Address determination method, apparatus, device, storage medium, and program product

Abstract

This application discloses a method, apparatus, device, storage medium, and program product for address determination, relating to the field of address management technology. The method first obtains address text input by a user; based on a preset spatial topology graph, the address text is replaced with the geographic entities corresponding to the target nodes to obtain a first address text; features are extracted from the first address text to obtain a sequence of target address elements; the sequence of target address elements is input into a gated network model to determine the probability of selecting a preset expert model; based on the preset model parameters of the target expert model whose probability is greater than a set threshold and multiple target address elements, correction data is determined, and the multiple target address elements are corrected to obtain a second address text; features are extracted from the second address text and coordinate transformation is performed to obtain a coordinate feature vector; the target address text is generated based on the second address text and the coordinate feature vector. The embodiments of this application improve the accuracy of address determination.

Description

Technical Field

[0001] This application belongs to the field of address management technology, and in particular relates to a method, apparatus, device, storage medium and program product for address determination. Background Technology

[0002] The efficient operation of network infrastructure plays a crucial role in data processing and information exchange. Address management technology, as a vital component, has played a key role in the development of the internet and is widely used in various network devices and systems. However, with the continuous expansion of network scale and the increasing complexity of application scenarios, existing technologies face numerous severe challenges. Existing address management technologies mainly fall into three categories: address resolution based on Hidden Markov Models (HMMs), address resolution utilizing geocoding services, and address resolution based on rule-based error correction engines. HMMs analyze address structures by segmenting address elements; geocoding services convert addresses into geographic coordinates; and rule-based error correction engines primarily rely on manually defined regular expressions to identify and correct address errors. However, because people's descriptions of addresses may be personal or regionally specific, or place names may change over time, the input address may differ significantly from the standard address description. HMMs and rule-based error correction engines struggle to identify and locate addresses outside the standard range. Furthermore, different map service providers use different coordinate systems, making it difficult for addresses in different coordinate systems to be mutually recognized. Therefore, existing technologies have low accuracy in identifying and determining user-input addresses. Summary of the Invention

[0003] This application provides a method, apparatus, device, storage medium, and program product for address determination, in order to solve the problem that the accuracy of prior art in identifying and determining user-input addresses is low.

[0004] In a first aspect, embodiments of this application provide a method for address determination, the method comprising: Get the address text entered by the user; Based on a preset spatial topology graph, the geographic entities associated with the target nodes in the address text are replaced with the geographic entities corresponding to the target nodes to obtain the first address text; the target nodes are preset standard geographic entities in the spatial topology graph. Feature extraction is performed on the first address text to obtain the target address element sequence, which includes multiple target address elements; The sequence of target address elements is input into the gated network model. Based on the preset weight matrix of the corresponding preset expert model in the gated network model, the probability of multiple target address elements selecting the corresponding preset expert model is determined. Select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and multiple target address elements, and correct multiple target address elements based on the correction data to obtain the second address text; Feature extraction is performed on the second address text, and the extracted second address text features are then subjected to coordinate transformation to obtain a coordinate feature vector; The target address text is generated based on the second address text and the coordinate feature vector.

[0005] Secondly, embodiments of this application provide an address determination device, the device comprising: The acquisition module is used to acquire the address text input by the user; The replacement module is used to replace the geographic entities associated with the target node in the address text with the geographic entities corresponding to the target node based on the preset spatial topology map, so as to obtain the first address text; the target node is a preset standard geographic entity in the spatial topology map. The extraction module is used to extract features from the first address text to obtain a sequence of target address elements, which includes multiple target address elements. The determination module is used to input the target address element sequence into the gated network model, and determine the probability of multiple target address elements selecting the corresponding preset expert model according to the preset weight matrix of the corresponding preset expert model in the gated network model; The determination module is also used to select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and multiple target address elements, and correct multiple target address elements based on the correction data to obtain the second address text. The extraction module is also used to extract features from the second address text and perform coordinate transformation on the extracted second address text features to obtain a coordinate feature vector; The generation module is used to generate the target address text based on the second address text and the coordinate feature vector.

[0006] Thirdly, embodiments of this application provide a terminal device, which includes: a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the address determination method as described in the first aspect.

[0007] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the address determination method as described in the first aspect.

[0008] Fifthly, embodiments of this application provide a computer program product in which instructions, when executed by a processor of an electronic device, cause the electronic device to perform the address determination method as described in the first aspect.

[0009] This application provides a method, apparatus, device, storage medium, and program product for address determination. The method first obtains address text input by a user; based on a preset spatial topology graph, it replaces the geographic entities associated with target nodes in the address text with the corresponding geographic entities in the spatial topology graph, obtaining a first address text; the target node is a preset standard geographic entity in the spatial topology graph; it extracts features from the first address text to obtain a sequence of target address elements, which includes multiple target address elements; it inputs the target address element sequence into a gated network model, and determines the probability of selecting the corresponding preset expert model for each target address element based on the preset weight matrices of different preset expert models in the gated network model; it selects a target expert model with a probability greater than a set threshold, determines correction data based on the preset model parameters of the target expert model and the multiple target address elements, and corrects the multiple target address elements based on the correction data, obtaining a second address text; it extracts features from the second address text and performs coordinate transformation on the extracted second address text features to obtain a coordinate feature vector; and it generates a target address text based on the second address text and the coordinate feature vector. This application embodiment uses a pre-defined spatial topology graph to uniformly replace associated geographic entities in the user-input address with standard geographic entities, achieving address text standardization from the source and reducing recognition errors caused by non-standard expressions. By extracting features from the standardized address, a regular sequence of address elements is obtained; then, a gating network is used to calculate the matching probability for each address element based on a weight matrix, and a high-confidence expert model is selected for precise correction, improving the rationality and correctness of the address elements. Based on the corrected address features, coordinate transformation is performed, fusing textual semantics and spatial location features, and finally, the target address is generated by combining standardized text and coordinate vectors, preserving the textual information of the address while adding the geographic coordinate dimension. Therefore, this application embodiment improves the accuracy of identifying and determining user-input addresses. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the structure of the address determination system provided in the embodiments of this application; Figure 2This is a flowchart illustrating the address determination method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the overall process for constructing a spatial topology graph provided in the embodiments of this application; Figure 4 This is a flowchart illustrating the method for determining the target address element sequence provided in an embodiment of this application; Figure 5 This is a flowchart illustrating the method for determining coordinate feature vectors provided in an embodiment of this application; Figure 6 This is a schematic diagram of the overall process of error correction using entropy values ​​provided in the embodiments of this application; Figure 7 This is a schematic diagram of the overall process for coordinate transformation provided in the embodiments of this application; Figure 8 This is a schematic diagram of the address determination device provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application. Detailed Implementation

[0012] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0013] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0014] In today's digital age, the efficient operation of network infrastructure plays a crucial role in data processing and information exchange. Address management technology, as a vital component, has played a key role in the development of the internet and is widely used in various network devices and systems. However, with the continuous expansion of network scale and the increasing complexity of application scenarios, existing technologies also face many severe challenges.

[0015] First, existing solutions mostly employ single-modal processing. Current address management often relies on rule engines or relatively simple neural network architectures. Their limitations become apparent when faced with complex and ever-changing real-world address descriptions. For example, unstructured addresses like "No. 3, Old Alleyway, Pudong Road" lack a clear, rule-compliant format, making traditional rule-based matching methods difficult to parse accurately. Rule engines typically require pre-setting a series of fixed formats and patterns, which are ineffective for such ambiguous and flexible expressions. Simple neural networks, due to limitations in their architecture and training data, struggle to extract accurate address elements, leading to errors or failures in address identification and location by network devices.

[0016] Secondly, existing solutions suffer from fragmented coordinate systems. In the current geographic information service market, different map service providers employ different coordinate systems. This independent coordinate system setup creates significant difficulties for cross-platform data fusion when frontline employees fill in equipment information or location data. When integrating data from different map platforms, the coordinate representation of the same geographical location is completely different on different platforms due to the differences in coordinate systems. This not only requires complex coordinate transformation calculations during data merging but also easily leads to accuracy loss during the transformation process, severely hindering efficient data utilization and increasing the cost and difficulty of data resource processing and maintenance.

[0017] Finally, existing solutions lack sufficient error correction capabilities. Current address error correction methods primarily rely on edit distance matching, which calculates the character differences between the input address and addresses in a known address database to determine and correct errors. However, this method is too simplistic and cannot adequately incorporate spatiotemporal context for reasoning. In reality, place names change over time. For example, in some historically significant areas, old place names are gradually replaced by new ones, but the old names are still used in older records or everyday expressions. Traditional methods cannot understand this historical shift, making accurate error correction difficult when encountering addresses containing historical place names. For instance, a region formerly called "Taiping Town" may now be renamed "Heping Town." If a user enters "a certain address in Taiping Town," traditional error correction methods may not correctly associate it with the current "Heping Town" address, thus affecting the accuracy and usability of the address.

[0018] To address the problems of existing technologies, embodiments of this application provide a method, apparatus, device, storage medium, and program product for address determination. The method first obtains address text input by a user; based on a preset spatial topology graph, it replaces the geographic entities associated with target nodes in the address text with the corresponding geographic entities in the spatial topology graph, obtaining a first address text; the target node is a preset standard geographic entity in the spatial topology graph; it extracts features from the first address text to obtain a sequence of target address elements, which includes multiple target address elements; it inputs the target address element sequence into a gated network model, and determines the probability of selecting the corresponding preset expert model for each target address element based on the preset weight matrices of different preset expert models in the gated network model; it selects a target expert model with a probability greater than a set threshold, determines correction data based on the preset model parameters of the target expert model and the multiple target address elements, and corrects the multiple target address elements based on the correction data, obtaining a second address text; it extracts features from the second address text and performs coordinate transformation on the extracted second address text features to obtain a coordinate feature vector; and it generates a target address text based on the second address text and the coordinate feature vector. This application embodiment uses a pre-defined spatial topology graph to uniformly replace associated geographic entities in the user-input address with standard geographic entities, achieving address text standardization from the source and reducing recognition errors caused by non-standard expressions. By extracting features from the standardized address, a regular sequence of address elements is obtained; then, a gating network is used to calculate the matching probability for each address element based on a weight matrix, and a high-confidence expert model is selected for precise correction, improving the rationality and correctness of the address elements. Based on the corrected address features, coordinate transformation is performed, fusing textual semantics and spatial location features, and finally, the target address is generated by combining standardized text and coordinate vectors, preserving the textual information of the address while adding the geographic coordinate dimension. Therefore, this application embodiment improves the accuracy of identifying and determining user-input addresses.

[0019] The following section first introduces the address determination system provided in the embodiments of this application, which is applied to the address determination method.

[0020] Figure 1 A schematic diagram of the structure of an address determination system according to an embodiment of this application is shown. Figure 1 As shown, the system may include: The system comprises a multimodal spatiotemporal large model module 101 and a quantum geocoding system 102. The multimodal spatiotemporal large model module 101 supports multi-source address input, including text, voice, and image ORC. Through preliminary processing and analysis of the input unstructured data, it dynamically schedules expert models to mine spatiotemporal features and semantic information in geographic information, outputting semantic parsing results (standardized address elements), spatial topological features, and dynamic scheduling weights, providing high-quality feature information for geocoding in the quantum geocoding system 102. The quantum geocoding system 102 employs a quantum geocoding system (QGS) to further achieve accurate geocoding, coordinate system transformation, address error correction, and standardization, ensuring the compatibility and consistency of geographic information across different systems and improving the quality and usability of address data.

[0021] The multimodal spatiotemporal large model module 101 includes a spatial topology modeling layer 1011, a semantic parsing layer 1012, and a dynamic scheduling layer 1013. The spatial topology modeling layer 1011 models the relationships between geographic entities (roads, buildings) based on the input raw geographic data, generating a topological network. The underlying model of the spatial topology modeling layer uses the DeepSeek-Math mathematical inference engine, constructing a graph structure for geographic entities through a Spatial-GCN network, where nodes represent entities and edges represent spatial relationships. The semantic parsing layer 1012 parses unstructured address data and extracts standardized address features. The base model integrates GPT-4o's multimodal understanding capabilities, supporting text, voice, and image OCR input. The dynamic scheduling layer 1013 dynamically calculates expert activation probabilities through a MoE gating network, combining the input raw multimodal data and intermediate features to dynamically schedule and select expert models (geocoding, semantic coding, and spatial reasoning experts), optimizing resource allocation to improve parsing efficiency and resource utilization in complex address scenarios. The quantized geocoding system 102 includes a semantic-spatial joint embedding layer 1021, a cross-coordinate system dynamic transformation layer 1022, a quantized error correction engine 1023, and an output layer 1024. The semantic-spatial joint embedding layer 1021 jointly embeds textual semantic information and spatial relationship information, mapping them to a unified vector space and generating a comprehensive feature vector, providing a more comprehensive and accurate feature representation for subsequent 3D coordinate prediction. The cross-coordinate system dynamic transformation layer 1022 is used to achieve lossless transformation between the three coordinate systems WGS84, GCJ02, and BD09 to ensure the consistency and accuracy of coordinate data across different platforms. The quantized error correction engine 1023 uses an entropy model to evaluate and correct geocoding results, dynamically identifying and correcting errors and ambiguities in address information. The output layer 1024 integrates and outputs 3D geographic coordinates and standardized addresses.

[0022] Figure 2A flowchart illustrating an address determination method according to an embodiment of this application is shown. Figure 2 As shown, the method may include the following steps: S201 to S207.

[0023] S201, Obtain the address text entered by the user.

[0024] The address text is the address description information entered by the user in text form.

[0025] In some embodiments, address description information input by the user in other forms can also be obtained and converted into address text. For example, address description information input by the user via voice can be converted into address text, or text extraction can be performed on an image input by the user that includes address description information to obtain the corresponding address text.

[0026] This application embodiment directly obtains the original address text input by the user, fully preserving the original address semantics and geographical information expressed by the user, providing complete information basis for subsequent address correction and coordinate transformation.

[0027] S202, based on the preset spatial topology diagram, replace the geographic entities in the address text that are associated with the target node in the spatial topology diagram with the geographic entities corresponding to the target node to obtain the first address text; the target node is a preset standard geographic entity in the spatial topology diagram.

[0028] Among them, the spatial topology graph is a topological map constructed with standard geographic entities as nodes and geographic spatiotemporal relationships as edges. The geographic spatiotemporal relationships can include actual geographic spatial relationships or historical time relationships.

[0029] In some embodiments, based on a pre-constructed spatial topology graph, various geographic entities in the address text can be traversed, and old place names, historical geographic entities, or non-standard geographic entities that are associated with the standard target node can be replaced with the standard geographic entities corresponding to the target node, thereby obtaining the first address text that has completed the initial standardization process.

[0030] In one example, when the input is “No. 3, 2nd Floor, Pudong Road Old Alley”, because the unstructured data contains the historical place name “Pudong Road”, the spatial topology modeling layer will associate the 1995 administrative division change record and building number mapping table, and associate the topological relationship between the historical name “Pudong Road” and the current “Pudong Avenue”.

[0031] This application embodiment uses a spatial topology graph to replace non-standard geographic entities with standard geographic entities, which can quickly correct historical and obsolete place names in address text, eliminate basic ambiguities in address expression, provide standardized basic text for subsequent address element extraction and expert model correction, and improve the accuracy of subsequent processing.

[0032] S203, perform feature extraction on the first address text to obtain the target address element sequence, which includes multiple target address elements.

[0033] The target address element sequence is an ordered set of address elements obtained by splitting the first address text. Address elements are the smallest semantic units that constitute an address, such as road names, building numbers, and floor numbers.

[0034] In some embodiments, semantic segmentation and feature extraction are performed on the first address text that has completed standard entity replacement. The text is decomposed into multiple independent address elements according to the composition structure of the address, and then arranged into an ordered sequence of target address elements according to the natural expression order of the address. This provides a structured processing object for expert model scheduling of the subsequent gated network model.

[0035] This application embodiment transforms unstructured address text into a structured sequence of address elements, which can decompose messy address information into independent units that can be recognized by the model, thereby achieving refined address correction processing.

[0036] S204, Input the target address element sequence into the gated network model, and determine the probability of multiple target address elements selecting the corresponding preset expert model according to the preset weight matrix of the corresponding preset expert model in the gated network model.

[0037] The gated network model is an allocation decision model used to schedule different expert models, and it can allocate processing weights based on input information. The pre-defined expert models are specialized processing models trained for different address problems. The pre-defined weight matrix is ​​a parameter matrix pre-trained in the gated network model used to calculate the matching degree between elements and the model.

[0038] In some embodiments, a structured sequence of target address elements is input into a pre-trained gating network model. The weight matrices of different types of preset expert models within the gating network are used to calculate the matching degree between each target address element and each preset expert model, and the probability of each address element selecting the corresponding expert model is presented in probabilistic form.

[0039] In some embodiments, the preset expert model may include a geocoding expert model, a semantic parsing expert model, and a spatial reasoning expert model.

[0040] This application embodiment calculates the matching probability between address elements and expert models through a gating network model, which enables intelligent dynamic scheduling of expert models and provides a quantitative basis for subsequent selection of the optimal expert model.

[0041] S205, Select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and multiple target address elements, and correct the multiple target address elements based on the correction data to obtain the second address text.

[0042] The threshold is a pre-set probability cutoff value used to filter effective expert models, for example, a threshold of 0.5. The target expert model is the expert model whose probability is greater than the set threshold and is selected to process the address element. The correction data is the standardized correction content generated by the expert model based on the address element. The second address text is the standardized address text obtained after correction by the expert model.

[0043] In some embodiments, based on the matching probability between each address element and the expert model calculated by the gated network model, target expert models with probabilities higher than a set threshold are selected, and then the address elements are processed using preset parameters of the target expert models to generate correction data. Finally, the address elements are corrected and integrated based on the correction data to form the second address text.

[0044] This application embodiment uses a target expert model with a high degree of matching for targeted correction, which can accurately solve the expression problems in the corresponding address text, further improve the standardization of address text, and avoid invalid models from participating in the processing, thereby improving the processing efficiency and accuracy of address correction.

[0045] S206, extract features from the second address text, and perform coordinate transformation on the extracted second address text features to obtain a coordinate feature vector.

[0046] The second address text features consist of semantic and spatial vectors extracted from the second address text for coordinate mapping. The coordinate feature vectors are semantic and spatial vectors representing the geographical location corresponding to the address.

[0047] In some embodiments, feature extraction is performed on the second address text corrected by the expert model to obtain text features that can represent address semantics. Then, the text features are transformed into coordinate feature vectors that can represent the corresponding geographical location through a preset coordinate transformation model, thereby realizing the mapping from address text to spatial coordinate features.

[0048] This application embodiment converts address text into coordinate feature vectors, which can realize the digital representation of address semantics to spatial location, and provide a spatial positioning basis for the subsequent generation of target address text containing precise coordinates.

[0049] S207, Generate target address text based on the second address text and coordinate feature vector.

[0050] The target address text is the final generated complete address result, which includes standard address descriptions and precise coordinate information.

[0051] In some embodiments, by using the first address text as a semantic reference and combining it with the precise geographic coordinate information and semantic information contained in the obtained coordinate feature vector, a target address text that combines standardized textual description and precise spatial positioning is finally generated.

[0052] This application's embodiment combines the original standardized basic text with coordinate feature vectors to generate the final result, which not only retains the original core semantics of the address but also incorporates accurate spatial positioning information, allowing the target address text to simultaneously possess readable textual descriptions and applicable spatial coordinates, thus meeting the needs of various geographic information application scenarios.

[0053] In summary, this application's embodiments first rely on spatial topology diagrams to standardize basic place names, and then use gating networks to achieve fine-grained scheduling and correction of expert models. This effectively solves the problems of non-standardization in address text, such as historical place names and dialect expressions. At the same time, it transforms address semantics into precise coordinate features and integrates them to generate complete target addresses. The overall process combines semantic standardization with spatial positioning accuracy. It can preserve the original core semantics of the address to avoid information loss, and improve the efficiency and accuracy of address correction and coordinate transformation through hierarchical processing and intelligent model scheduling. Thus, it provides a standardized, accurate address data foundation that combines text and coordinates, adapting to address determination and positioning needs in multiple scenarios.

[0054] In some embodiments, before replacing the geographic entity associated with the target node in the address text with the geographic entity corresponding to the target node, the method may further include: Acquire raw geographic data, which includes geographic entities and the relationships between them; Using geographic entities as nodes and the relationships between geographic entities as edges, an initial spatial topology graph is constructed. For each node in the initial spatial topology graph, a fusion feature is determined based on the node's characteristics and the characteristics of its connected neighboring nodes. The node is then updated based on the fusion feature to obtain the spatial topology graph.

[0055] In one example, the original geographic data includes “Pudong Road”, “Pudong Avenue”, “Building No. 3”, and “Old Alley”. These four geographic entities are treated as independent nodes, and “Pudong Road and Pudong Avenue correspond to the same area”, “Building No. 3 is located next to Pudong Road”, and “Old Alley belongs to the Pudong Road area” are used as edges to construct an initial spatial topology graph containing these four nodes and three edges.

[0056] This application embodiment obtains geographic entities and the relationships between these geographic entities, providing a real and complete data source for the construction of the subsequent initial spatial topology graph, ensuring that the construction of the subsequent topology graph conforms to the actual geographic scene, and then transforms the abstract geographic relationships into an intuitive node edge structure through the topology graph, providing a reliable spatial topology foundation for address modification in subsequent steps, and reducing semantic deviations and information loss in the address processing process.

[0057] In some embodiments, the fusion features are determined based on the characteristics of the node and the characteristics of the neighboring nodes connected to the node. The calculation formula can be: in, For nodes In the Features of the layer For nodes The set of neighboring nodes (usually including) itself), The learnable weight matrix parameters for layer l and The learnable bias vector parameters for layer l. It is a non-linear activation function.

[0058] In some embodiments, for nodes Every neighbor ∈ ( First, set the features of the neighboring nodes at the current layer. Multiply by the weight matrix For each neighboring node's features, a linear projection is made onto the new feature space to obtain the transformed feature vector. Sum the transformed feature vectors of all neighboring nodes and add the bias vector to the aggregation result. This increases the model's expressive power. Finally, the results are input into the activation function. In this process, the introduction of nonlinearity enables the network to learn more complex patterns and obtain node... New features at the next layer ( +1).

[0059] In this embodiment, each node updates its own features by examining the features of all its neighbors, summarizing their information, and then performing a linear transformation and nonlinear activation. Through multiple layers of this operation, each node's features gradually fuse higher-order neighborhood information. All nodes share the same weight matrix. ( This greatly reduces the number of parameters.

[0060] In some embodiments, determining fusion features for each node in the initial spatial topology graph based on the node's features and the features of its connected neighboring nodes may include: For each node, based on the node's characteristics and the characteristics of its connected neighboring nodes, the importance score of each neighboring node is determined, and neighboring nodes with an importance score greater than a set threshold are selected as the node's target neighboring nodes. Here, the node characteristics are the information of the geographic entity corresponding to the node; the importance score represents the degree of influence or relevance of the neighboring node on the current node, and is calculated through an attention mechanism. The fusion features are determined based on the characteristics of the node and the characteristics of the target neighbor nodes connected to the node.

[0061] In some embodiments, the formula for calculating the importance score can be: in, For learnable parameters, Represents a node After weight matrix Transformed features This indicates a splicing operation. Represents the normalized exponential function, This represents the activation function.

[0062] In some embodiments, a shared weight matrix is ​​first used. Node and nodes The features are respectively mapped to and This maps the original features to an attention space, allowing the model to learn which feature combinations are more important. The transformed two vectors are then concatenated into a single long vector. and learnable vectors Taking the dot product of the transpose and concatenated vector yields the attention coefficients, representing the unnormalized values ​​and reflecting the neighboring nodes. For nodes The original importance scores are obtained. Finally, the LeakyReLU activation function is applied and softmax normalization is performed to obtain the final importance score. The LeakyReLU activation function introduces non-linearity, allowing a small number of negative values ​​to be passed, which helps to preserve subtle differences in attention. Softmax ensures that the sum of the importance scores of all neighbors is 1, making the score probabilistically interpretable and facilitating subsequent weighted aggregation.

[0063] In one example, when the input is “No. 3, 2nd Floor, Pudong Road Old Alley”, because the unstructured data contains the historical place name “Pudong Road”, the spatial topology modeling layer will associate the 1995 administrative division change record and building number mapping table, and associate the topological relationship between the historical name “Pudong Road” and the current “Pudong Avenue”.

[0064] This application embodiment filters out target neighbor nodes that have little impact on the node by calculating importance scores, reducing interference and improving efficiency. Then, the node and target neighbor nodes are fused to obtain a more comprehensive and accurate node representation. The whole process simplifies the data and enhances the expressive power of features, making the final spatial topology graph more accurate, thereby improving the efficiency and accuracy of address determination.

[0065] In one example, such as Figure 3 As shown, the overall process of constructing a spatial topology graph can include: inputting geographic entity data to be processed in step S301; then proceeding to step S302, constructing a spatial graph network based on these geographic entities, where nodes represent each geographic entity and edges represent the spatial relationships between entities; next, in step S303, inputting the constructed spatial graph network into a spatial graph convolutional network; in step S304, assigning attention weights to different nodes and edges through a dynamic weight allocation mechanism to quantify the importance of each entity and its spatial relationships; then, in step S305, using a judgment branch of "whether the attention weight exceeds a set threshold", nodes with weights higher than the threshold are identified as key nodes and retained in step S306, while nodes with weights lower than the threshold are identified as noise nodes and filtered out in step S307; finally, in step S308, based on the filtered key nodes and retained edges, outputting an optimized spatial topology graph, achieving accurate modeling and noise filtering of the spatial structure of geographic entities, providing a reliable topological foundation for subsequent geographic information processing.

[0066] In some embodiments, such as Figure 4 As shown, feature extraction is performed on the first address text to obtain the target address element sequence, which may include: S401 to S404.

[0067] S401, Semantic feature extraction is performed on the first address text to obtain multiple address element features.

[0068] Among them, the address element feature is the smallest core semantic unit feature that constitutes an address.

[0069] In one example, when the input is "2nd Floor, No. 3, Old Lane, Pudong Road", the key address elements "Pudong Road" and "No. 3, Lane" are extracted after parsing.

[0070] This application embodiment extracts semantic features to decompose the first address text, which may contain redundant information, into structured and independent address element features, accurately preserving the core semantic information of the address and avoiding interference from redundant information in subsequent processing.

[0071] S402, for each address element feature, determine the deep semantic features of the address element feature based on the address element feature and all previous address element features, including the address element feature, in the first address text.

[0072] Among them, deep semantic features are comprehensive features generated by combining the semantics of the current address element's own features with the contextual semantic association of all its previous address element features, thus possessing both its own features and contextual logic. In some embodiments, deep semantic features are modeled using Transformer-XL long sequence modeling to handle long text dependencies, and the calculation formula can be: in, Represents the time from the beginning of the sequence to time. The input sequence, Indicates the first The hidden state at a given moment carries information about the previous sequence.

[0073] This application uses Transformer-XL, which incorporates the hidden state information of the previous segment into the calculation of the current segment by introducing relative position encoding and a loop mechanism. This effectively solves the problem of long text dependencies, enabling the model to process long sequence data and capture long-distance dependencies.

[0074] S403, determine the semantic score value of each address element feature sequence based on the deep semantic features of all address element features.

[0075] The address element feature sequence is an ordered set composed of multiple extracted address element features. The semantic score is a score that quantifies the semantic completeness and logical coherence of the address element feature sequence.

[0076] In some embodiments, to optimize the label sequence in conjunction with the CRF layer, BERT-wwm (Whole Word Masking) entity recognition is introduced, and the calculation formula can be: in, It is the input text. It is the corresponding label sequence. This represents each position in the output of the BERT-wwm model. The hidden state, and These are the weights and bias parameters learned by the model, used to calculate the current label. The association with the hidden state and the relationship between adjacent labels.

[0077] This application's embodiments, by combining a CRF layer, can consider the dependencies between tags, optimize the prediction of tag sequences, and thus improve the accuracy of entity recognition. Furthermore, through quantified semantic scoring values, the normalization and completeness of the feature sequence of each address element can be objectively and accurately measured. Simultaneously, the semantic scoring, based on deep semantic features, can fully consider the contextual associations and semantic logic of each element, ensuring the accuracy and reasonableness of the scoring results.

[0078] S404. Select the address element feature sequence with the highest semantic score as the target address element sequence.

[0079] This application embodiment ensures that the final target address element sequence is the most standardized, complete, and semantically accurate by selecting the address element sequence with the highest semantic score. This reduces problems such as subsequent address correction errors and coordinate transformation deviations caused by selecting semantically incomplete or logically disordered sequences, thereby improving the efficiency and reliability of the determined address.

[0080] Since the expert model scheduling of the traditional MoE architecture relies on static weights, this approach cannot be adapted to all scenarios. Therefore, this application uses the MoE gated network (E1) to dynamically calculate the expert activation probability and combines the spatiotemporal context (the original multimodal data and intermediate features of the input) to dynamically schedule and select expert models, thereby optimizing resource allocation and improving the parsing efficiency and resource utilization in complex address scenarios.

[0081] In some embodiments, the probability of multiple target address elements selecting the corresponding preset expert model is determined according to the preset weight matrices of different preset expert models in the gated network model. The calculation formula can be: in, These are input features. and These are the weights and bias parameters learned by the model. It is the output of the gating unit, and its value is between 0 and 1.

[0082] In one example, when the input is "No. 3, 2nd Floor, Pudong Road Old Alley", the gating network model determines to call a semantic parsing expert to analyze the specific meaning of the dialect word "alley" and convert it into a standard address.

[0083] In some embodiments, feature extraction of the second address text may include: The second address text is subjected to semantic feature extraction and spatial feature extraction respectively to obtain semantic feature vector and spatial feature vector; The semantic feature vector and the spatial feature vector are concatenated, and the concatenated feature vector is determined as the second address text feature.

[0084] The semantic feature vector represents the semantic logic and core meaning of the address, while the spatial feature vector represents the spatial location and spatial relationships of the address. The second address text feature is a joint embedding vector obtained by concatenating the semantic and spatial feature vectors.

[0085] In one example, Transformer-XL is introduced to process the address text during semantic feature extraction. A self-attention mechanism is used to capture long-distance dependencies in the text to obtain semantic feature vectors, thus achieving semantic embedding. For spatial feature extraction, GraphSAGE is introduced to model entities and their relationships in geospatial space, generating spatial vectors. The representations of nodes are updated by aggregating information from neighboring nodes, thereby achieving spatial embedding. Finally, the semantic feature vectors and spatial feature vectors are concatenated to obtain the second address text feature.

[0086] In some embodiments, the formula for calculating the concatenated semantic feature vector and spatial feature vector can be: in, For semantic feature vectors, These are spatial feature vectors.

[0087] In one example, when the input is "No. 3, 2nd Floor, Old Alleyway, Pudong Road", the semantic-spatial joint embedding layer will ultimately output a joint vector, including the semantic vector {0.32, 0.15, 0.87, ...} and spatial vector {longitude deviation} x = 0.0001°, latitude deviation y = 0.0002°, elevation h=0.5m}.

[0088] In some embodiments, such as Figure 5 As shown, the extracted second address text features are subjected to coordinate transformation to obtain a coordinate feature vector, which may include: S501 to S503.

[0089] S501, the extracted second address text features are subjected to coordinate transformation to obtain the initial coordinate feature vector.

[0090] The initial coordinate feature vector is the unoptimized and uncalibrated coordinate feature vector obtained after coordinate transformation; coordinate transformation is the process of converting the coordinates from the original coordinate system to the desired target coordinate system.

[0091] In some embodiments, the coordinate transformation formula can be: in, These are the coordinate values ​​in the original coordinate system. These are the coordinate values ​​in the target coordinate system after transformation. It is a coordinate transformation function. These parameters are obtained through continuous optimization during the reinforcement learning process and are used to adjust the specific method of coordinate transformation so that the transformed coordinates are closer to the true target coordinates.

[0092] This application embodiment achieves unification between different coordinate systems, enabling subsequent coordinates to adapt to common systems such as maps and positioning, ensuring standard coordinate format, and relying on complete address features to make the conversion more closely match the real location.

[0093] S502, determine the entropy values ​​of the second address text features and the initial coordinate feature vector based on the second address text features and the initial coordinate feature vector.

[0094] Among them, entropy is used to measure the degree of uncertainty or disorder in an address.

[0095] In some embodiments, the entropy values ​​of the second address text features are summed to obtain the total entropy value of the entire second address text. The higher the entropy value, the greater the uncertainty in the address representation, and the more likely there are errors or non-standardizations; the lower the entropy value, the more standardized the address representation.

[0096] This application embodiment calculates the entropy value of the second address text feature and the initial coordinate feature vector, providing an objective basis for determining whether subsequent correction is needed, avoiding the use of fixed distance thresholds or manual rules. At the same time, the entropy value calculation can integrate complex multimodal interaction information.

[0097] S503, when the total entropy value exceeds the set threshold, the initial coordinate feature vector is modified according to the preset parameters to obtain the coordinate feature vector.

[0098] The total entropy is the sum of the entropy values ​​of the second address text features and the initial coordinate feature vector. The preset parameters are pre-trained, learnable correction parameters, such as the weights of a corrected network or a linear transformation matrix.

[0099] In some embodiments, the calculation formula of the total entropy value may be: wherein n represents the number of address elements, is an address element, such as specific address components like "longtang" (alleyway), "building No." and so on. is the occurrence probability of the address element.

[0100] The embodiment of the present application adopts an entropy value-triggered condition correction mechanism, which only modifies coordinates when necessary, avoids additional errors that may be introduced by non-discriminative correction on all samples, can effectively improve the coordinate accuracy of high-entropy samples, and maintains the calculation efficiency of low-entropy samples.

[0101] In one example, if , dialect / historical name correction (e.g., "longtang" → "hutong") is triggered, is the set high entropy threshold; if , the standard address is directly mapped, is the low entropy threshold. For example: when inputting "No. 3, Old Longtang, Pudong Road, 2nd Floor", since the address entropy H_address=1.8 and τ_high=1.2, it is corrected to "No. 3, Hutong, Pudong Avenue, 2nd Floor".

[0102] In one example, as shown in Figure 6 , the overall process of error correction through entropy value may include: firstly, inputting a coordinate feature vector in step S601, then calculating the total entropy value corresponding to the coordinate feature vector in step S602; next, determining whether the total entropy value exceeds a set threshold in step S603, if the determination result is "yes", then entering step S604 to trigger the error correction process, performing dialect correction in step S605, historical name correction in step S606 and coordinate offset correction in step S607 respectively, and performing comprehensive correction from two dimensions of address expression specification and spatial position accuracy; if the determination result is "no", then directly entering step 608 for direct output; after error correction processing or direct output, uniformly entering step S609 to generate a standardized address, and finally outputting the error-corrected and optimized standardized address result in step S610, realizing entropy-driven collaborative error correction of address text and spatial coordinates, and ensuring the standardization of the output address and the accuracy of the position.

[0103] In some embodiments, performing coordinate conversion on the extracted second address text features to obtain an initial coordinate feature vector may include: inputting the second address text features into a coordinate conversion model, and determining the initial coordinate feature vector according to coordinate conversion parameters of the coordinate conversion model and the second address text features.

[0104] In some embodiments, the coordinate transformation model is a reinforcement learning-proximal policy optimization (PPO) algorithm model.

[0105] In some embodiments, before inputting the second address text features into the coordinate transformation model and determining the initial coordinate feature vector based on the coordinate transformation parameters of the coordinate transformation model and the second address text features, the method may further include: Obtain the training sample set. Each training sample in the training sample set includes an address text feature sample and its corresponding standard coordinates. Input the address text feature samples into the initial coordinate transformation model, and determine the training coordinate feature vector based on the initial coordinate transformation parameters of the initial coordinate transformation model and the address text feature samples; The reward value is determined based on the training coordinate feature vector and the corresponding standard coordinate, and the model parameters of the initial coordinate transformation model are updated based on the reward value until the updated value of the model parameters is less than the first set threshold, thus obtaining the coordinate transformation model; wherein, when the updated value of the model parameters exceeds the second set threshold, the model parameters are updated to the preset update value.

[0106] In some embodiments, the formula for calculating the reward value can be: in, These are the target coordinates. It predicts coordinate values. This is a weighting coefficient used to balance the effects of the distance and entropy terms. It is a strategy Entropy measures the uncertainty or randomness of a strategy. The entropy term is introduced to encourage exploration, prevent the algorithm from prematurely converging to a local optimum, and allow the model to try more different strategies during the optimization process to find better transformation parameters.

[0107] The embodiments of this application construct a coordinate transformation model by using the PPO algorithm model and introduce hierarchical reinforcement learning and entropy feedback expansion, so that the model can adaptively explore the optimal coordinates during the decision-making process, while maintaining training stability and sample efficiency.

[0108] In one example, such as Figure 7As shown, the overall process of coordinate transformation can include: First, inputting the second address text in step S701; then, performing semantic embedding in step S702 and spatial embedding in step S703 in parallel to extract the semantic and spatial features of the address, respectively; next, performing joint feature concatenation in step S704 to form a comprehensive feature vector that integrates semantic and spatial information; then, performing 3D coordinate prediction in step S705, and then transforming the predicted coordinates to the target coordinate system through cross-coordinate system transformation in step S706; in step S707, determining whether the coordinate error is less than a set threshold; if not, proceeding to the reinforcement learning optimization stage in step S708 to optimize the cross-coordinate system... The coordinate system transformation process iteratively corrects and returns to step S706 until the error meets the requirements. When the coordinate error meets the conditions, the optimized coordinates are output in step S709, and then the entropy error correction engine in step S710 is entered to calculate and evaluate the matching entropy value between the address and the coordinates. Then, in step S711, it is determined whether the entropy value is greater than the set threshold. If the entropy value exceeds the threshold, step S712 is entered to trigger error correction. If it does not exceed the threshold, step S713 is entered to directly output the value. Finally, a standardized address is generated in step S714 to complete the entire quantum geocoding (QGS) process, realizing the end-to-end accurate processing from address text to three-dimensional coordinates and then to standardized address.

[0109] In some embodiments, generating target address text based on the second address text and coordinate feature vectors may include: Regarding coordinates, the coordinates in the target coordinate system obtained after dynamic cross-coordinate system transformation are further determined as three-dimensional geographic coordinates, specifying longitude, latitude, elevation, and other information to accurately represent the geographical location. Regarding the address, the standardized address obtained through a quantum error correction engine removes non-standard expressions such as dialects and historical names, presenting it in a standardized form, such as "15F-B12, No. 1 Pudong Avenue". Finally, the integrated three-dimensional geographic coordinates and standardized address are output.

[0110] Specifically, on the one hand, the completed basic address standardization result (second address text) is combined with the optimized parts triggered by excessive entropy values ​​in the coordinate feature vector (dialect correction and historical name correction) to form the final standardized address text. This completely eliminates non-standard content such as dialect expressions, historical names, and colloquial abbreviations, outputting address text with a unified format, complete elements, and conforming to geographic service standards, such as "1 Pudong Avenue 15F-B12". On the other hand, the longitude, latitude, and elevation information in the coordinate feature vector are integrated into three-dimensional geographic coordinates, ensuring that the coordinates accurately reflect the real geographical location in the target coordinate system and completely match the semantics of the standardized address, without spatial offset or semantic conflict. Finally, the standardized address is strongly associated with the corresponding three-dimensional geographic coordinates, establishing a unique mapping relationship between the address text and spatial coordinates, forming complete geographic information data.

[0111] Figure 8 An address determination device 800 according to an embodiment of this application is shown. The device may include: The acquisition module 801 is used to acquire the address text input by the user; The replacement module 802 is used to replace the geographic entities associated with the target node in the address text with the geographic entities corresponding to the target node based on the preset spatial topology map, so as to obtain the first address text; the target node is a preset standard geographic entity in the spatial topology map. The extraction module 803 is used to extract features from the first address text to obtain a target address element sequence, which includes multiple target address elements. The determination module 804 is used to input the target address element sequence into the gated network model and determine the probability of multiple target address elements selecting the corresponding preset expert model according to the preset weight matrix of the corresponding preset expert model in the gated network model. The determination module 804 is also used to select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and multiple target address elements, and correct the multiple target address elements based on the correction data to obtain the second address text. The extraction module 803 is also used to extract features from the second address text and perform coordinate transformation on the extracted second address text features to obtain a coordinate feature vector. The generation module 805 is used to generate the target address text based on the second address text and the coordinate feature vector.

[0112] In some embodiments, the address determination device 800 may further include: The acquisition module 801 is also used to acquire raw geographic data, which includes geographic entities and the relationships between geographic entities. The building module is used to construct an initial spatial topology graph, with geographic entities as nodes and the relationships between geographic entities as edges. The determination module 804 is also used to determine the fusion features of each node in the initial spatial topology graph based on the characteristics of the node and the characteristics of the neighboring nodes connected to the node, and update the node based on the fusion features to obtain the spatial topology graph.

[0113] In some embodiments, the determining module 804 is further configured to, for each node, determine the importance score of each neighbor node of the node based on the characteristics of the node and the characteristics of the neighbor nodes connected to the node, and select the neighbor nodes with an importance score greater than a set threshold as the target neighbor nodes of the node. The fusion features are determined based on the characteristics of the node and the characteristics of the target neighbor nodes connected to the node.

[0114] In some embodiments, the extraction module 803 is further configured to extract semantic features from the first address text to obtain multiple address element features; The determination module 804 is further configured to, for each address element feature, determine the deep semantic features of the address element feature based on the address element feature and all previous address element features, including the address element feature, in the first address text. The determination module 804 is also used to determine the semantic score value of each address element feature sequence based on the deep semantic features of all address element features; The determination module 804 is also used to select the address element feature sequence with the largest semantic score as the target address element sequence.

[0115] In some embodiments, the extraction module 803 is further configured to perform semantic feature extraction and spatial feature extraction on the second address text respectively to obtain semantic feature vector and spatial feature vector; The determination module 804 is also used to concatenate semantic feature vectors and spatial feature vectors, and to determine the concatenated feature vector as the second address text feature.

[0116] In some embodiments, the address determination device 800 may further include: The conversion module is used to perform coordinate transformation on the extracted second address text features to obtain an initial coordinate feature vector; The determining module 804 is also used to determine the entropy values ​​of the second address text features and the initial coordinate feature vector based on the second address text features and the initial coordinate feature vector; The modification module is used to modify the initial coordinate feature vector according to preset parameters when the total entropy value exceeds a set threshold, so as to obtain the coordinate feature vector.

[0117] In some embodiments, the determining module 804 is further configured to input the second address text features into the coordinate transformation model, and determine the initial coordinate feature vector based on the coordinate transformation parameters of the coordinate transformation model and the second address text features.

[0118] In some embodiments, the acquisition module 801 is further configured to acquire a training sample set, wherein each training sample in the training sample set includes an address text feature sample and a corresponding standard coordinate. The determination module 804 is also used to input the address text feature samples into the initial coordinate transformation model, and determine the training coordinate feature vector based on the initial coordinate transformation parameters of the initial coordinate transformation model and the address text feature samples; The determination module 804 is also used to determine the reward value based on the training coordinate feature vector and the corresponding standard coordinate, and update the model parameters of the initial coordinate transformation model according to the reward value until the updated value of the model parameters is less than the first set threshold, thus obtaining the coordinate transformation model; wherein, when the updated value of the model parameters exceeds the second set threshold, the model parameters are updated to the preset update value.

[0119] Figure 8 The various modules in the illustrated device can achieve Figure 2 The various steps involved, and the corresponding technical effects achieved, will not be elaborated upon here for the sake of brevity.

[0120] Figure 9 A schematic diagram of the hardware structure of the terminal device provided in an embodiment of this application is shown.

[0121] The terminal device may include a processor 901 and a memory 902 storing computer program instructions.

[0122] Specifically, the processor 901 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0123] Memory 902 may include mass storage for data or instructions. For example, and not limitingly, memory 902 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 902 may include removable or non-removable (or fixed) media, or memory 902 may be non-volatile solid-state memory. Memory 902 may be internal or external to the integrated gateway disaster recovery device.

[0124] In one example, memory 902 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Thus, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method determined according to the address in this disclosure.

[0125] The processor 901 reads and executes computer program instructions stored in the memory 902 to achieve... Figure 2 The address determination method in the illustrated embodiment.

[0126] In one example, the terminal device may also include a communication interface 903 and a bus 904. Wherein, for example... Figure 9 As shown, the processor 901, memory 902, and communication interface 903 are connected through bus 904 and complete communication with each other.

[0127] The communication interface 903 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0128] Bus 904 includes hardware, software, or both, that couples components of an end device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 904 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0129] Furthermore, in conjunction with the address determination methods in the above embodiments, this application embodiment can provide a computer storage medium for implementation. This computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the address determination methods in the above embodiments.

[0130] This application also provides a computer program product, including a computer program, which, when executed, implements any of the address determination methods described in the above embodiments.

[0131] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0132] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or text segments used to perform the required tasks. Programs or text segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Text segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0133] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0134] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in 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, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0135] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for determining an address, characterized in that, include: Get the address text entered by the user; Based on a preset spatial topology graph, the geographic entities associated with the target nodes in the address text are replaced with the geographic entities corresponding to the target nodes to obtain the first address text; the target nodes are preset standard geographic entities in the spatial topology graph. Feature extraction is performed on the first address text to obtain a target address element sequence, wherein the target address element sequence includes multiple target address elements; The sequence of target address elements is input into the gated network model, and the probability of selecting the corresponding preset expert model for the multiple target address elements is determined according to the preset weight matrix of the corresponding preset expert model in the gated network model. Select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and the multiple target address elements, and correct the multiple target address elements based on the correction data to obtain the second address text; Feature extraction is performed on the second address text, and the extracted second address text features are then subjected to coordinate transformation to obtain a coordinate feature vector; The target address text is generated based on the second address text and the coordinate feature vector.

2. The address determination method according to claim 1, characterized in that, Before replacing the geographic entities associated with the target node in the address text with the geographic entities corresponding to the target node, the method further includes: Acquire raw geographic data, which includes geographic entities and the relationships between geographic entities; Using the geographic entities as nodes and the relationships between the geographic entities as edges, an initial spatial topology graph is constructed. For each node in the initial spatial topology graph, a fusion feature is determined based on the characteristics of the node and the characteristics of the neighboring nodes connected to the node, and the node is updated based on the fusion feature to obtain the spatial topology graph.

3. The address determination method according to claim 2, characterized in that, For each node in the initial spatial topology graph, a fusion feature is determined based on the node's characteristics and the characteristics of its connected neighboring nodes, including: For each node, based on the characteristics of the node and the characteristics of the neighboring nodes connected to the node, the importance score of each neighboring node of the node is determined, and the neighboring nodes with an importance score greater than a set threshold are selected as the target neighboring nodes of the node. The fusion features are determined based on the characteristics of the node and the characteristics of the target neighbor nodes connected to the node.

4. The address determination method according to claim 1, characterized in that, The step of extracting features from the first address text to obtain the target address element sequence includes: Semantic feature extraction is performed on the first address text to obtain multiple address element features; For each address element feature, a deep semantic feature is determined based on the address element feature and all previous address element features in the first address text, including the address element feature. The semantic score value of each address element feature sequence is determined based on the deep semantic features of all address element features. The address element feature sequence with the largest semantic score is selected as the target address element sequence.

5. The address determination method according to claim 1, characterized in that, The feature extraction of the second address text includes: The second address text is subjected to semantic feature extraction and spatial feature extraction respectively to obtain semantic feature vector and spatial feature vector; The semantic feature vector and the spatial feature vector are concatenated, and the concatenated feature vector is determined as the second address text feature.

6. The address determination method according to claim 1, characterized in that, The extracted second address text features are subjected to coordinate transformation to obtain a coordinate feature vector, including: The extracted second address text features are subjected to coordinate transformation to obtain the initial coordinate feature vector; The entropy values ​​of the second address text features and the initial coordinate feature vector are determined based on the second address text features and the initial coordinate feature vector. If the total entropy value exceeds a set threshold, the initial coordinate feature vector is modified according to preset parameters to obtain the coordinate feature vector.

7. The address determination method according to claim 6, characterized in that, The step of performing coordinate transformation on the extracted second address text features to obtain an initial coordinate feature vector includes: The second address text feature is input into the coordinate transformation model, and the initial coordinate feature vector is determined based on the coordinate transformation parameters of the coordinate transformation model and the second address text feature.

8. The address determination method according to claim 7, characterized in that, Before inputting the second address text features into the coordinate transformation model and determining the initial coordinate feature vector based on the coordinate transformation parameters of the coordinate transformation model and the second address text features, the method further includes: Obtain a training sample set, wherein each training sample in the training sample set includes an address text feature sample and its corresponding standard coordinates; The address text feature samples are input into the initial coordinate transformation model, and the training coordinate feature vector is determined based on the initial coordinate transformation parameters of the initial coordinate transformation model and the address text feature samples. The reward value is determined based on the training coordinate feature vector and the corresponding standard coordinate, and the model parameters of the initial coordinate transformation model are updated based on the reward value until the updated value of the model parameters is less than a first set threshold, thus obtaining the coordinate transformation model; wherein, when the updated value of the model parameters exceeds a second set threshold, the model parameters are updated to a preset update value.

9. An address determination device, characterized in that, include: The acquisition module is used to acquire the address text input by the user; The replacement module is used to replace the geographic entities associated with the target node in the address text with the geographic entities corresponding to the target node based on a preset spatial topology graph, thereby obtaining the first address text; the target node is a preset standard geographic entity in the spatial topology graph. The extraction module is used to extract features from the first address text to obtain a target address element sequence, wherein the target address element sequence includes multiple target address elements; The determination module is used to input the target address element sequence into the gated network model, and determine the probability of selecting the corresponding preset expert model for the multiple target address elements according to the preset weight matrix of the corresponding preset expert model in the gated network model. The determination module is also used to select a target expert model with a probability greater than a set threshold, determine correction data based on the preset model parameters of the target expert model and the plurality of target address elements, and correct the plurality of target address elements based on the correction data to obtain the second address text; The extraction module is also used to extract features from the second address text and perform coordinate transformation on the extracted second address text features to obtain a coordinate feature vector; The generation module is used to generate target address text based on the second address text and the coordinate feature vector.

10. A terminal device, characterized in that, The device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the address determination method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the address determination method as described in any one of claims 1-8.

12. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device performs the address determination method as described in any one of claims 1-8.

Images