Intelligent address resolution method and device, equipment and medium

Abstract

The invention provides an intelligent address resolution method, apparatus and device, and a medium. The method comprises the steps of obtaining a target address on a logistics express sheet by using an identification model; constructing an address-coordinate bidirectional mapping system to obtain geographic coordinates of the target address; extracting address entity information in the target address to generate a standardized address vector; respectively calculating Euclidean distances and semantic similarities with peripheral POIs meeting preset requirements on the basis of the geographic coordinates and the standardized address vectors of the target addresses, and generating multi-dimensional spatial feature vectors on the basis of the Euclidean distances, the semantic similarities and attribute features of the peripheral POIs; based on the standardized address vector and the multi-dimensional spatial feature vector, capturing address sequence dependence, a relationship between the region and peripheral POIs meeting preset requirements and a one-hot code of a target POI meeting the preset requirements through hierarchical modeling; and fusing the text feature vector and the enhanced spatial feature vector through an attention mechanism based on the target POI one-hot code meeting a preset requirement to obtain a final address representation vector.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent address resolution, and specifically, to an intelligent address resolution method, device, equipment and medium. Background Art

[0002] Traditional address resolution methods mostly rely on simple keyword matching or rule templates, and can only process structured addresses with standardized formats, such as "No. XX, XX Street, XX District, XX City, XX Province"; however, in practical applications, address information often exists in unstructured forms: for example, the address on a logistics label may include handwritten corrections, abbreviations, such as using "Jing" to represent "Beijing", reversed word order, such as "No. 123, XX Road, XX District", typos or redundant information, such as mixed with phone numbers and remarks, resulting in frequent failures of keyword matching and insufficient parsing accuracy.

[0003] In addition, traditional methods lack in-depth modeling of the hierarchical relationship of addresses, and cannot map unstructured addresses to a standardized hierarchical structure, thus making it difficult to achieve precise association with geographical coordinates, seriously restricting the automation and intelligence levels of downstream services such as logistics sorting and intelligent distribution. Summary of the Invention

[0004] The main purpose of the present invention is to solve the problem in the prior art that it mostly relies on simple keyword matching or rule templates, can only process structured addresses with standardized formats, and has insufficient parsing accuracy.

[0005] The first aspect of the present invention provides an intelligent address resolution method, including: Constructing an identification model, and obtaining the target address on the logistics label by using the constructed identification model; Constructing an address-coordinate two-way mapping system for obtaining the geographical coordinates of the target address; Extracting the address entity information in the target address, and mapping the extracted address entity information to a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector; Using the R-tree spatial index algorithm to construct a regional grid, retrieving the grid cell to which the target address belongs based on the regional grid; obtaining the surrounding POIs that meet the preset requirements based on the grid cell to which the target address belongs; Calculating the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements respectively based on the geographical coordinates and standardized address vector of the target address, and extracting the attribute features of the surrounding POIs that meet the preset requirements, and generating a multi-dimensional spatial feature vector based on the Euclidean distance, semantic similarity and attribute features; Based on standardized address vectors and multidimensional spatial feature vectors, hierarchical modeling is used to capture address sequence dependencies, the relationship between the region and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements. Text feature vectors are obtained based on the captured address sequence dependencies. The multidimensional spatial feature vectors are enhanced based on the relationship between the region and surrounding POIs that meet preset requirements to obtain enhanced spatial feature vectors. Based on the one-hot encoding of target POIs that meet preset requirements, the text feature vectors and enhanced spatial feature vectors are fused through an attention mechanism to obtain the final address representation vector.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the step of constructing an identification model and using the constructed identification model to obtain the target address on the logistics waybill includes: Obtain the original image of the logistics waybill, and preprocess the obtained original image to obtain the preprocessed image; The preprocessed image is then used to detect and locate the address bar region using a single-stage object detection algorithm. The address bar area is tilted and corrected using region of interest alignment technology to obtain the corrected address bar area. The corrected address bar region is used for feature extraction through a convolutional recurrent neural network to obtain a feature sequence; The feature sequence is used to obtain the complete address text through a concatenation-time classification algorithm; The preprocessed image is used to detect and locate the address bar region through a single-stage target detection algorithm, and the method also includes fusing multi-scale features through a feature pyramid network.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the construction of the address-coordinate bidirectional mapping system includes: a one-way conversion from the target address to geographic coordinates and a reverse resolution from geographic coordinates to the target address; The one-way conversion from target address to geographic coordinates includes: using an N-gram model to correct the target address to obtain a corrected target address; using a trie to match a standard address library to obtain a corresponding standard address; and using the standard address to obtain the corresponding geographic coordinates. The reverse resolution from geographic coordinates to target address includes: dividing geographic regions using Thiessen polygons, with each region uniquely corresponding to a network address; determining the polygon to which the geographic coordinates to be resolved belong, thereby obtaining the network address corresponding to the current polygon; When constructing the address-coordinate bidirectional mapping system, Harman coding is used to compress the mapping relationship.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the step of extracting address entity information from the target address and mapping the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector includes: Construct an entity information recognition model to extract address entity information from the target address; The entity information recognition model includes an architecture that combines the BERT model with a bidirectional long short-term memory network and a conditional random field to recognize address entity information in the target address. The BERT model is used to capture deep semantics in the target address; the bidirectional long short-term memory network is used to process the sequence features in the target address; and the conditional random field is used to optimize boundary recognition. Meanwhile, for nested entities, higher attention weights are assigned through an attention mechanism to enhance the ability to identify nested entities; an entity enhancement module is set up, including an administrative division dictionary and a logistics domain knowledge graph, which are used to match the identified entities with the knowledge graph; A five-layer tree-structured encoder is constructed to map address entity information into a multi-dimensional vector space. Graph embedding technology is used to capture semantic relationships between entities, and weights are assigned to different levels through a hierarchical attention mechanism to finally generate standardized address vectors.

[0009] Optionally, in the fourth implementation of the first aspect of the present invention, the surrounding POIs that meet the preset requirements include: using a quadtree for spatial partitioning and combining it with timestamps to realize dynamic data management of POIs; constructing a spatiotemporal index structure based on the dynamic data of POIs to obtain POIs within a certain range around the target address; A similar hash algorithm is used to generate a feature hash for the attributes and location of POIs. The Hamming distance is calculated based on the feature hash, and the Hamming distance is used to determine whether the POI has changed. When the POI changes, the Delaunay triangulation algorithm is used to take advantage of the property that POIs with shared triangle edges are adjacent to each other. The triangulation is locally updated when the POI changes, and the adjacency relationship is dynamically maintained.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the Euclidean distance and semantic similarity with surrounding POIs that meet preset requirements are calculated based on the geographic coordinates and standardized address vector of the target address, respectively, and attribute features of surrounding POIs that meet preset requirements are extracted. A multi-dimensional spatial feature vector is generated based on the Euclidean distance, semantic similarity, and attribute features, including: Convert the geographic coordinates of the target address to UTM projected coordinates, and calculate the Euclidean distance to the surrounding POIs that meet the preset requirements based on the UTM projected coordinates; Calculate the semantic similarity between the target address and the names of surrounding POIs that meet preset requirements; Extract the attribute features of surrounding POIs that meet preset requirements, including static attributes and dynamic attributes; wherein, the static data includes: POI service type, POI service range, and regional affiliation; the dynamic attributes include: real-time pedestrian flow at the POI; The obtained Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements are normalized to obtain the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements; a multidimensional spatial feature vector is constructed based on the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements.

[0011] Optionally, in the sixth implementation of the first aspect of the present invention, the step of capturing address sequence dependencies, the relationship between the region and surrounding POIs that meet preset requirements, and the one-hot encoding of the target POI that meets preset requirements through hierarchical modeling based on standardized address vectors and multi-dimensional spatial feature vectors; obtaining text feature vectors based on captured address sequence dependencies; strengthening the multi-dimensional spatial feature vectors based on the relationship between the region and surrounding POIs that meet preset requirements to obtain strengthened spatial feature vectors; and fusing the text feature vectors and strengthened spatial feature vectors through an attention mechanism based on the one-hot encoding of the target POI that meets preset requirements to obtain the final address representation vector, including: A three-layer Transformer architecture is used to implement hierarchical modeling, including: the bottom layer uses BiLSTM based on standardized address vectors to capture address sequence dependencies; the middle layer embeds the relationship between related regions and surrounding POIs that meet preset requirements through knowledge graph embedding; Text feature vectors are obtained based on the captured address sequence dependencies; The enhanced spatial feature vector is obtained by strengthening the multidimensional spatial feature vector based on the relationship between the region and the surrounding POIs that meet the preset requirements. The top layer selects target POIs based on text feature vectors and enhanced spatial feature vectors according to preset requirements, and performs one-hot encoding to obtain the one-hot encoding of the target POI; The association weights of the text feature vector and the enhanced spatial feature vector are clearly assigned based on the one-hot encoding of the target POI; the text feature vector and the enhanced spatial feature vector are fused based on the association weights through an attention mechanism to generate the final address representation vector.

[0012] A second aspect of the present invention provides an intelligent address resolution device, comprising: The target address identification module is used to build an identification model and use the built identification model to obtain the target address on the logistics waybill; The geographic coordinate acquisition module is used to build a two-way address-coordinate mapping system to obtain the geographic coordinates of the target address; The standardized address vector generation module is used to extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector. The surrounding POI acquisition module is used to construct a regional grid using the R-tree spatial indexing algorithm, retrieve the grid cell to which the target address belongs based on the regional grid, and acquire surrounding POIs that meet preset requirements based on the grid cell to which the target address belongs. The multidimensional spatial feature vector generation module is used to calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements based on the geographic coordinates and standardized address vector of the target address, and to extract the attribute features of the surrounding POIs that meet the preset requirements. Based on the Euclidean distance, semantic similarity and attribute features, a multidimensional spatial feature vector is generated. The final address representation vector generation module is used to capture address sequence dependencies, the relationship between the region and surrounding POIs that meet preset requirements, and the one-hot encoding of the target POI that meets preset requirements, based on standardized address vectors and multi-dimensional spatial feature vectors through hierarchical modeling. It obtains text feature vectors based on captured address sequence dependencies; strengthens the multi-dimensional spatial feature vectors based on the relationship between the region and surrounding POIs that meet preset requirements to obtain strengthened spatial feature vectors; and fuses the text feature vectors and strengthened spatial feature vectors through an attention mechanism based on the one-hot encoding of the target POI that meets preset requirements to obtain the final address representation vector.

[0013] Optionally, in a first implementation of the second aspect of the present invention, the target address identification module includes: The preprocessing submodule is used to acquire the original image of the logistics waybill, preprocess the acquired original image, and obtain the preprocessed image. The address bar region acquisition submodule is used to obtain the address bar region by performing address bar detection and localization on the preprocessed image through a single-stage object detection algorithm. The correction submodule is used to correct the tilt of the address bar area using region of interest alignment technology to obtain the corrected address bar area. The feature sequence acquisition submodule is used to extract features from the corrected address bar region through a convolutional recurrent neural network to obtain a feature sequence. The address text acquisition submodule is used to obtain the complete address text from the feature sequence through a connection-time classification algorithm. The preprocessed image is used to detect and locate the address bar region through a single-stage target detection algorithm, and the method also includes fusing multi-scale features through a feature pyramid network.

[0014] Optionally, in a second implementation of the second aspect of the present invention, the construction of the address-coordinate bidirectional mapping system includes: a one-way conversion from the target address to geographic coordinates and a reverse resolution from geographic coordinates to the target address; The one-way conversion from target address to geographic coordinates includes: using an N-gram model to correct the target address to obtain a corrected target address; using a trie to match a standard address library to obtain a corresponding standard address; and using the standard address to obtain the corresponding geographic coordinates. The reverse resolution from geographic coordinates to target address includes: dividing geographic regions using Thiessen polygons, with each region uniquely corresponding to a network address; determining the polygon to which the geographic coordinates to be resolved belong, thereby obtaining the network address corresponding to the current polygon; When constructing the address-coordinate bidirectional mapping system, Harman coding is used to compress the mapping relationship.

[0015] Optionally, in a third implementation of the second aspect of the present invention, the standardized address vector generation module includes: Construct an entity information recognition model to extract address entity information from the target address; The entity information recognition model includes an architecture that combines the BERT model with a bidirectional long short-term memory network and a conditional random field to recognize address entity information in the target address. The BERT model is used to capture deep semantics in the target address; the bidirectional long short-term memory network is used to process the sequence features in the target address; and the conditional random field is used to optimize boundary recognition. Meanwhile, for nested entities, higher attention weights are assigned through an attention mechanism to enhance the ability to identify nested entities; an entity enhancement module is set up, including an administrative division dictionary and a logistics domain knowledge graph, which are used to match the identified entities with the knowledge graph; A five-layer tree-structured encoder is constructed to map address entity information into a multi-dimensional vector space. Graph embedding technology is used to capture semantic relationships between entities, and weights are assigned to different levels through a hierarchical attention mechanism to finally generate standardized address vectors.

[0016] Optionally, in the fourth implementation of the second aspect of the present invention, the surrounding POIs that meet the preset requirements include: using a quadtree for spatial partitioning and combining it with timestamps to realize dynamic data management of POIs; constructing a spatiotemporal index structure based on the dynamic data of POIs to obtain POIs within a certain range around the target address; A similar hash algorithm is used to generate a feature hash for the attributes and location of POIs. The Hamming distance is calculated based on the feature hash, and the Hamming distance is used to determine whether the POI has changed. When the POI changes, the Delaunay triangulation algorithm is used to take advantage of the property that POIs with shared triangle edges are adjacent to each other. The triangulation is locally updated when the POI changes, and the adjacency relationship is dynamically maintained.

[0017] Optionally, in a fifth implementation of the second aspect of the present invention, the multidimensional spatial feature vector generation module includes: The Euclidean distance calculation submodule is used to convert the geographic coordinates of the target address into UTM projected coordinates, and calculate the Euclidean distance to the surrounding POIs that meet the preset requirements based on the UTM projected coordinates. The semantic similarity acquisition submodule is used to calculate the semantic similarity between the target address and the names of surrounding POIs that meet preset requirements; The attribute feature acquisition submodule is used to extract attribute features of surrounding POIs that meet preset requirements, including static attributes and dynamic attributes; wherein, the static data includes: POI business type, POI service range and regional affiliation; the dynamic attributes include: real-time pedestrian flow of POI; The spatial feature vector acquisition submodule is used to normalize the acquired Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet preset requirements, to obtain the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet preset requirements; and to construct a multidimensional spatial feature vector based on the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet preset requirements.

[0018] Optionally, in a sixth implementation of the second aspect of the present invention, the final address representation vector generation module includes: A three-layer Transformer architecture is used to implement hierarchical modeling, including: the bottom layer uses BiLSTM based on standardized address vectors to capture address sequence dependencies; the middle layer embeds the relationship between related regions and surrounding POIs that meet preset requirements through knowledge graph embedding; Text feature vectors are obtained based on the captured address sequence dependencies; The enhanced spatial feature vector is obtained by strengthening the multidimensional spatial feature vector based on the relationship between the region and the surrounding POIs that meet the preset requirements. The top layer selects target POIs based on text feature vectors and enhanced spatial feature vectors according to preset requirements, and performs one-hot encoding to obtain the one-hot encoding of the target POI; The association weights of the text feature vector and the enhanced spatial feature vector are clearly assigned based on the one-hot encoding of the target POI; the text feature vector and the enhanced spatial feature vector are fused based on the association weights through an attention mechanism to generate the final address representation vector.

[0019] A third aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the intelligent address resolution method described above.

[0020] 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, wherein the computer program, when executed by the processor, implements the steps of the intelligent address resolution method described above.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a single-stage target detection algorithm combined with a feature pyramid network to accurately locate the address bar, and uses a convolutional recurrent neural network and a connection-time classification algorithm to process multiple font variations and tilted waybills, which solves the problem of low accuracy of traditional OCR in recognizing non-standard waybills, improves the accuracy of logistics waybill address recognition, and significantly reduces the cost of manual verification. 2. This invention uses an N-gram model to correct address errors, a trie to match a standard address library, and Thiessen polygons to achieve inverse coordinate parsing, solving the problem that traditional keyword matching cannot handle fuzzy addresses and abbreviations; at the same time, it reduces storage space and improves the efficiency and accuracy of bidirectional address-coordinate conversion by compressing the mapping relationship through Huffman coding. 3. This invention extracts address entities based on the BERT+BiLSTM+CRF architecture and generates standardized address vectors through a five-layer tree structure encoder, effectively capturing the hierarchical dependency relationship of "province-city-district-street-house number". This solves the problem that traditional methods cannot map unstructured addresses to a standardized hierarchical structure, and provides a unified address semantic benchmark for downstream businesses. 4. This invention uses a quadtree and timestamp to construct a spatiotemporal index, and combines similar hashing algorithm and Delaunay triangulation to realize real-time detection of POI changes and dynamic maintenance of adjacency relationships. This enables the POI retrieval response time within a 500-meter radius to be controlled at the millisecond level, solving the problems of traditional spatial indexes being unable to handle dynamic changes of POIs and having low retrieval efficiency. 5. This invention integrates textual and spatial features through a hierarchical Transformer architecture and dynamically allocates weights using an attention mechanism, so that the final address representation vector simultaneously contains address semantic information and spatial association with surrounding POIs. This solves the limitations of traditional methods that process text or geographic information in isolation, providing more accurate decision-making basis for downstream applications such as logistics sorting and route planning, and improving the efficiency of intelligent delivery. Attached Figure Description

[0022] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a first flowchart of the intelligent address resolution method provided in an embodiment of the present invention.

[0023] Figure 2 This is a second flowchart of the intelligent address resolution method provided in an embodiment of the present invention.

[0024] Figure 3 This is a third flowchart of the intelligent address resolution method provided in an embodiment of the present invention.

[0025] Figure 4 This is a fourth flowchart of the intelligent address resolution method provided in an embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram of a smart address resolution device provided in an embodiment of the present invention.

[0027] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] This invention provides an intelligent address resolution method, apparatus, device, and medium, comprising: constructing an identification model; using the constructed identification model to obtain the target address on a logistics waybill; constructing an address-coordinate bidirectional mapping system to obtain the geographic coordinates of the target address; extracting address entity information from the target address and mapping the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector; constructing a regional grid using the R-tree spatial indexing algorithm and retrieving the grid cell to which the target address belongs based on the regional grid; obtaining surrounding POIs that meet preset requirements based on the grid cell to which the target address belongs; and calculating the Euclidean coordinates of the target address and the surrounding POIs that meet the preset requirements based on the geographic coordinates of the target address and the standardized address vector. This invention utilizes distance and semantic similarity to extract attribute features from surrounding POIs that meet preset requirements. A multi-dimensional spatial feature vector is generated based on Euclidean distance, semantic similarity, and attribute features. Based on standardized address vectors and the multi-dimensional spatial feature vector, hierarchical modeling is used to capture address sequence dependencies, the relationship between the region and surrounding POIs meeting preset requirements, and the one-hot encoding of the target POI meeting preset requirements. Text feature vectors are obtained based on the captured address sequence dependencies. The multi-dimensional spatial feature vector is strengthened based on the relationship between the region and surrounding POIs meeting preset requirements to obtain a strengthened spatial feature vector. Finally, the text feature vector and the strengthened spatial feature vector are fused using an attention mechanism based on the one-hot encoding of the target POI meeting preset requirements to obtain the final address representation vector. This invention solves the problem that existing technologies often rely on simple keyword matching or rule templates, can only handle structured addresses with standardized formats, and have insufficient parsing accuracy.

[0029] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The first embodiment of the intelligent address resolution method in this invention includes: 101. Construct an identification model and use the constructed identification model to obtain the target address on the logistics waybill; In this embodiment, the recognition model adopts end-to-end optical character recognition technology, combined with convolutional recurrent neural network and connection-time classification algorithm, and is trained on hundreds of thousands of logistics waybill images to achieve high-precision recognition of 72 font variants. More specifically, the original image of the logistics waybill is obtained, and the obtained original image is preprocessed to obtain the preprocessed image; The preprocessed image is used to detect and locate the address bar region through a single-stage target detection algorithm; the single-stage target detection algorithm also includes fusing multi-scale features through a feature pyramid network to improve the detection capability of small target text. The address bar area is corrected for tilt using region of interest alignment technology to obtain the corrected address bar area, ensuring the correct recognition of tilted waybills; The corrected address bar region is used for feature extraction through a convolutional recurrent neural network to obtain a feature sequence; The feature sequence is used to obtain the complete address text through a connection-based temporal classification algorithm.

[0031] 102. Construct a bidirectional address-coordinate mapping system to obtain the geographic coordinates of the target address; In this embodiment, a one-way conversion from target address to geographic coordinates and a reverse resolution from geographic coordinates to target address are performed. The one-way conversion from target address to geographic coordinates includes: using an N-gram model to correct the target address to obtain a corrected target address; using a trie to match a standard address library to obtain a corresponding standard address; and using the standard address to obtain the corresponding geographic coordinates. The reverse resolution from geographic coordinates to target address includes: dividing geographic regions using Thiessen polygons, with each region uniquely corresponding to a network point address; determining the polygon to which the geographic coordinates to be resolved belong, thereby obtaining the network point address corresponding to the current polygon and realizing coordinate error correction; When constructing the address-coordinate bidirectional mapping system, Harman coding is used to compress the mapping relationship, reduce the storage space of the mapping relationship, and improve query efficiency.

[0032] 103. Extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector; In this embodiment, an entity information recognition model is constructed to extract address entity information from the target address; The entity information recognition model includes an architecture that combines the BERT model with a bidirectional long short-term memory network and a conditional random field to recognize address entity information in the target address. The BERT model is used to capture deep semantics in the target address; the bidirectional long short-term memory network is used to process the sequence features in the target address; and the conditional random field is used to optimize boundary recognition. Meanwhile, for nested entities, higher attention weights are assigned through an attention mechanism to enhance the ability to identify nested entities; an entity enhancement module is set up, including an administrative division dictionary and a logistics domain knowledge graph, which are used to match the identified entities with the knowledge graph; A five-layer tree-structured encoder is constructed to map address entity information into a multi-dimensional vector space. Graph embedding technology is used to capture semantic relationships between entities, and weights are assigned to different levels through a hierarchical attention mechanism to finally generate standardized address vectors.

[0033] 104. Construct a regional grid using the R-tree spatial indexing algorithm, and retrieve the grid cell to which the target address belongs based on the regional grid; obtain the surrounding POIs that meet the preset requirements based on the grid cell to which the target address belongs; In this embodiment, an R-tree spatial index is used to achieve fast grid retrieval. A grid encoding algorithm is designed to convert coordinates into integer codes, supporting efficient range queries and spatial join operations.

[0034] The surrounding POIs that meet the preset requirements include: using a quadtree for spatial partitioning and combining it with timestamps to achieve dynamic POI data management; constructing a spatiotemporal index structure based on the dynamic POI data to obtain POIs within a certain range around the target address, thereby achieving millimeter-level retrieval of points of interest within a certain distance radius; A similar hash algorithm is used to generate a feature hash for the attributes and location of POIs. The Hamming distance is calculated based on the feature hash, and the Hamming distance is used to determine whether the POI has changed. When the POI changes, the Delaunay triangulation algorithm is used to take advantage of the property that POIs with shared triangle edges are adjacent to each other. The triangulation is locally updated when the POI changes, and the adjacency relationship is dynamically maintained.

[0035] 105. Calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements based on the geographic coordinates and standardized address vector of the target address, respectively, and extract the attribute features of the surrounding POIs that meet the preset requirements. Generate a multi-dimensional spatial feature vector based on the Euclidean distance, semantic similarity and attribute features. In this embodiment, the geographic coordinates of the target address are converted into UTM projected coordinates, and the Euclidean distance to the surrounding POIs that meet the preset requirements is calculated based on the UTM projected coordinates. The semantic space of the logistics domain is trained based on word vector model, and the semantic similarity between the target address and the names of surrounding POIs that meet the preset requirements is calculated by cosine similarity. Extract the attribute features of surrounding POIs that meet preset requirements, including static attributes and dynamic attributes; wherein, the static data includes: POI service type, POI service range, and regional affiliation; the dynamic attributes include: real-time pedestrian flow at the POI; The obtained Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements are normalized to obtain the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements; a multidimensional spatial feature vector is constructed based on the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements.

[0036] 106. Based on standardized address vectors and multidimensional spatial feature vectors, hierarchical modeling is used to capture address sequence dependencies, the relationship between regions and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements; text feature vectors are obtained based on the captured address sequence dependencies; the multidimensional spatial feature vectors are enhanced based on the relationship between regions and surrounding POIs that meet preset requirements to obtain enhanced spatial feature vectors; the text feature vectors and enhanced spatial feature vectors are fused through an attention mechanism based on the one-hot encoding of target POIs that meet preset requirements to obtain the final address representation vector; In this embodiment, a three-layer Transformer architecture is used to implement hierarchical modeling, including: the bottom layer uses BiLSTM based on standardized address vectors to capture address sequence dependencies; the middle layer embeds the relationship between related regions and surrounding POIs that meet preset requirements through knowledge graph embedding; Text feature vectors are obtained based on the captured address sequence dependencies; The enhanced spatial feature vector is obtained by strengthening the multidimensional spatial feature vector based on the relationship between the region and the surrounding POIs that meet the preset requirements. The top layer selects target POIs based on text feature vectors and enhanced spatial feature vectors according to preset requirements, and performs one-hot encoding to obtain the one-hot encoding of the target POI; The association weights of the text feature vector and the enhanced spatial feature vector are clearly assigned based on the one-hot encoding of the target POI; the text feature vector and the enhanced spatial feature vector are fused based on the association weights through an attention mechanism to generate the final address representation vector.

[0037] Please see Figure 2 The second embodiment of the intelligent address resolution method in this invention includes: 201. Construct an identification model and use the constructed identification model to obtain the target address on the logistics waybill; 202. Construct a bidirectional address-coordinate mapping system to obtain the geographic coordinates of the target address; 203. Extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector; 204. Construct a regional grid using the R-tree spatial indexing algorithm, and retrieve the grid cell to which the target address belongs based on the regional grid; obtain the surrounding POIs that meet the preset requirements based on the grid cell to which the target address belongs; In this embodiment, obtaining surrounding POIs that meet preset requirements includes: Receiving addresses typically include province, city, district, street, house number, and surrounding landmark information, such as "XX Building" or "XX Shopping Mall." First, the address needs to be broken down to extract key components. For example, "No. 1 Zhongguancun Street, Haidian District, Beijing, Xinzhongguan Shopping Center" can be broken down into "Zhongguancun Street," "No. 1," and "Xinzhongguan Shopping Center." Many addresses directly contain AOI names. For example, names like "Wanda Plaza" and "XX University" are themselves clear points of interest. By establishing a database of common AOI keywords, such as "building," "school," and "hospital," candidate names can be directly matched from the address text; for example, "Tencent Building" in "Tencent Building, Science Park Road" can be used as an AOI point.

[0038] If an address doesn't explicitly mention a landmark, it can be converted to latitude and longitude and then linked to surrounding POI data. For example, the address "No. 93 Jianguo Road, Chaoyang District" can be linked to "Beijing Wanda Plaza" through location services because the actual building name corresponding to these coordinates is included in the map data. Some AOI names are hidden within area divisions. For example, "Alibaba Xixi Park" in "Alibaba Xixi Park, Yuhang District, Hangzhou" is a corporate park-level AOI. By identifying company names, community names, or specific area names in the address, such as "Financial Street" or "Software Park," it can be directly linked to a pre-defined AOI database. Multiple addresses in the same area may point to the same AOI. For example, "No. 1288 Lujiazui Ring Road, Pudong New Area, Shanghai" and "Shanghai Tower" both point to the same landmark. By establishing a mapping database between addresses and AOIs and regularly updating it with new landmarks or changes, matching accuracy can be improved.

[0039] 205. Calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements based on the geographic coordinates and standardized address vector of the target address, respectively, and extract the attribute features of the surrounding POIs that meet the preset requirements. Generate a multi-dimensional spatial feature vector based on the Euclidean distance, semantic similarity and attribute features. 206. Based on standardized address vectors and multidimensional spatial feature vectors, hierarchical modeling is used to capture address sequence dependencies, the relationship between regions and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements; text feature vectors are obtained based on the captured address sequence dependencies; the multidimensional spatial feature vectors are enhanced based on the relationship between regions and surrounding POIs that meet preset requirements to obtain enhanced spatial feature vectors; the text feature vectors and enhanced spatial feature vectors are fused through an attention mechanism based on the one-hot encoding of target POIs that meet preset requirements to obtain the final address representation vector.

[0040] Please see Figure 3 The third embodiment of the intelligent address resolution method in this invention includes: 301. Obtain the final address representation vector based on the intelligent address resolution method; 302. Perform sorting operations based on the obtained final address representation vector; 303. Establish a distributed log collection system to collect dynamic sorting behavior data generated during the sorting process based on the constructed distributed log collection system; In this embodiment, a distributed log collection system is established, using message queues to achieve real-time collection of high-concurrency sorting behavior data. Data cleaning and transformation are performed through a log processing framework, and Bloom filters are used to quickly identify abnormal operations. Geo-hash encoding is used to compress geographic location information to reduce storage costs. The offline cache adopts a two-level caching strategy: the first-level cache stores frequently accessed data based on an in-memory database, and the second-level cache stores all historical data based on an embedded database. Incremental synchronization mechanism ensures data consistency and supports efficient querying in weak network environments.

[0041] Please see Figure 4 The fourth embodiment of the intelligent address resolution method in this invention includes: 401. Obtain the final address representation vector based on the intelligent address resolution method; 402. Construct a Markov decision process. The state space includes the final address representation vector, historical correction records, and area familiarity. The action space is a set of candidate AOIs. The ε-greedy strategy is used to balance exploration and utilization. The state value function is updated through the Q-learning algorithm. A hierarchical reward mechanism is designed to realize dynamic optimization of sorting decisions based on real-time state.

[0042] The intelligent address resolution method in the embodiments of the present invention has been described above. The intelligent address resolution device in the embodiments of the present invention will be described below. Please refer to [link / reference]. Figure 5 One embodiment of the intelligent address resolution device in this invention includes: The target address identification module 501 is used to build an identification model and use the built identification model to obtain the target address on the logistics waybill; In this embodiment, the target address identification module 501 includes: The preprocessing submodule 5011 is used to acquire the original image of the logistics waybill, preprocess the acquired original image, and obtain the preprocessed image. Address bar region acquisition submodule 5012 is used to obtain the address bar region by performing address bar detection and localization on the preprocessed image through a single-stage target detection algorithm. The correction submodule 5013 is used to perform tilt correction processing on the address bar region using region of interest alignment technology to obtain the corrected address bar region. The feature sequence acquisition submodule 5014 is used to extract features from the corrected address bar region through a convolutional recurrent neural network to obtain a feature sequence. Address text acquisition submodule 5015 is used to obtain complete address text from feature sequences through a connection-time classification algorithm; The preprocessed image is used to detect and locate the address bar region through a single-stage target detection algorithm, and the method also includes fusing multi-scale features through a feature pyramid network.

[0043] The geographic coordinate acquisition module 502 is used to construct a bidirectional address-coordinate mapping system to obtain the geographic coordinates of the target address; In this embodiment, the geographic coordinate acquisition module 502 includes: The construction of the address-coordinate bidirectional mapping system includes: one-way conversion from target address to geographic coordinates and reverse resolution from geographic coordinates to target address; The one-way conversion from target address to geographic coordinates includes: using an N-gram model to correct the target address to obtain a corrected target address; using a trie to match a standard address library to obtain a corresponding standard address; and using the standard address to obtain the corresponding geographic coordinates. The reverse resolution from geographic coordinates to target address includes: dividing geographic regions using Thiessen polygons, with each region uniquely corresponding to a network address; determining the polygon to which the geographic coordinates to be resolved belong, thereby obtaining the network address corresponding to the current polygon; When constructing the address-coordinate bidirectional mapping system, Harman coding is used to compress the mapping relationship.

[0044] The standardized address vector generation module 503 is used to extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector. In this embodiment, the standardized address vector generation module 503 includes: Construct an entity information recognition model to extract address entity information from the target address; The entity information recognition model includes an architecture that combines the BERT model with a bidirectional long short-term memory network and a conditional random field to recognize address entity information in the target address. The BERT model is used to capture deep semantics in the target address; the bidirectional long short-term memory network is used to process the sequence features in the target address; and the conditional random field is used to optimize boundary recognition. Meanwhile, for nested entities, higher attention weights are assigned through an attention mechanism to enhance the ability to identify nested entities; an entity enhancement module is set up, including an administrative division dictionary and a logistics domain knowledge graph, which are used to match the identified entities with the knowledge graph; A five-layer tree-structured encoder is constructed to map address entity information into a multi-dimensional vector space. Graph embedding technology is used to capture semantic relationships between entities, and weights are assigned to different levels through a hierarchical attention mechanism to finally generate standardized address vectors.

[0045] The surrounding POI acquisition module 504 is used to construct a regional grid using the R-tree spatial indexing algorithm, retrieve the grid cell to which the target address belongs based on the regional grid, and acquire surrounding POIs that meet preset requirements based on the grid cell to which the target address belongs. In this embodiment, the surrounding POIs that meet the preset requirements include: using a quadtree for spatial partitioning and combining it with timestamps to achieve dynamic POI data management; and constructing a spatiotemporal index structure based on the dynamic POI data to obtain POIs within a certain range around the target address. A similar hash algorithm is used to generate a feature hash for the attributes and location of POIs. The Hamming distance is calculated based on the feature hash, and the Hamming distance is used to determine whether the POI has changed. When the POI changes, the Delaunay triangulation algorithm is used to take advantage of the property that POIs with shared triangle edges are adjacent to each other. The triangulation is locally updated when the POI changes, and the adjacency relationship is dynamically maintained.

[0046] The multidimensional spatial feature vector generation module 505 is used to calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements based on the geographic coordinates and standardized address vector of the target address, and to extract the attribute features of the surrounding POIs that meet the preset requirements. Based on the Euclidean distance, semantic similarity and attribute features, a multidimensional spatial feature vector is generated. In this embodiment, the multidimensional spatial feature vector generation module 505 includes: The Euclidean distance calculation submodule 5051 is used to convert the geographic coordinates of the target address into UTM projected coordinates, and calculate the Euclidean distance with the surrounding POIs that meet the preset requirements based on the UTM projected coordinates. The semantic similarity acquisition submodule 5052 is used to calculate the semantic similarity between the target address and the names of surrounding POIs that meet preset requirements; The attribute feature acquisition submodule 5053 is used to extract attribute features of surrounding POIs that meet preset requirements, including static attributes and dynamic attributes; wherein, the static data includes: POI business type, POI service range and regional affiliation; the dynamic attributes include: real-time pedestrian flow of POI; The spatial feature vector acquisition submodule 5054 is used to normalize the acquired Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements, to obtain the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements; and to construct a multidimensional spatial feature vector based on the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements.

[0047] The final address representation vector generation module 506 is used to capture address sequence dependencies, the relationship between the region and surrounding POIs that meet preset requirements, and the one-hot encoding of the target POI that meets preset requirements through hierarchical modeling based on standardized address vectors and multi-dimensional spatial feature vectors; obtain text feature vectors based on captured address sequence dependencies; enhance the multi-dimensional spatial feature vectors based on the relationship between the region and surrounding POIs that meet preset requirements to obtain enhanced spatial feature vectors; and fuse the text feature vectors and enhanced spatial feature vectors through an attention mechanism based on the one-hot encoding of the target POI that meets preset requirements to obtain the final address representation vector. In this embodiment, the final address representation vector generation module 506 includes: A three-layer Transformer architecture is used to implement hierarchical modeling, including: the bottom layer uses BiLSTM based on standardized address vectors to capture address sequence dependencies; the middle layer embeds the relationship between related regions and surrounding POIs that meet preset requirements through knowledge graph embedding; Text feature vectors are obtained based on the captured address sequence dependencies; The enhanced spatial feature vector is obtained by strengthening the multidimensional spatial feature vector based on the relationship between the region and the surrounding POIs that meet the preset requirements. The top layer selects target POIs based on text feature vectors and enhanced spatial feature vectors according to preset requirements, and performs one-hot encoding to obtain the one-hot encoding of the target POI; The association weights of the text feature vector and the enhanced spatial feature vector are clearly assigned based on the one-hot encoding of the target POI; the text feature vector and the enhanced spatial feature vector are fused based on the association weights through an attention mechanism to generate the final address representation vector.

[0048] above Figure 5 The intelligent address resolution device in the embodiments of the present invention will be described in detail from the perspective of modular functional entities. The electronic device in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0049] Figure 6This is a schematic diagram of the structure of an electronic device 700 provided in an embodiment of the present invention. The electronic device 700 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 710 (e.g., one or more processors) and a memory 720, and one or more storage media 730 (e.g., one or more mass storage devices) for storing application programs 733 or data 732. The memory 720 and storage media 730 can be temporary or persistent storage. The program stored in the storage media 730 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the electronic device 700. Furthermore, the processor 710 may be configured to communicate with the storage media 730 and execute the series of instruction operations in the storage media 730 on the electronic device 700.

[0050] Electronic device 700 may also include one or more power supplies 740, one or more wired or wireless network interfaces 750, one or more input / output interfaces 750, and / or one or more operating systems 731, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 6 The illustrated electronic device structure does not constitute a limitation on electronic devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0051] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the intelligent address resolution method.

[0052] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0053] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0054] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart address resolution method, characterized in that, include: Construct a recognition model and use it to obtain the target address on the logistics waybill; Construct a bidirectional address-coordinate mapping system to obtain the geographic coordinates of a target address; Extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector; The R-tree spatial indexing algorithm is used to construct a regional grid, and the grid cell to which the target address belongs is retrieved based on the regional grid; Obtain surrounding POIs that meet preset requirements based on the grid cell to which the target address belongs; Based on the geographic coordinates and standardized address vector of the target address, calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements, extract the attribute features of the surrounding POIs that meet the preset requirements, and generate a multi-dimensional spatial feature vector based on the Euclidean distance, semantic similarity and attribute features. Based on standardized address vectors and multi-dimensional spatial feature vectors, hierarchical modeling is used to capture address sequence dependencies, the relationship between regions and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements. The text feature vector is obtained based on the captured address sequence dependency; the multi-dimensional spatial feature vector is enhanced based on the relationship between the region and the surrounding POIs that meet the preset requirements; the text feature vector and the enhanced spatial feature vector are fused through an attention mechanism based on the one-hot encoding of the target POI that meets the preset requirements to obtain the final address representation vector.

2. The intelligent address resolution method according to claim 1, characterized in that, The construction of the identification model, which is used to obtain the target address on the logistics waybill, includes: Obtain the original image of the logistics waybill, and preprocess the obtained original image to obtain the preprocessed image; The preprocessed image is then used to detect and locate the address bar region using a single-stage object detection algorithm. The address bar area is tilted and corrected using region of interest alignment technology to obtain the corrected address bar area. The corrected address bar region is used for feature extraction through a convolutional recurrent neural network to obtain a feature sequence; The feature sequence is used to obtain the complete address text through a concatenation-time classification algorithm; The preprocessed image is used to detect and locate the address bar region through a single-stage target detection algorithm, and the method also includes fusing multi-scale features through a feature pyramid network.

3. The intelligent address resolution method according to claim 1, characterized in that, The construction of the address-coordinate bidirectional mapping system includes: one-way conversion from target address to geographic coordinates and reverse resolution from geographic coordinates to target address; The one-way conversion from target address to geographic coordinates includes: using an N-gram model to correct the target address to obtain a corrected target address; using a trie to match a standard address library to obtain a corresponding standard address; and using the standard address to obtain the corresponding geographic coordinates. The reverse resolution from geographic coordinates to target address includes: dividing geographic regions using Thiessen polygons, with each region uniquely corresponding to a network address; determining the polygon to which the geographic coordinates to be resolved belong, thereby obtaining the network address corresponding to the current polygon; When constructing the address-coordinate bidirectional mapping system, Harman coding is used to compress the mapping relationship.

4. The intelligent address resolution method according to claim 1, characterized in that, The step of extracting address entity information from the target address and mapping the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector includes: Construct an entity information recognition model to extract address entity information from the target address; The entity information recognition model includes an architecture that combines the BERT model with a bidirectional long short-term memory network and a conditional random field to recognize address entity information in the target address. The BERT model is used to capture deep semantics in the target address; the bidirectional long short-term memory network is used to process the sequence features in the target address; and the conditional random field is used to optimize boundary recognition. Meanwhile, for nested entities, higher attention weights are assigned through an attention mechanism to enhance the ability to identify nested entities; an entity enhancement module is set up, including an administrative division dictionary and a logistics domain knowledge graph, which are used to match the identified entities with the knowledge graph; A five-layer tree-structured encoder is constructed to map address entity information into a multi-dimensional vector space. Graph embedding technology is used to capture semantic relationships between entities, and weights are assigned to different levels through a hierarchical attention mechanism to finally generate standardized address vectors.

5. The intelligent address resolution method according to claim 1, characterized in that, The surrounding POIs that meet the preset requirements include: using a quadtree for spatial partitioning and combining it with timestamps to achieve dynamic POI data management; and constructing a spatiotemporal index structure based on the dynamic POI data to obtain POIs within a certain range around the target address. A similar hash algorithm is used to generate a feature hash for the attributes and location of POIs. The Hamming distance is calculated based on the feature hash, and the Hamming distance is used to determine whether the POI has changed. When the POI changes, the Delaunay triangulation algorithm is used to take advantage of the property that POIs with shared triangle edges are adjacent to each other. The triangulation is locally updated when the POI changes, and the adjacency relationship is dynamically maintained.

6. The intelligent address resolution method according to claim 1, characterized in that, The Euclidean distance and semantic similarity between the target address's geographic coordinates and standardized address vector and surrounding POIs that meet preset requirements are calculated, and attribute features of surrounding POIs that meet preset requirements are extracted. A multi-dimensional spatial feature vector is generated based on the Euclidean distance, semantic similarity, and attribute features, including: Convert the geographic coordinates of the target address to UTM projected coordinates, and calculate the Euclidean distance to the surrounding POIs that meet the preset requirements based on the UTM projected coordinates; Calculate the semantic similarity between the target address and the names of surrounding POIs that meet preset requirements; Extract the attribute features of surrounding POIs that meet preset requirements, including static attributes and dynamic attributes; wherein, the static data includes: POI service type, POI service range, and regional affiliation; the dynamic attributes include: real-time pedestrian flow at the POI; The obtained Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements are normalized to obtain the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements; a multidimensional spatial feature vector is constructed based on the processed Euclidean distance, semantic similarity, and attribute features of surrounding POIs that meet the preset requirements.

7. The intelligent address resolution method according to claim 1, characterized in that, The method, based on standardized address vectors and multidimensional spatial feature vectors, captures address sequence dependencies, the relationship between regions and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements through hierarchical modeling. Text feature vectors are obtained based on the captured address sequence dependencies; multi-dimensional spatial feature vectors are enhanced based on the relationship between the region and surrounding POIs that meet preset requirements, resulting in enhanced spatial feature vectors; one-hot encoding of target POIs that meet preset requirements fuses the text feature vectors and enhanced spatial feature vectors through an attention mechanism to obtain the final address representation vector, including: A three-layer Transformer architecture is used to implement hierarchical modeling, including: the bottom layer uses BiLSTM based on standardized address vectors to capture address sequence dependencies; the middle layer embeds the relationship between related regions and surrounding POIs that meet preset requirements through knowledge graph embedding; Text feature vectors are obtained based on the captured address sequence dependencies; The enhanced spatial feature vector is obtained by strengthening the multidimensional spatial feature vector based on the relationship between the region and the surrounding POIs that meet the preset requirements. The top layer selects target POIs based on text feature vectors and enhanced spatial feature vectors according to preset requirements, and performs one-hot encoding to obtain the one-hot encoding of the target POI; The association weights of the text feature vector and the enhanced spatial feature vector are clearly assigned based on the one-hot encoding of the target POI; the text feature vector and the enhanced spatial feature vector are fused based on the association weights through an attention mechanism to generate the final address representation vector.

8. An intelligent address resolution device, characterized in that, include: The target address identification module is used to build an identification model and use the built identification model to obtain the target address on the logistics waybill; The geographic coordinate acquisition module is used to build a two-way address-coordinate mapping system to obtain the geographic coordinates of the target address; The standardized address vector generation module is used to extract address entity information from the target address and map the extracted address entity information into a 5-level tree structure including province-city-district-street-house number to generate a standardized address vector. The surrounding POI acquisition module is used to construct a regional grid using the R-tree spatial indexing algorithm, and to retrieve the grid cell to which the target address belongs based on the regional grid; Obtain surrounding POIs that meet preset requirements based on the grid cell to which the target address belongs; The multidimensional spatial feature vector generation module is used to calculate the Euclidean distance and semantic similarity with the surrounding POIs that meet the preset requirements based on the geographic coordinates and standardized address vector of the target address, and to extract the attribute features of the surrounding POIs that meet the preset requirements. Based on the Euclidean distance, semantic similarity and attribute features, a multidimensional spatial feature vector is generated. The final address representation vector generation module is used to capture address sequence dependencies, the relationship between the region and surrounding POIs that meet preset requirements, and one-hot encoding of target POIs that meet preset requirements based on standardized address vectors and multi-dimensional spatial feature vectors through hierarchical modeling. The text feature vector is obtained based on the captured address sequence dependency; the multi-dimensional spatial feature vector is enhanced based on the relationship between the region and the surrounding POIs that meet the preset requirements; the text feature vector and the enhanced spatial feature vector are fused through an attention mechanism based on the one-hot encoding of the target POI that meets the preset requirements to obtain the final address representation vector.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent address resolution method according to any one of claims 1 to 7.

10. 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 computer program is executed by the processor, it implements the steps of the intelligent address resolution method according to any one of claims 1 to 7.

Images