Address retrieval method and system based on space-time semantics

Abstract

The invention discloses an address retrieval method based on spatio-temporal semantics, which comprises the following steps of: aiming at address description information input by a client, combining a BERT model and a BiGRU gating mechanism, and converting an oral and fuzzy address text into a structured semantic unit which can be understood by a machine by a unified structured generation normal form; and then, for time class, space class and scene class semantic units output by word segmentation and semantic classification, a corresponding space-time knowledge base is called to carry out semantic analysis, information expansion and responsibility association, space, time and scene analysis results are fused, and finally, an optimal retrieval sequence is presented to a user for selection. According to the method, accurate and reliable address semantic understanding and information retrieval capability can be provided in key scenes such as city management, emergency command and livelihood service.

Description

Technical Field

[0001] This invention relates to the field of geographic information data processing technology, and in particular to an address retrieval method and system based on spatiotemporal semantics. Background Technology

[0002] Mega-cities (such as Beijing, Shanghai, Guangzhou, Chongqing, and Wuhan) are characterized by dense populations, complex spatial structures, and dynamic and ever-changing governance elements. Their core urban governance needs are to achieve "refined, intelligent, and real-time" control of these elements. Address retrieval, as a spatiotemporal link connecting "governance events - spatial locations - responsible entities - disposal resources," is facing increasingly prominent technological bottlenecks. Specifically, this manifests in the high spatiotemporal semantic complexity of governance events, the strong multi-source nature of governance elements, and the high demand for spatiotemporal correlation in cross-departmental collaborative governance. Traditional address retrieval relies heavily on structured address string matching (such as keyword-based fuzzy queries and regular expression matching). However, in practical applications, user-input addresses often suffer from unstructured, semantically ambiguous, and missing or mixed spatiotemporal information. For example, a user querying "fresh food supermarket near the Olympic Sports Center" includes not only the spatial information of "Olympic Sports Center" but also the semantic attribute of "fresh food supermarket." Traditional retrieval methods struggle to simultaneously parse and integrate multi-dimensional information, resulting in poor relevance and a high false negative rate in search results.

[0003] In the context of governance in megacities, a common approach is address retrieval based on spatial modeling. This involves associating geographic entities with location descriptions and using spatial reasoning techniques (such as topological reasoning and distance reasoning) to extract and match spatial information from user queries. The core logic of this approach revolves around "structured processing of address information and matching of main keywords." Its advantages include simple implementation and fast response time. However, it suffers from drawbacks such as poor handling of fuzzy addresses and inability to disambiguate semantic ambiguities. Furthermore, grassroots governance personnel and the public often use colloquial or simplified address expressions, leading to frequent instances of mismatches or inability to match, thus limiting the applicability of this approach in such scenarios.

[0004] Existing address retrieval based on spatial modeling primarily involves defining regular expression rules for address formats, validating and extracting the structure of user-input addresses, and then searching an address database based on the extracted structured fields to match and locate the target address. For fuzzy addresses containing descriptions like "nearby," spatial buffer analysis is triggered to retrieve similar addresses, followed by a simple ranking based on recommended addresses, and finally, the user interacts to select the desired address. However, for addresses containing directional terms, descriptions along routes plus distance descriptions, and colloquial expressions, such as "YY 100 meters east of XX," generally only a single keyword can be matched, failing to accurately pinpoint the spatial relationship between the two, and the address cannot be effectively identified and extracted. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides an address retrieval method and system based on spatiotemporal semantics. The technical problem to be solved by this invention is achieved through the following technical solution: The first aspect of this invention provides an address retrieval method based on spatiotemporal semantics, comprising the following steps: The address description information is sequentially input into the BERT model, the bidirectional LSTM model, and the CRF model for word segmentation, and the output is multiple address entities as well as the latitude and longitude coordinates and type labels of the address entities; Multiple address entities are input into a semantic processing model for semantic classification to obtain spatial address entities, temporal address entities, and address-type address entities. The standard topology and standard direction information of the spatial address entity are determined based on the spatial semantic knowledge base, and the time distance information of the time address entity from the current time is determined based on the time knowledge base. The address subject and address object of the address class address entity are determined based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information; The accurate address subject is determined based on the latitude and longitude coordinates of the address subject; The accurate address object is determined based on the address object and the entity knowledge base; Based on the accurate address subject, the time distance, and the preset association conditions, candidate address object information is determined from the accurate address objects; The address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order based on the address score.

[0006] In one embodiment of the present invention, determining the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information includes: The standard address entity of the address class address entity is determined by entity linking based on the entity knowledge base; The association information of standard address entities is determined based on the preset address knowledge graph; The address subject and address object are determined based on the association information of the standard address entity, the standard topology information, and the standard direction information.

[0007] In one embodiment of the present invention, determining the accurate address object based on the address object and the entity knowledge base includes: Based on the address object, the corresponding accurate address object is determined in the entity knowledge base.

[0008] In one embodiment of the present invention, a comprehensive address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order according to the comprehensive address score, including: The straight-line distance between the candidate address object information and the accurate address subject is calculated using the Haversine algorithm. The address score is obtained by weighting the straight-line distance according to the preset semantic weight and the preset time adaptability weight. The candidate address object information is sorted in descending order based on the comprehensive address score.

[0009] A second aspect of this invention provides an address retrieval system based on spatiotemporal semantics, comprising: The word segmentation module is used to input address description information into the BERT model, bidirectional LSTM model and CRF model in sequence for word segmentation, and output multiple address entities as well as the latitude and longitude coordinates and type labels of the address entities; The classification module is used to input multiple address entities into the semantic processing model for semantic classification, and obtain spatial address entities, time-based address entities, and address-based address entities. The first determining module is used to determine the standard topology information and standard direction information of the spatial address entity based on the spatial semantic knowledge base, and to determine the time distance information of the time address entity from the current time based on the time knowledge base. The second determining module is used to determine the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information and the standard direction information; The third determining module is used to determine the accurate address subject based on the latitude and longitude coordinates of the address subject; The fourth determining module is used to determine the accurate address object based on the address object and the entity knowledge base; The fifth determining module is used to determine candidate address object information from the accurate address objects based on the accurate address subject, the time distance, and preset association conditions; The sorting module is used to calculate a comprehensive address score based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight, and to sort the candidate address object information in descending order based on the comprehensive address score.

[0010] In one embodiment of the present invention, determining the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information includes: The standard address entity of the address class address entity is determined by entity linking based on the entity knowledge base; The association information of standard address entities is determined based on the preset address knowledge graph; The address subject and address object are determined based on the association information of the standard address entity, the standard topology information, and the standard direction information.

[0011] In one embodiment of the present invention, determining the accurate address object based on the address object and the entity knowledge base includes: Based on the address object, the corresponding accurate address object is determined in the entity knowledge base.

[0012] In one embodiment of the present invention, a comprehensive address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order according to the comprehensive address score, including: The straight-line distance between the candidate address object information and the accurate address subject is calculated using the Haversine algorithm. The address score is obtained by weighting the straight-line distance according to the preset semantic weight and the preset time adaptability weight. The candidate address object information is sorted in descending order based on the comprehensive address score.

[0013] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement an address retrieval method based on spatiotemporal semantics provided in the first aspect of the present invention.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an address retrieval method based on spatiotemporal semantics provided in the first aspect of the present invention.

[0015] The beneficial effects of this invention are: This invention employs an end-to-end model architecture of "bidirectional LSTM + CRF word segmentation + pre-trained BERT classification." First, it accurately identifies proper noun boundaries (such as place names, POIs, roads, etc.) in address text using a bidirectional LSTM and CRF layer. Then, it fine-tunes a pre-trained BERT model based on massive contextualized labeled corpora (including urban governance event data), effectively capturing contextual semantic relationships using a multi-head attention mechanism. This achieves comprehensive and fine-grained recognition of multi-dimensional semantics in address text, including spatial, temporal, and scene dimensions, enabling accurate matching of multi-level tags. This clarifies the geographic entity category of each part and its location within the address structure. It can accurately locate the address subject, and combined with the core spatial entity range of the address subject, retrieves objects conforming to spatial association rules within a time range using a spatial index. These objects are then sorted by spatial distance, and multiple optimal sequences are output for user selection. This invention effectively solves problems such as the difficulty in recognizing ambiguous and colloquial addresses, inaccurate retrieval, and fragmented information in the governance of megacities, providing accurate and reliable address semantic understanding and information retrieval capabilities for key scenarios such as urban management, emergency command, and public services.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating an address retrieval method based on spatiotemporal semantics provided in an embodiment of the present invention; Figure 2 This is a block diagram of an address retrieval system based on spatiotemporal semantics provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0020] like Figure 1 As shown, the first aspect of this invention provides an address retrieval method based on spatiotemporal semantics, comprising the following steps: Step 11: Input the address description information into the BERT model, bidirectional LSTM model and CRF model in sequence for word segmentation, and output multiple address entities as well as the latitude and longitude coordinates and type labels corresponding to the address entities.

[0021] Step 12: Input multiple address entities into the semantic processing model for semantic classification to obtain spatial address entities, time address entities, and address-type address entities.

[0022] Step 13: Determine the standard topology and standard direction information of spatial address entities based on the spatial semantic knowledge base, and determine the time distance information of time-based address entities from the current time based on the time knowledge base.

[0023] Step 14: Determine the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, standard topology information, and standard direction information.

[0024] Step 15: Determine the accurate address subject based on the latitude and longitude coordinates of the address subject.

[0025] Step 16: Determine the accurate address object based on the address object and entity knowledge base.

[0026] Step 17: Determine candidate address object information from the accurate address objects based on the accurate address subject, time distance, and preset association conditions.

[0027] Step 18: Calculate the comprehensive address score based on the straight-line distance between the accurate address subject and the candidate address object information, the preset semantic weight, and the preset time adaptability weight, and sort the candidate address object information in descending order according to the comprehensive address score.

[0028] In this embodiment, an end-to-end model architecture of "bidirectional LSTM + CRF word segmentation + pre-trained BERT classification" is adopted. First, the bidirectional LSTM and CRF layers accurately identify the boundaries of proper nouns (such as place names, POIs, roads, etc.) in address text. Then, the pre-trained BERT model is fine-tuned based on massive scene-specific labeled corpus (including urban governance event data). A multi-head attention mechanism is used to effectively capture contextual semantic relationships, achieving comprehensive and fine-grained recognition of multi-dimensional semantics in address text, including spatial, temporal, and scene dimensions. This enables accurate matching of multi-level tags, clarifying the geographical entity category of each part and its location information within the address structure. It can accurately locate the address subject, and combined with the core spatial entity range of the address subject, retrieves objects that conform to spatial association rules within a time range through spatial indexing, sorts them according to spatial distance, and finally outputs multiple optimal sequences for the user to choose from. This invention effectively solves the problems of fuzzy addresses, difficulty in recognizing colloquial addresses, inaccurate retrieval, and fragmented information in the governance of megacities, providing accurate and reliable address semantic understanding and information retrieval capabilities for key scenarios such as urban management, emergency command, and public services.

[0029] Based on the first aspect of the present invention, the second aspect of the present invention provides a more detailed description of an address retrieval method based on spatiotemporal semantics. The second aspect of the present invention provides an address retrieval method based on spatiotemporal semantics, comprising the following steps: Step 21: The input address information is processed by the BERT model through the Transformer to capture the semantics of the context and output a feature vector containing deep semantics. The LSTM receives the feature vector from the BERT and captures the sequential dependencies between characters through a bidirectional loop structure (converting characters into meaningful phrases and assigning them labels). The CRF performs verification and optimization on the basis of the LSTM to divide the phrases and labels.

[0030] In this step, the address description information is a natural language description, which is the text input by the user. To address potential issues such as redundant information, inconsistent formatting, synonyms, abbreviations and full names, and differences between ancient and modern names in the user-input address description information text, the address description information is filtered and preprocessed according to preprocessing mapping rules before being input into the BERT model. The Transformer architecture in BERT gives higher weights to important tokens in the address. By jointly training tasks such as entity recognition and relation extraction, and utilizing the semantic associations between tasks (such as the correctness of "relationship" depending on "entity"), cross-validation and filtering of unreasonable noise annotations are performed. Redundant expressions such as "I", "this", and "that" are automatically removed, core address elements are retained and address completion is performed. Dropout and L2 regularization algorithms are used to reduce the model's overfitting to noise, and finally the natural language address description information is transformed into a standardized address expression.

[0031] The standardized address representation is input into a bidirectional LSTM model, which outputs the probability distribution of address entities and basic labels. This probability distribution is then input into a CRF model for word segmentation, yielding segmented address entities and their corresponding type labels. Type labels include, but are not limited to, "administrative division," "POI," "road," and "address." During word segmentation, the CRF model first injects an "address entity type library" (containing types such as "administrative division," "POI," "road," and "address") into the pre-trained model. Secondly, during word segmentation, the model dynamically adjusts the segmentation boundaries through entity type prediction (e.g., determining that "XX District Social Security Center" belongs to "POI - Government Services") to ensure that complete address entities are not split. Finally, the segmentation result of "address entity - type label" is output, overcoming the low accuracy of traditional word segmentation tools in recognizing proper nouns such as place names and POIs.

[0032] Step 22: Input multiple address entities into the semantic processing model for semantic classification to obtain spatial address entities, time address entities, and address-type address entities.

[0033] In this step, a multi-level label classification method driven by prompts and zero-shot classification is employed. First, a prompt template specific to address semantics is designed (e.g., "Which category does {address entity} belong to? Options: Time / Spatial / Address, Subclass: {Specific Subclass}"). Second, the word segmentation results (address entities) from the previous step are concatenated with the prompt template and input into the PaddleUIE model. Based on its pre-trained semantic understanding capabilities, the model outputs classification results through "option matching - probability calculation" (e.g., "tomorrow morning" → Time category - relative time, "XX District Social Security Center" → Address category - POI - Government Services). Finally, a confidence calibration algorithm triggers manual verification for results with low confidence (<90%) to ensure classification accuracy. The classification results are three categories of address entities: Time, Spatial, and Address. Each category (Spatial, Time, and Address) contains at least one address entity.

[0034] Step 23: Determine the standard topology information and standard direction information of spatial address entities based on the spatial semantic knowledge base, and determine the time distance information of time address entities from the current time based on the time knowledge base.

[0035] In this step, a spatial semantic knowledge base is pre-constructed. Topological relationships in the spatial semantic knowledge base can be categorized as completely equal (equal, identical, equivalent, etc.) and intersecting (intersection, entrance, end, etc.). Orientation relationships can be specified down to a concrete angular range (e.g., northeast, southeast, etc.), thus forming a spatial semantic knowledge base (orientation knowledge base, spatial topology knowledge base). For example, the spatial semantic knowledge base is shown in Table 1. Table 1 Spatial address entities are classified according to the mapping relationship of the spatial semantic knowledge base, that is, spatial address entities are mapped to standard topological information and standard directional information. Spatial address entities correspond to the natural language description part in Table 1. Based on the spatial semantic knowledge base, a clear semantic representation is defined for each type of spatial relationship. Vector representation and natural language description are used to clearly characterize the connotation and characteristics of the relationship, realizing spatial entity localization and spatial relationship quantification. Combining with the existing MCP framework, the spatial analysis capability and orientation calculation capability are semantically converted into spatial MCP capability. Furthermore, based on entity and relationship recognition, the spatial semantic conversion capability is fully utilized to quickly realize semantic spatial understanding.

[0036] The time-based address entity is converted into a description of a time range based on the time knowledge base. An example of the time knowledge base is shown in Table 2. Table 2 Step 24: Determine the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, standard topology information, and standard direction information.

[0037] Step 24 includes steps 241-242: Step 241: Determine the standard address entity of the address class address entity through entity links based on the entity knowledge base.

[0038] In this step, the entity knowledge base contains a large number of databases or knowledge graphs of standard geographic entities and their related information, such as POI databases, Gaode / Baidu map data, Wikidata, OpenStreetMap, and custom address databases. Address-related terms (address-type address entities) are precisely matched with standard entities in the entity knowledge base through entity links to obtain the standard address entities corresponding to the address-type address entities.

[0039] Step 242: Determine the association information of standard address entities based on the preset address knowledge graph.

[0040] Based on accurate standard address entities, corresponding entities, related entities, relationships, and entity attributes can be found in the preset address knowledge graph.

[0041] Step 243: Determine the address subject and address object based on the association information, standard topology information and standard direction information of the standard address entity.

[0042] In this step, the address subject is the main element in the address entered by the user that determines the destination, and the address object is the destination, that is, the target that the user needs. For example, if the B convenience store is opposite the north gate of shopping mall A, then shopping mall A is the address subject, the entrance of the shopping mall is the address subject, and the B convenience store is the user's target, that is, the address object.

[0043] For example, in the word segmentation process, the category label for "A shopping mall" is determined as address, the category label for "opposite" is determined as location, and the category label for "convenience store" is determined as address. At the same time, during the training process, some large POI categories such as "shopping mall", "school" and "hospital" are trained as large addresses with core positioning significance (i.e., subjects), while some small POIs such as "pharmacy", "convenience store" and "noodle shop" are trained as small addresses with auxiliary positioning significance (i.e., objects).

[0044] Here, the hierarchical relationship between standard address entities can be determined based on the association information between them, and the address subject and address object in the address can be determined based on the standard topology information and standard direction information.

[0045] Step 25: Determine the accurate address subject based on the latitude and longitude coordinates of the address subject.

[0046] In step 21, the latitude and longitude coordinates of the address entity are obtained. Both the address subject and the address object are address entities. Accordingly, the latitude and longitude coordinates of the address subject can be obtained, and the accurate address subject can be obtained.

[0047] Step 26: Determine the accurate address object based on the address object and entity knowledge base.

[0048] Specifically, the accurate address object is determined in the entity knowledge base based on the address object.

[0049] Specifically, firstly, multiple entities are matched in the entity knowledge base based on the address object. Then, the multiple entities are enhanced to have a stronger perception ability of geographically relevant features through the multi-head attention mechanism in the Transformer model. When calculating the attention weights, in addition to considering the conventional semantic similarity of address text, geographical hierarchical similarity and spatial coordinate distance factors are additionally introduced to determine the weight allocation and score calculation.

[0050] Calculation example: Suppose the original semantic similarity scores of words i and j are obtained by conventional dot product calculation. Geographic similarity score Spatial coordinate distance score is calculated by comparing the hierarchical encoding vectors of the two (e.g., using methods such as cosine similarity). Based on the latitude and longitude coordinates of the two objects (the closer the distance, the higher the score), the comprehensive attention weighting is calculated as follows: λ1 and λ2 are adjustable weight parameters used to balance the influence of different factors on attention weight.

[0051] For example, for two place name entities, if they are geographically related (such as a municipal administrative region belonging to a provincial administrative region) or are close in spatial coordinates, they will be given corresponding weights when calculating attention weights, so that the model will be more inclined to focus on geographically related information when processing text.

[0052] Step 27: Determine multiple candidate address object information from the accurate address objects based on the accurate address subject, time distance, and preset association conditions.

[0053] By combining the core spatial entity range of the accurate address subject and determining the timestamp in the user's address description information based on time distance (e.g., the exact arrival time within one hour), the system retrieves objects within this time range that meet preset association conditions. Then, an entity recognition filtering module removes objects that do not conform to the time rules. Finally, it outputs a list of candidate address object information after time-space collaborative filtering. The preset association conditions can be a preset range; for example, searching for convenience stores within 1000 meters of shopping mall A will yield multiple convenience stores as candidate address object information.

[0054] Step 28: Calculate the comprehensive address score based on the straight-line distance between the accurate address subject and each candidate address object information, the preset semantic weight, and the preset time adaptability weight, and sort the candidate address object information in descending order according to the comprehensive address score.

[0055] Step 28 includes steps 281-283: Step 281: Calculate the straight-line distance between the candidate address object information and the accurate address subject using the Haversine algorithm.

[0056] In this step, the latitude and longitude of the core spatial entity of the accurate address subject are used as the reference point to calculate the straight-line distance between the latitude and longitude coordinates in the candidate address object information and the reference point.

[0057] Step 282: Calculate the overall address score by weighting the straight-line distance according to the preset semantic weight and preset time adaptability weight.

[0058] In this step, the straight-line distance is weighted and calculated. For example, in an "emergency scenario", the weight of distance within 500 meters is 100%, and the weight of 500-1000 meters is 60%. The preset time adaptability weight can be, for example, whether the business is currently open. If it is open, the corresponding weight is higher. The two weights are multiplied by the straight-line distance respectively and then added together to obtain the address comprehensive score.

[0059] Step 283: Sort the candidate address object information in descending order according to the address comprehensive score.

[0060] Finally, the data is sorted in descending order of the overall score, and objects that are "close in distance but do not match in time" are removed using the anomaly detection module.

[0061] This invention is based on a contextualized multi-level tagging system for semantic classification. The protection focuses on the core innovations in the "semantic classification and annotation" stage, encompassing the design method of the contextualized multi-level tagging system, the deep learning-driven semantic classification model architecture, and the dynamic optimization mechanism. This invention employs a multi-source fusion spatiotemporal knowledge base construction method, including three steps: data cleaning (filtering invalid characters, correcting address typos and traditional Chinese characters), semantic mapping (unifying the semantic representation of multi-source data into standard knowledge base tags), and spatiotemporal alignment (associating data from different data sources through spatial encoding, such as aligning POI addresses with administrative division codes). During the fusion process, high-quality data is selected through "data quality scoring" (based on completeness, accuracy, and timeliness) to ensure the quality of the knowledge base data. This invention uses an interactive method between semantic parsing and the knowledge base in address retrieval. After the semantic classification and annotation module outputs multi-level tags, it automatically matches the corresponding knowledge base interface based on the tag type (e.g., spatial tags call the location interface of the spatially related knowledge base). Through the "tag encoding - knowledge base index" mapping relationship, target data is quickly located, achieving a second-level query response.

[0062] like Figure 2 As shown, a third aspect of the present invention provides an address retrieval system based on spatiotemporal semantics, comprising: The word segmentation module 31 is used to input the address description information into the BERT model, the bidirectional LSTM model and the CRF model in sequence for word segmentation, and output multiple address entities as well as the latitude and longitude coordinates and type labels corresponding to the address entities. The classification module 32 is used to input multiple address entities into the semantic processing model for semantic classification, and obtain spatial address entities, time address entities, and address-type address entities. The first determining module 33 is used to determine the standard topology information and standard direction information of spatial address entities based on the spatial semantic knowledge base, and to determine the time distance information of time-class address entities from the current time based on the time knowledge base. The second determining module 34 is used to determine the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information and the standard direction information; The third determining module 35 is used to determine the accurate address subject based on the latitude and longitude coordinates of the address subject; The fourth determination module 36 is used to determine the accurate address object based on the address object and the entity knowledge base; The fifth determination module 37 is used to determine candidate address object information from the accurate address objects based on the accurate address subject, time distance, and preset association conditions; The sorting module 38 is used to calculate the comprehensive address score based on the straight-line distance between the accurate address subject and the candidate address object information, the preset semantic weight, and the preset time adaptability weight, and sort the candidate address object information in descending order according to the comprehensive address score.

[0063] In one embodiment of the present invention, determining the address subject and address object of an address-class address entity based on an entity knowledge base, a preset address knowledge graph, standard topology information, and standard direction information includes: The standard address entity is determined by entity links based on the entity knowledge base; The association information of standard address entities is determined based on the preset address knowledge graph; The address subject and address object are determined based on the association information, standard topology information, and standard direction information of the standard address entity.

[0064] In one embodiment of the present invention, determining the accurate address object based on the address object and an entity knowledge base includes: The corresponding accurate address object is determined in the entity knowledge base based on the address object.

[0065] In one embodiment of the present invention, a comprehensive address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order according to the comprehensive address score, including: The Haversine algorithm is used to calculate the straight-line distance between the candidate address object information and the accurate address subject. The address score is calculated by weighting the straight-line distance based on preset semantic weights and preset time adaptability weights. The candidate address information is sorted in descending order based on the overall address score.

[0066] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the address retrieval method based on spatiotemporal semantics provided by the present invention described above.

[0067] A fifth aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the address retrieval method based on spatiotemporal semantics provided in the above-described embodiments of the present invention.

[0068] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage system located remotely from the aforementioned processor.

[0069] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware systems.

[0070] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0071] For system / electronic device embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be found in the description of the method embodiments.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0073] 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.

[0074] 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.

[0075] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An address retrieval method based on spatiotemporal semantics, characterized in that, Includes the following steps: The address description information is sequentially input into the BERT model, the bidirectional LSTM model, and the CRF model for word segmentation, and the output is multiple address entities as well as the latitude and longitude coordinates and type labels of the address entities; Multiple address entities are input into a semantic processing model for semantic classification to obtain spatial address entities, temporal address entities, and address-type address entities. The standard topology and standard direction information of the spatial address entity are determined based on the spatial semantic knowledge base, and the time distance information of the time address entity from the current time is determined based on the time knowledge base. The address subject and address object of the address class address entity are determined based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information; The accurate address subject is determined based on the latitude and longitude coordinates of the address subject; The accurate address object is determined based on the address object and the entity knowledge base; Based on the accurate address subject, the time distance, and the preset association conditions, candidate address object information is determined from the accurate address objects; The address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order based on the address score.

2. The method as described in claim 1, characterized in that, The step of determining the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information includes: The standard address entity of the address class address entity is determined by entity linking based on the entity knowledge base; The association information of standard address entities is determined based on the preset address knowledge graph; The address subject and address object are determined based on the association information of the standard address entity, the standard topology information, and the standard direction information.

3. The method as described in claim 1, characterized in that, Determining the accurate address object based on the address object and the entity knowledge base includes: Based on the address object, the corresponding accurate address object is determined in the entity knowledge base.

4. The method as described in claim 1, characterized in that, A comprehensive address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order based on the comprehensive address score, including: The straight-line distance between the candidate address object information and the accurate address subject is calculated using the Haversine algorithm. The address score is obtained by weighting the straight-line distance according to the preset semantic weight and the preset time adaptability weight. The candidate address object information is sorted in descending order based on the comprehensive address score.

5. An address retrieval system based on spatiotemporal semantics, characterized in that, include: The word segmentation module is used to input address description information into the BERT model, bidirectional LSTM model and CRF model in sequence for word segmentation, and output multiple address entities as well as the latitude and longitude coordinates and type labels of the address entities; The classification module is used to input multiple address entities into the semantic processing model for semantic classification, and obtain spatial address entities, time-based address entities, and address-based address entities. The first determining module is used to determine the standard topology information and standard direction information of the spatial address entity based on the spatial semantic knowledge base, and to determine the time distance information of the time address entity from the current time based on the time knowledge base. The second determining module is used to determine the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information and the standard direction information; The third determining module is used to determine the accurate address subject based on the latitude and longitude coordinates of the address subject; The fourth determining module is used to determine the accurate address object based on the address object and the entity knowledge base; The fifth determining module is used to determine candidate address object information from the accurate address objects based on the accurate address subject, the time distance, and preset association conditions; The sorting module is used to calculate a comprehensive address score based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight, and to sort the candidate address object information in descending order based on the comprehensive address score.

6. The system as described in claim 5, characterized in that, The step of determining the address subject and address object of the address class address entity based on the entity knowledge base, the preset address knowledge graph, the standard topology information, and the standard direction information includes: The standard address entity of the address class address entity is determined by entity linking based on the entity knowledge base; The association information of standard address entities is determined based on the preset address knowledge graph; The address subject and address object are determined based on the association information of the standard address entity, the standard topology information, and the standard direction information.

7. The system as described in claim 5, characterized in that, Determining the accurate address object based on the address object and the entity knowledge base includes: Based on the address object, the corresponding accurate address object is determined in the entity knowledge base.

8. The system as described in claim 5, characterized in that, A comprehensive address score is calculated based on the straight-line distance between the accurate address subject and the candidate address object information, a preset semantic weight, and a preset time adaptability weight. The candidate address object information is then sorted in descending order according to the comprehensive address score, including: The straight-line distance between the candidate address object information and the accurate address subject is calculated using the Haversine algorithm. The address score is obtained by weighting the straight-line distance according to the preset semantic weight and the preset time adaptability weight. The candidate address object information is sorted in descending order based on the comprehensive address score.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the address retrieval method based on spatiotemporal semantics as described in any one of claims 1 to 4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the address retrieval method based on spatiotemporal semantics as described in any one of claims 1 to 4.

Images