby achieving semantic alignment between different modalities.

[0067] Specifically, let the text semantic feature vector be... The spatial geometric eigenvectors are , where d is the dimension of the unified vector space. This is achieved through a mapping function. Project them into the same semantic space and calculate their cross-modal semantic similarity. :

[0068]

[0069] in, Represents the vector dot product. This represents the L2 norm.

[0070] Objective function of adversarial training Aiming to minimize the discriminator The ability to distinguish the source of vectors while maximizing the alignment effect of the generator:

[0071]

[0072] Ultimately, the joint query indicates Generated by weighted concatenation of aligned feature vectors:

[0073]

[0074] in, and For the process Feature vectors after mapping and alignment These are modal weighting coefficients, which can be dynamically adjusted based on the query intent (e.g., when spatial queries are the primary focus). Lower it). Represents the semantic feature vector of the text. The mathematical expectation is used to calculate the average probability of the discriminator on the text modality data distribution; Represents spatial geometric eigenvectors Used for calculating generators The average probability over the spatial modal data distribution.

[0075] When the text description "Zhonglou District CBD area" and the polygonal features of the CBD boundary are input simultaneously, the cosine similarity of the two types of vectors in the latent space is highly stable after training.

[0076] Step S2-4: Handling Special Data Conflict Scenarios. When the quality of a certain modality data is low (e.g., OCR recognition confidence < 0.6), the system automatically reduces the weight of that modality to 0.3, using the high-confidence modality as the dominant alignment benchmark. When there are reasonable differences in cross-modal semantics (e.g., the text describes "within the cell" but the spatial data is displayed at the cell boundary), constraint rules are added by registering the knowledge base to allow semantic coexistence in boundary cases, avoiding excessive forced alignment.

[0077] Step S2-5: Spatial query parameter standardization. The parsed spatial geometric features are converted into query parameters recognizable by the spatial index.

[0078] Bounding box extraction: Calculate the minimum bounding rectangle (MBR) of the spatial geometry to generate (x min y min x max y max A quadruple is used for range queries in an R-tree index.

[0079] Topological constraint coding: Mapping natural language descriptions such as "adjacent", "containment", and "intersection" to nine-intersection model topological predicate coding.

[0080] Step S3: Based on the joint query representation, perform multi-level hybrid retrieval in the real estate registration data resources. The retrieval includes structured field matching, spatial semantic similarity calculation, and graph expansion retrieval, such as... Figure 2 As shown.

[0081] Furthermore, the specific implementation process of multi-level hybrid retrieval includes:

[0082] Step S3-1: Perform precise field matching based on structured indexes. For fields with unique identifiers such as real estate unit number, owner's ID number, and certificate number, a B+ tree index is constructed to achieve fast retrieval with O(logn) complexity. For complex query conditions, an optimization strategy combining bitmap indexes and Bloom filters is adopted. When the user enters the complete property certificate number "Su (2026) Changzhou Real Estate Certificate No. 1234567", the system generates the following SQL statement:

[0083] SELECT * FROM property_units WHERE certificate_number = 'Su20261234567'

[0084] And quickly return accurate results.

[0085] Alternative strategy: When exact matching conditions are insufficient, use a combination query of rights holder and spatial range to improve query efficiency through composite indexes.

[0086] Step S3-2: Perform spatial range queries and topological relationship matching based on the spatial index.

[0087] Furthermore, based on the spatial constraint features in the query request, the system first performs spatial information retrieval to quickly locate the candidate set of geometric ranges:

[0088] Step S3-2-1: Spatial Range Query. The system parses the spatial geometric feature vectors in the joint query representation. Extract the query geometry object q (Supports point, line, and polygon geometry). Perform efficient range queries using R-tree spatial indexes (GiST indexes) in the PostgreSQL spatial database:

[0089] SELECT unit_id, geom

[0090] FROM real_estate_units

[0091] WHERE ST_Within(geom, ST_GeomFromText('POLYGON(...)', 4490))

[0092] OR ST_DWithin(geom, ST_GeomFromText('POINT(...)', 4490), 200);

[0093] For the query "Find all real estate units within the Zhonglou District CBD area", the system quickly filters out the set of units located within the polygon boundary using spatial indexing. .

[0094] Step S3-2-2: Topological Relationship Matching. For queries involving spatial topological relationships (such as "find adjacent to the target parcel" or "find all parcels containing a certain point"), the system performs accurate topological relationship calculation based on the nine-intersection model:

[0095] Let the query geometry be... Geometry of candidate units The topological relation matrix is The system verifies whether the cells satisfy the specified topological predicate. Only cells that satisfy the topological constraints are retained for subsequent fusion retrieval.

[0096] Step S3-2-3: Multi-scale spatial indexing strategy. The system adopts a hierarchical indexing mechanism:

[0097] Coarse filtering layer: Based on R-tree index, it quickly filters units that are clearly outside the query range;

[0098] Fine screening layer: Performs precise geometric calculations on the coarse screening results to verify complex topological relationships (such as "shared boundary length exceeding 10 meters").

[0099] Step S3-3: Perform spatial semantic fuzzy matching based on vector similarity calculation. Input the joint query representation vector into the Milvus vector database, and use the HNSW (Hierarchical Navigable Small World) approximate nearest neighbor search algorithm to construct a hybrid index structure (spatial R-tree index + vector HNSW index). The HNSW algorithm employs an asymmetric fusion strategy, with the specific formula as follows:

[0100]

[0101] in, The semantic similarity (cosine similarity) is used as a basic retrieval capability, with a fixed weight of 1. Let be the spatial geometric similarity function, used as an enhancement term; λ is the spatial-semantic fusion weight coefficient and 0 ≤ λ ≤ 1; where the spatial geometric similarity function Defined as:

[0102]

[0103] in This is the geometric distance (such as Euclidean distance). This is the maximum permissible distance.

[0104] When the query does not contain spatial semantic keywords, λ=0, and pure semantic retrieval is performed to maintain the basic capabilities of semantic retrieval; when the query contains spatial semantic keywords (such as "nearby" or "surrounding area"), the system dynamically adjusts λ>0 to enhance the contribution of spatial similarity without weakening the basic weight of semantic similarity.

[0105] Step S3-4: Perform entity and association information retrieval based on the real estate knowledge graph. Perform multi-hop Cypher queries in the Neo4j graph database, using a graph embedding-based traversal algorithm that supports path reasoning with more than 5 hops.

[0106] MATCH (u: unit) - [: rights holder association *1..3] -> (o: rights holder)

[0107] MATCH (u)-[:spatial relation{type:"contains"|"adjacent"}]->(p:spatial object)

[0108] MATCH (u)-[:Business Association{Type:"Mortgage"|"Seizure"}]->(r:Restriction Record)

[0109] RETURN DISTINCT u LIMIT 30

[0110] Extended operations include: rights holder association extension (finding all units under the same rights holder), spatial topology extension (discovering adjacent / contained units), and business association extension (tracking the same transaction sequence).

[0111] Step S3-5: Retrieval Result Fusion and Optimization. The intersection of multi-level hybrid retrieval results is taken as the final candidate set, and a credibility arbitration mechanism is adopted: specifically, for a candidate result i, its comprehensive credibility score C i The result is calculated by the weighted sum of the four search results:

[0112]

[0113] in: , , , , representing the normalized scores (range [0,1]) of candidate i in spatial retrieval matching (spatial overlap rate or distance inverse normalization), structured field matching, vector similarity-based text retrieval, and graph-based entity and association information retrieval.

[0114] , , , These are the weight coefficients for the four-way retrieval, satisfying... and The system dynamically adjusts the weighting coefficients of each retrieval path based on the intent characteristics output by the query parsing module.

[0115] (1) If the query contains spatial semantic keywords (such as "nearby" or "surrounding area"), the spatial retrieval weight is enhanced:

[0116]

[0117] (2) If the query contains attribute keywords (such as "area" or "price"), the attribute retrieval weight is enhanced:

[0118]

[0119] in, , Indicator variables for intent characteristics , , For adjustment coefficient ( ).

[0120] Finally, the system returned... Result set Score based on overall credibility Sort in descending order.

[0121] If the final number of candidates is still too large (>100), DBSCAN clustering will be automatically executed (ε=0.15, min_samples=5), and the central units of each category will be displayed first after clustering by spatial distribution.

[0122] Step S4: Construct a context-enhanced representation by fusing search results. The context-enhanced representation includes ownership time-series information, spatial topology associations, and business rule constraints.

[0123] Furthermore, the specific implementation process of constructing context-enhanced representations includes:

[0124] Step S4-1: Encoding Spatial Retrieval Results. Receive the spatial candidate set returned in step S3-2 and construct a spatial context representation:

[0125] Spatial matching metric: For a class of queries in the buffer, calculate the normalized distance score. ,in The actual distance between the element geometry and the query point. The buffer radius;

[0126] Topological relation structuring: The calculation results of the nine-intersection model are encoded into relation triples (unit A, topological relation type, unit B), such as (Su2026XXXX, adjacent, Su2026YYYY), and context-enhanced representation is injected;

[0127] Spatial conflict detection: If no unit satisfies the topological constraints in the search results (e.g., the user queries "contained in a certain area" but the geometric calculation shows only partial overlap), a spatial conflict marker is generated for the large model to generate an interpretive response (e.g., "The area you specified partially exceeds the boundary of the registered unit, and the actual overlapping area is XX square meters").

[0128] Step S4-2: Construct the ownership timeline, such as... Figure 3 As shown, ownership change information is extracted from scattered registration records, and ownership change events are represented by a time-series graph model. Nodes represent registration events (initial registration, transfer registration, change registration, mortgage registration, seizure registration, cancellation registration), and edges represent time-series relationships. The system automatically identifies key event nodes to display the entire process of ownership evolution, aiding in understanding the historical changes in the status of real estate units and improving the completeness of information in question-and-answer sessions.

[0129] Step S4-3: Constructing a spatial relationship network, such as... Figure 4 As shown, a multi-level spatial correlation network is generated based on the nine-intersection topology model.

[0130] The nine-intersection model is a standard mathematical model for describing the topological relationship between two spatial objects a and b. It defines the relationship by examining nine binary intersections (empty / non-empty) between the object's interior (Ia), boundary (Ba), and exterior (Ea).

[0131] Specifically, its relation matrix is ​​defined as follows:

[0132] In this matrix, the value of each element indicates whether the corresponding intersection is empty (0 for empty, 1 for non-empty). This matrix allows for the formal determination of topological relationships such as "adjacent," "containment," "intersection," and "disjoint." For example, the matrix characteristics of an "adjacent" relationship are: and .

[0133] Based on this model, the system constructs relationship matrices for spatial objects at different levels, such as parcels, buildings, and rooms, and encodes them as edge attributes in a graph structure, forming a computable spatial relationship network. At the parcel level, an adjacency relationship network is built to record adjacent, contained, and overlapping relationships between parcels; at the building level, a hierarchical structure of "building-unit-household" is displayed; and within a unit, the connectivity relationships between rooms are recorded. The network construction uses the NetworkX graph computing framework, supporting efficient topological relationship queries and path analysis, providing structured relationship support for spatial semantic question answering.

[0134] Step S4-4: Business Rule Perception and Interpretation. The system has a built-in structured knowledge base of real estate registration regulations, including the "Provisional Regulations on Real Estate Registration" and its implementing rules. The system automatically extracts business rule clauses relevant to the current query and retrieval results through a rule parsing unit. Further, the system performs semantic parsing on the rules, identifying the conditions, subjects, behaviors, and constraints within the rules, and matching and associating them with the entities, states, and operational intentions in the query. The matched rules are injected into the context-enhanced representation in a structured semantic representation to support the question-and-answer system's understanding of business scenarios, explanation of compliance, and description of operational feasibility, thereby significantly enhancing the accuracy and interpretability of the answers.

[0135] Steps S4-5: Information Completeness Optimization and Credibility Fusion. The system automatically detects missing information in the context and initiates a supplementary retrieval process to retrieve information from relevant databases to improve the completeness of the question-and-answer session. Weights are dynamically assigned based on the reliability of the data sources (e.g., register weight 0.1, land ownership survey results weight 0.9). When discrepancies exist between information from different sources, DS evidence theory is used for fusion calculation, outputting a unified information description after fusion to improve the consistency and credibility of the question-and-answer answers.

[0136] Step S5: Input the context-enhanced representation into the large language model and output structured response information containing real estate unit location, ownership status resolution, and business rule interpretation.

[0137] Furthermore, the specific implementation process of language large-scale model reasoning includes:

[0138] Step S5-1: Entity Recognition and Information Aggregation. Based on context-enhanced representation, the system accurately identifies target real estate units and confirms their unique identifiers and core attributes. The system employs an attention mechanism to focus on high-weighted evidence. When multiple potentially relevant units exist, disambiguation and selection are performed based on the evidence fusion results to ensure the accuracy of the response.

[0139] Step S5-2: State Resolution and Correlation Analysis. A comprehensive analysis is performed on the target unit's current ownership status, historical evolution chain, spatial topological relationships, and related business rules and clauses. The system generates a multi-dimensional state feature summary to clearly and structurally describe the unit's status and explain the meaning of various states (such as mortgage and seizure) and their possible associated business rule backgrounds.

[0140] Step S5-3: Structured Response Generation. Based on the aforementioned analysis, the system encapsulates the answers generated by the language model into a unified JSON format. This format strictly adheres to a predefined response pattern, ensuring complete information, clear semantics, and machine readability, facilitating front-end page parsing and display. Core content includes:

[0141] Unit summary information: basic attributes such as unit identifier, location, area, and purpose.

[0142] A comprehensive overview of ownership status, including the current rights holder, type of rights, whether there are any restrictions on the rights, and details thereof.

[0143] Rule association explanation: Summarize and explain the business rules involved or associated with the status in the query, clarifying their contextual meaning;

[0144] Spatial positioning information: includes the precise coordinates of the unit and its relative position to the queried geometry.

[0145] This invention also provides a spatial semantic enhancement system for retrieving, reasoning, and generating real estate registration data, such as... Figure 5 As shown, it includes the following modules:

[0146] The data acquisition module is used to receive user query requests, supports multimodal input (text, hand-drawn sketches, voice), and performs preprocessing and standardization.

[0147] The query parsing module is used to perform spatial semantic parsing in step S2, including named entity recognition, knowledge graph query and cross-modal alignment, and to generate a joint query representation vector.

[0148] The hybrid retrieval module is used to perform the multi-level hybrid retrieval in step S3, including a precise retrieval submodule, a spatial retrieval submodule, a semantic retrieval submodule, and a graph retrieval submodule.

[0149] The context enhancement module is used to perform the context enhancement representation construction in step S4. It integrates the time-series database TimescaleDB to manage the ownership time axis, the graph computing engine NetworkX to process the spatial relationship network, and injects the interpretation of related business rules.

[0150] The generation management module is used to perform language large model inference in step S5, deploy the domain-adaptive fine-tuned Qwen-32B model, use 4-bit quantization technology to compress the model size, generate structured response information including unit location, state parsing and rule interpretation, and record the question-and-answer full-link audit log to ensure that the process is traceable and the results are verifiable.

[0151] Embodiments of this application may be provided as methods or computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0152] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A spatial semantic enhancement method for real estate registration knowledge retrieval enhancement, characterized in that, The method includes the following steps: S1. Obtain a user query request, wherein the query request contains real estate registration business and spatial information described in natural language; S2. Perform spatial semantic parsing on the query request, extract spatial geometric and topological constraint features, ownership business features and text semantic features, and generate a joint query representation in a unified vector space; S3. Based on the joint query representation, perform multi-level hybrid retrieval in the real estate registration multimodal knowledge base. The multi-level hybrid retrieval includes structured field matching, spatial range query and topological relationship matching based on spatial index, spatial semantic fuzzy matching based on vector similarity, and entity and association information retrieval based on graph. S4. Construct a context-enhanced representation by fusing search results, wherein the context-enhanced representation includes ownership time-series information, spatial topological associations, and business rule constraints; S5. The context-enhanced representation is used to input the large language model, generate a structured response, and complete the output.

2. The method according to claim 1, characterized in that: In step S2, the spatial semantic parsing specifically includes: Named entity recognition is used to extract real estate unit identifiers, rights holder information, and spatial relationship descriptions. Identify implicit business constraints in query intent based on knowledge graphs in the real estate sector; The joint query representation is generated by cross-modal alignment of textual semantic features and spatial geometric features.

3. The method according to claim 1, characterized in that: In step S3, the multi-level hybrid retrieval specifically includes: Perform precise field matching based on structured indexes; Spatial range queries and topological relationship matching are performed based on spatial indexes, including R-tree indexes, quadtree indexes, or grid indexes, which are used to retrieve and query real estate units whose spatial geometry satisfies specific topological relationships. Perform spatial semantic fuzzy matching based on vector similarity calculation; Entity and related information retrieval is performed based on real estate knowledge graph.

4. The method according to claim 3, characterized in that: The specific steps of performing spatial range queries and topological relationship matching based on spatial indexes include: The spatial geometric features in the joint query representation are analyzed, including geometric type features, spatial coordinate features, geometric metric features, and spatial reference frame identifiers. The geometric boundaries and topological constraint types of the query range are extracted, whereby the geometric boundaries include the minimum bounding rectangle (MBR) quadruple (x...). min , y min , x max , y max The topological constraint types include at least one of the nine basic topological relationships defined based on the nine-intersection model: disjoint, connected, intersecting, contained, included, overlapping, equal, covering, and covered, and are mapped to corresponding spatial predicate operators to perform matching. Spatial objects are quickly filtered using spatial indexes. A nine-intersection model topological relationship matrix between the spatial objects and the query geometry is calculated to verify whether the specified topological constraints are met. The nine-intersection model topological relationship matrix is ​​defined as follows: Where Ia represents the interior of spatial object a, Ba represents the boundary of spatial object a, and Ea represents the exterior of spatial object a. Represents the interior of spatial object b. Represents the boundary of spatial object b. The matrix represents the space outside object b. The value of each matrix element indicates whether the corresponding intersection is empty, with 0 indicating empty and 1 indicating non-empty.

5. The method according to claim 3, characterized in that: The vector similarity calculation employs the Hierarchical Navigable Small World Graph (HNSW) algorithm, and its vector similarity calculation formula is as follows: ； in, For query vector, For candidate space vectors, (·) is the spatial geometric similarity function. To query the spatial geometric feature components in vector q, Candidate space vector The spatial geometric feature components, where λ is the spatial-semantic fusion weight coefficient and 0 ≤ λ ≤ 1; the spatial geometric similarity function (·) is defined as: ； in Geometric distance This is the maximum allowed distance; The configuration parameters for the HNSW algorithm include: maximum number of connections M=16, and the size of the dynamic candidate set built during the indexing phase ef. Construction =200, dynamic candidate set size during the query phase (ef) Search =50, level decay coefficient levelMult=1 / ln(M), maximum number of levels in the graph structure maxLevel=5; The method further includes: performing initial spatial range screening based on R-tree index, performing HNSW vector similarity fine ranking on the retained candidate set, and returning Top-K matching results; dynamically adjusting the λ value when the query contains spatial semantic keywords.

6. The method according to claim 3, characterized in that: The entity and related information retrieval based on real estate knowledge graph can be performed by encoding the topological and business relationships between real estate units through graph neural networks, or it can be implemented in a graph database. It adopts a graph embedding-based traversal algorithm and supports path reasoning with more than 5 hops.

7. A spatial semantic enhancement system for real estate registration knowledge retrieval and generation, used to implement the method as described in claim 1, characterized in that, include: The query parsing module is used to obtain user query requests, perform spatial semantic parsing on the query requests, extract spatial geometric and topological constraint features, ownership business features and text semantic features, and generate a joint query representation; The hybrid retrieval module is used to perform structured field matching, spatial range and topological relationship retrieval based on spatial index, text retrieval based on vector similarity, and entity and association information retrieval based on graph in the real estate registration multimodal knowledge base based on the joint query representation. The context enhancement module is used to integrate search results to construct a context enhancement representation, which includes ownership time sequence information, spatial topology association and business rule constraints. The generation management module is used to take the context-enhanced representation of the input language model, generate structured responses, and encapsulate and output them in a standardized manner.

8. The system according to claim 7, characterized in that: The hybrid retrieval module includes: The precise retrieval submodule is used to perform structured field matching; The spatial retrieval submodule is used to retrieve and query real estate units whose spatial geometry satisfies specific topological relationships based on spatial indexes. The semantic retrieval submodule is used to perform spatial semantic similarity calculation; The graph retrieval submodule is used to perform cross-entity relationship reasoning.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images