Creation method for creating building spatiotemporal objects

Abstract

The present invention belongs to the field of computer technology and discloses a method for creating a spatiotemporal building object. The method is based on the following model framework: an address resolution model and an address matching model, wherein the address resolution model and the address matching model share an encoding network, the encoding network input is a text address, and the encoding network output is a text-geographic fusion vector; the address resolution model and the address matching model are trained using a multi-task joint training method; the address resolution model is used to convert address text in multi-source building data into an address label sequence; the address matching model is used to identify address text indicating the same building in the multi-source building data; and the building spatiotemporal building object of the target building includes building information and an address label sequence extracted from the multi-source building data indicating the target building. Thus, multi-task learning is used to achieve accurate semantic parsing and full spatiotemporal association modeling of cross-departmental, multi-source, heterogeneous data.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method for creating a spatiotemporal object of a building. Background Art

[0002] Data on buildings, especially urban buildings, often have multi-dimensional complexity, including physical attribute information (such as building structure, area, and floor height), multi-format address expressions (both standardized and non-standardized), and multi-source business data (such as building maintenance records, elevator installation projects, and historical renovation and expansion information).

[0003] When processing multi-source, heterogeneous building data, it's difficult to effectively integrate geographic knowledge, resulting in inaccurate and inconsistent semantic parsing. This is especially true when data standards aren't unified across departments, leading to low precision in identifying building entities and extracting relationships. Furthermore, given the multidimensional nature of object attribute information, traditional models have limited ability to model the dynamic associations of all spatiotemporal objects, making it difficult to capture the complex semantic relationships of buildings across time and space. This impacts the construction of spatiotemporal knowledge graphs and the ability to align semantics across data sources. Summary of the Invention

[0004] This disclosure section is provided to briefly introduce concepts that will be described in detail in the detailed description section below. This disclosure section is not intended to identify key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0005] In a first aspect, an embodiment of the present application provides a model framework for creating a building spatiotemporal object, the model framework comprising: an address resolution model and an address matching model, wherein the address resolution model and the address matching model share an encoding network, the input of the encoding network is a text address, and the output of the encoding network is a text geographic fusion vector; the address resolution model and the address matching model are trained by multi-task joint training; the address resolution model comprises an encoding network and an address resolution network; the input of the address resolution model is an address text, and the output of the address resolution model is an address label sequence; the address resolution model is used to convert the address text in the building multi-source data into an address label sequence; the address matching model comprises an encoding network and an address matching network; the input of the address matching model is a paired address text, and the output of the address matching model is a relationship category between the paired address texts; the address matching model is used to identify address texts indicating the same building in the building multi-source data; the building spatiotemporal object of the target building comprises building information and an address label sequence extracted from the building multi-source data indicating the target building.

[0006] Optionally, the encoding network includes a word segmenter, an address label embedding module, a geographic coordinate encoding module and a feature extraction layer; the word segmenter is used to convert the input address text into an address text vector; the address label embedding module is used to convert the address elements in the address text into an address label vector; the geographic coordinate encoding module is used to convert the address elements in the address text into a geographic coordinate vector; the input of the feature extraction layer is a text-geometry splicing vector generated based on the address text vector, the address label vector and the geographic coordinate vector, and the output of the feature extraction layer is a text-geometry fusion vector.

[0007] Optionally, the geographic coordinate encoding adopts spatial continuity preservation technology to convert longitude and latitude coordinates into multi-dimensional continuous vectors; the feature extraction layer adopts a 12-layer Transformer encoder structure, each layer has a 768-dimensional hidden state and 12 attention heads.

[0008] Optionally, the building spatiotemporal object includes: building identification, which is the unique identifiable identification of the building object; building spatiotemporal reference, which is the spatiotemporal reference of the building object; building spatial location; building spatial form; building attribute characteristics; building behavioral capabilities, the building's response to changes in the natural environment and to social planning measures; and a set of building changes.

[0009] Optionally, the address resolution network includes an expanded convolutional neural network, a bidirectional long short-term memory network and a conditional random field layer; the expanded convolutional neural network is used to extract local multi-scale features from the text-geography fusion vector output by the encoding network; the bidirectional long short-term memory network is used to extract dependencies in the text-geography fusion vector; the input of the conditional random field layer is a first fusion vector, which is obtained by splicing the outputs of the expanded convolutional neural network and the bidirectional long short-term memory network; the output of the conditional random field layer is an address label sequence.

[0010] Optionally, the input of the address matching network is the paired text-geography fusion vector output by the encoding network; the address matching network includes a pooling layer, a feature interaction layer and a fully connected network; the pooling layer is used to pool the input paired text-geography fusion vector to obtain a first pooling result vector and a second pooling result vector; the feature interaction layer is used to fuse the first pooling result vector and the second pooling result vector to obtain a second fusion vector; the fully connected network is used to map the second fusion vector to the address category matching space.

[0011] Optionally, the pooling layer performs average pooling and maximum pooling, and fuses the results of average pooling and maximum pooling; the feature interaction layer performs: calculating the difference vector between the first pooling result vector and the second pooling result vector, calculating the dot product vector between the first pooling result vector and the second pooling result vector; splicing the first pooling result vector, the second pooling result vector, the difference vector and the dot product vector to obtain a second fused vector.

[0012] Optionally, the loss function of the address resolution model is a conditional random field loss function L parse The loss function of the address matching model adopts the multi-class cross entropy loss function L match The joint training process of the address resolution model and the address matching model includes: constructing a joint loss function; wherein the joint loss function is

[0013] L=αL parse +βL match +λ||θ|| 2

[0014] Among them, α and β are task weights, which are initially set to 0.5, and λ is the regularization coefficient, which is set to 1×10 -5 , θ is the model parameter.

[0015] Optionally, the joint training process of the address resolution model and the address matching model includes: determining a performance change rate during the joint training process; and adjusting the first task weight and the second task weight based on the performance change rate; wherein the performance change rate is defined as:

[0016]

[0017] Among them, perf task，t is the performance indicator of round t, where the accuracy is used for parsing tasks and the F1 score is used for matching tasks;

[0018] Among them, the weight update formula is:

[0019] α t+1 =α t ·(1+γ(Δperf parse -Δperf match ))

[0020] β t+1 =1-α t+1

[0021] Here, γ is the adjustment rate, which is set to 0.1.

[0022] In a second aspect, an embodiment of the present application provides a method for creating a building spatiotemporal object, comprising: obtaining heterogeneous data from different departments and text data containing addresses; for the text data, extracting structured address information through an address resolution model to generate standard address data; inputting the standard address data together with the heterogeneous data from different departments into an address matching model; the address matching model performs matching processing on the received standard address data and the heterogeneous data from different departments; if the match is successful, extracting the matching address pairs, and constructing the building spatiotemporal object based on the matching address pairs and the heterogeneous data corresponding to the address pairs, further integrating the spatiotemporal attributes, and storing the generated spatiotemporal object; if the match fails, recording the unmatched data and feeding it back for manual verification; wherein, the address matching model and / or the address resolution model is determined based on the model framework in any one of the first aspects.

[0023] The model framework and method for creating spatiotemporal building objects provided in this application effectively enhance the generalization and robustness of the model framework through a joint training mechanism, establishing a unified semantic understanding standard for cross-departmental building data. Furthermore, the integration of geographic knowledge encoding networks and multi-task learning significantly improves the accuracy, consistency, and efficiency of semantic parsing and spatiotemporal correlation modeling for cross-departmental, multi-source, and heterogeneous urban building data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and other features, advantages, and aspects of the various embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the originals and elements are not necessarily drawn to scale.

[0025] Figure 1 is a flow chart of an embodiment of a model framework for creating a spatiotemporal object of a building according to the present application;

[0026] Figure 2 It is a schematic diagram of the coding network structure that integrates geographical knowledge;

[0027] Figure 3 This is a schematic diagram of the structure of the address resolution model with multi-scale feature fusion and sequence labeling;

[0028] Figure 4 This is a schematic diagram of the structure of the building address matching model with deep feature interaction;

[0029] Figure 5 is a flowchart of an embodiment of a method for creating a building spatiotemporal object according to the present application;

[0030] Figure 6 It is a schematic diagram of the basic structure of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0031] The following describes embodiments of the present application in more detail with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as limited to the embodiments described herein. Instead, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are for illustrative purposes only and are not intended to limit the scope of protection of the present application.

[0032] It should be understood that the various steps described in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0033] As used herein, the term "including" and its variations are open-ended, i.e., "including but not limited to." The term "based on" means "based at least in part on." The term "one embodiment" means "at least one embodiment," the term "another embodiment" means "at least one additional embodiment," and the term "some embodiments" means "at least some embodiments." Relevant definitions of other terms are provided in the following description.

[0034] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0035] It should be noted that the modifications of "one" and "multiple" mentioned in this application are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0036] The names of the messages or information exchanged between multiple devices in the embodiments of the present application are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0037] In one or more embodiments of the present application, a method for semantic association of urban buildings based on a large language model is proposed, which aims to achieve accurate semantic parsing and full-time and space association modeling of cross-departmental multi-source heterogeneous data by integrating geographic knowledge and multi-task learning. The core technical solution adopts a shared encoder (such as GeoRoBERTa) to generate a token-level representation with both semantic and geographic features by injecting address tag embedding and geographic coordinate encoding. In the address parsing branch, IDCNN and BiLSTM are used to extract multi-scale features, combined with CRF to predict place name label sequences, to achieve high-precision building entity recognition and full-time and space object construction of buildings. In the address matching branch, the address pair features output by GeoRoBERTa are pooled, and matching categories are predicted to ensure semantic alignment across data sources. Through multi-task joint training, address parsing and matching tasks are optimized simultaneously to improve model efficiency and generalization ability. This solution builds a unified semantic standard by deeply integrating geographic knowledge with the semantic analysis capabilities of a large language model, supports the dynamic representation of full-time and space objects of urban buildings and the precise construction of spatiotemporal knowledge graphs, and solves the consistency problem in cross-departmental data integration.

[0038] In one or more embodiments of the present application, an implementation method of a semantic association method for urban buildings based on a large language model is proposed, which realizes semantic parsing of cross-departmental multi-source heterogeneous data and full-time and space object association modeling through a shared encoder and a multi-task learning framework. The overall implementation plan can use GeoRoBERTa as the basic encoder, inject geographic knowledge (address tag embedding and geographic coordinate encoding), and generate a token-level representation with both semantic and geographic features. The system includes two branches: the address parsing branch uses IDCNN and BiLSTM to extract multi-scale features, and combines CRF to predict place name label sequences; the address matching branch pools the address pair features and predicts matching categories. The two branches are optimized through multi-task joint training to ensure the efficiency and generalization ability of the model.

[0039] Please refer to Figure 1 , which shows a process of an embodiment of creating a model framework of a building spatiotemporal object according to the present application. Figure 1 Shown is the model framework for creating building spatiotemporal objects.

[0040] The model framework includes an address resolution model and an address matching model.

[0041] In this embodiment, the address resolution model and the address matching model share an encoding network, the input of the encoding network is a text address, and the output of the encoding network is a text-geographic fusion vector; the address resolution model and the address matching model are trained using a multi-task joint training method.

[0042] In this embodiment, the address resolution model includes an encoding network and an address resolution network; the input of the address resolution model is address text, and the output of the address resolution model is an address label sequence; the address resolution model is used to convert address text in multi-source building data into an address label sequence.

[0043] In this embodiment, the address matching model includes an encoding network and an address matching network; the input of the address matching model is paired address text, and the output of the address matching model is the relationship category between the paired address text; the address matching model is used to identify address texts indicating the same building in multi-source building data.

[0044] In this embodiment, the building spatiotemporal object can integrate multi-dimensional information of the building's physical store to support the expression, association, and dynamic change analysis of the building.

[0045] For example, BuildingObj represents a building object.

[0046] BuildingObj

[0047] ={Id, Refer, Locations, Shapes, Features, Action, Variations}

[0048] Among them, Id refers to the ID value of the building object, which is the unique identifiable identifier of the building object. Refer is the spatiotemporal reference of the building object. Locations represents the spatial location of the building object. Shapes represents the spatial form of the building object. Features represents the attribute characteristics of the building object. Actions represents the behavioral capabilities of the building object. Variations represents the set of changes of the building object.

[0049] It is important to note that the multi-dimensional complexity of building data, including physical attribute information (such as building structure, area, and floor height), multi-format address expressions (both standardized and non-standardized), and multi-source business data (such as building maintenance records, elevator installation projects, and historical renovation and expansion information), is effectively enhanced by the multi-task learning framework through a joint training mechanism. This framework effectively enhances the generalization and robustness of the model, establishing a unified semantic understanding standard for cross-departmental building data. Furthermore, the integration of geographic knowledge encoding networks and multi-task learning significantly improves the accuracy, consistency, and efficiency of semantic parsing and spatiotemporal correlation modeling for cross-departmental, multi-source, and heterogeneous urban building data.

[0050] In some embodiments, the building spatiotemporal object can also be referred to as a full spatiotemporal object model. The building spatiotemporal object has a higher level of logical and cognitive abstraction and can construct a unified data framework for spatiotemporal entities. For multi-dimensional building entities, the model can integrate their attributes to support the expression, association, and dynamic change analysis of the building. The core of the object-oriented expression of a building is to treat it as an independent entity and describe and manage information in the form of a spatiotemporal object. This method converts attributes such as the building's history, current status, and development potential into computable data streams, which not only reconstructs the information system but also provides a theoretical basis for full spatiotemporal data modeling.

[0051] In some embodiments, a building spatiotemporal object may include: a building identification, which is a uniquely identifiable identification of a building object; a building spatiotemporal reference, which is a spatiotemporal reference of the building object; a building spatial location; a building spatial form; building attribute characteristics; the building's behavioral capabilities, the building's response to changes in the natural environment and to social planning measures; and a set of building changes.

[0052] When establishing a data model for building objects, it is necessary to use them as core entity categories and build a systematic framework based on the key elements of object description. The specific modeling method examples are as follows:

[0053] BuildingObj

[0054] ={Id, Refer, Locations, Shapes, Features, Action, Variations}

[0055] Among them, Id refers to the ID value of the building object, which is the unique identifiable identifier of the building object. Refer is the spatiotemporal reference of the building object. Locations represents the spatial location of the building object. Shapes represents the spatial form of the building object. Features represents the attribute characteristics of the building object. Actions represents the behavioral capabilities of the building object. Variations represents the set of changes of the building object.

[0056] Optionally, the spatiotemporal reference can determine the temporal and spatial position of all spatiotemporal objects in the real world. Taking the spatiotemporal reference as a benchmark, the reference can be divided into time and space, which can be expressed as the formula:

[0057] Refer = {TemporalRefer, SpatialRefer}, where TemporalRefer represents the temporal reference of the building object, and SpatialRefer represents the spatial reference of the building object. Commonly used temporal reference systems include International Atomic Time (TAI), Coordinated Universal Time (UTC), Greenwich Mean Solar Time (GMST), and BeiDou Time (BDT). The spatial reference of a building can be based on commonly used spatial reference systems in China, such as the China Geodetic Coordinate System (CGCS2000) and WGS84.

[0058] Optionally, the spatial location attribute determines the spatial range and centroid coordinates of the building object. The spatial location is expressed as follows:

[0059] Loactions={SpatialExtent, Centroid}

[0060] Among them, SpatialExtent represents the two-dimensional spatial extent of the building object, and Centroid represents the two-dimensional centroid coordinates of the building. SpatialExtent in the formula is the frame of the spatial extent, which is formally expressed as the formula:

[0061] SpatialExtent={[x1, y1], [x2, y2]}

[0062] Where [x1, y1], [x2, y2] are the coordinates of the upper left corner and lower right corner of the two-dimensional square space range of the building object.

[0063] Optionally, spatial form specifies the two-dimensional spatial boundaries, shape, and representation of a building object. The representation of a full-time and space-time building object varies at different scales, angles, and states. The spatial form representation is as follows:

[0064]

[0065] Among them, Form represents the different forms of building objects. For the display of different forms, it is necessary to specify the Type type, MinScale represents the minimum scale for displaying the form, MaxScale represents the maximum scale for displaying the form, and Geometry represents the display method of the form.

[0066] Optionally, attribute features are used to describe specific properties of an object, such as building area, actual use of the building, construction year, building structure, building height, number of floors above and below ground, etc. Attribute features can be expressed as the formula:

[0067]

[0068] Among them, Features represents the set of attribute features of the building object, Feature1, Feature2, and Feature3 represent an attribute feature in the attribute feature set, Name represents the name of a specific attribute feature, and Value represents the attribute value corresponding to the specific attribute feature.

[0069] Optionally, the building's behavioral capacity describes how the building responds to changes over time and social planning development, including the building's response to changes in the natural environment and to social planning measures. For example, the building's structure will change with human-economic or natural factors such as reconstruction, expansion, renovation, and natural disasters. Behavioral capacity can be expressed as the formula:

[0070] Actions

[0071] ={Type, Name, Parameters, Conditions, Triggers, Receptors, Models}

[0072] Among them, Type and Name represent the type and name of the behavior, Parameters, Conditions, Triggers, Receptors, and Models represent behavior capability parameters, environmental influencing factors, behavior triggering conditions, behavior action objects, and behavior calculation models, respectively.

[0073] Optionally, a change set describes the dynamic changes of building objects over time, such as urban planning can change the building area and form of the building structure. The change set can be expressed as a formula:

[0074]

[0075] Among them, Variations represents the set of all changes of building objects, Variation represents each specific change, Time is the time when a certain change occurs, Refer, Locations, Shapes, Features, and Action represent the characteristic values ​​corresponding to the spatiotemporal reference, spatial position, spatial form, attribute characteristics, and behavioral capabilities after the change, respectively.

[0076] Therefore, the full-time and space-time object model provides a unified framework for the accurate and multi-dimensional description and management of building entities in computers. It can express the spatial form of building objects, describe the attribute characteristics of building objects and record the dynamic changes of building objects. This provides a basic framework and subsequent data management guidance for cross-departmental multi-source heterogeneous data processing, address matching and semantic association of building entity objects.

[0077] In some embodiments, the encoding network includes a word segmenter, an address tag embedding module, a geographic coordinate encoding module, and a feature extraction layer. The word segmenter is configured to convert an input address text into an address text vector. The address tag embedding module is configured to convert address elements in the address text into an address tag vector. The geographic coordinate encoding module is configured to convert address elements in the address text into a geographic coordinate vector.

[0078] The input of the feature extraction layer is a text-geometry splicing vector generated based on the address text vector, the address label vector and the geographic coordinate vector, and the output of the feature extraction layer is a text-geometry fusion vector.

[0079] It's important to note that the encoding network achieves precise semantic modeling of multi-source, heterogeneous data by infusing the encoder with geographic knowledge and a multi-layered Transformer architecture. By integrating semantic and geographic features, this module establishes a deep connection between language and space, providing high-quality contextualized feature representations for subsequent address parsing and matching tasks. This significantly improves the accuracy and generalization of building semantic associations.

[0080] As an example, the encoding network can be pre-trained based on a large language model.

[0081] As an example, the encoding network may include a 12-layer Transformer encoder structure, each layer having 768-dimensional hidden states and 12 attention heads, and is optimized by Dynamic Masked Language Modeling (MLM) technology.

[0082] Dynamic masking thus applies a different masking pattern in each training batch, significantly enhancing the model's ability to capture the semantics and generalize to complex text. The core of the Transformer architecture lies in its self-attention mechanism, which allows the model to consider information at all positions when processing a sequence, thereby capturing long-range semantic dependencies in building description text.

[0083] Optionally, the Transformer architecture can be deeply modified to incorporate geographic knowledge, thereby enhancing the model's ability to model address and spatial information to meet the specific needs of tasks involving semantically related building data. Geographic knowledge infusion is a multimodal information fusion technique that significantly improves the model's performance on geo-related tasks by combining semantic and spatial information.

[0084] Geographic knowledge injection mainly includes two key parts: address tag embedding and geographic coordinate encoding using spatial continuity preservation technology.

[0085] In one embodiment, address embedding can be achieved by the following method: Design a 64-dimensional label embedding vector for the address components (such as province, city, district, street, house number). Let the address component set be C = {c1, c2, ..., c n}, each component c i Mapping to embedding vector

[0086] Through the learnable embedding matrix The introduction of address tag embedding not only enhances the semantic representation of tokens, but also injects prior knowledge of address structure into the model, improving the sensitivity to address information and parsing accuracy.

[0087] In one embodiment, the geographic coordinate encoding uses a spatial continuity preservation technology to convert longitude and latitude coordinates into multi-dimensional continuous vectors.

[0088] Geographic coordinate encoding uses spatial continuity preservation technology to convert longitude and latitude coordinates (lat, lon) into 128-dimensional continuous vectors and uses sine-cosine position encoding:

[0089]

[0090] In this expression, pos is the normalized coordinate position, d = 128 is the encoded latitude, and i is the dimension index. The core advantage of sine-cosine encoding is that it preserves spatial continuity, ensuring that similar geographic locations have similar representations in the embedding space. This is crucial for geographic clustering of buildings and reasoning about spatial relationships. Furthermore, this encoding method is scale-invariant, adapting to location representation requirements at varying geographic granularities.

[0091] In one embodiment, the feature extraction layer adopts a 12-layer Transformer encoder structure, each layer having 768-dimensional hidden states and 12 attention heads.

[0092] like Figure 2 As shown in the figure, the encoding network adopts the GeoRoBERTa model, which deeply integrates semantic and geographic features to provide high-quality token-level representation for subsequent address parsing and matching tasks.

[0093] In terms of basic model construction, this solution can use RoBERTa-base as the pre-training basis. The model contains a 12-layer Transformer encoder structure, each layer has a 768-dimensional hidden state and 12 attention heads. RoBERTa is optimized through dynamic masked language modeling (MLM) technology. Compared with the static masking strategy, dynamic masking applies different masking patterns in each training batch, which significantly enhances the model's semantic capture ability and generalization of complex texts. The core of the Transformer architecture lies in its self-attention mechanism, which allows the model to consider information at all positions when processing the sequence, thereby capturing long-range semantic dependencies in the building description text. The mathematical basis of self-attention calculation is expressed as:

[0094]

[0095] Where Q, K, and V represent query, key, and value matrices, respectively, which are generated from the same input data through linear transformation. k The dimension of the key vector (64) is used to scale the dot product operation to prevent the gradient vanishing problem. The uniqueness of this attention mechanism is that it can adaptively assign weights to each token, enabling the model to dynamically adjust its focus based on the context, which is crucial for complex address parsing and matching tasks.

[0096] To meet the specific task requirements of semantic associations of urban buildings, this solution deeply transforms RoBERTa by infusing geographic knowledge, thereby enhancing the model's ability to model addresses and spatial information. Geographic knowledge infusion is a multimodal information fusion technology that significantly improves the model's performance on geography-related tasks by combining semantic and spatial information. Specifically, geographic knowledge infusion consists of two key components:

[0097] Address label embedding: Design a 64-dimensional label embedding vector for address components (such as province, city, district, street, house number). Let the address component set be C = {c1, c2, ..., c n}, each component c i Mapping to embedding vector

[0098] Through the learnable embedding matrix The introduction of address tag embedding not only enhances the semantic representation of tokens, but also injects prior knowledge of address structure into the model, improving the sensitivity to address information and parsing accuracy.

[0099] Geographic coordinate encoding uses spatial continuity preservation technology to convert longitude and latitude coordinates (lat, lon) into 128-dimensional continuous vectors and uses sine-cosine position encoding:

[0100]

[0101] In this expression, pos is the normalized coordinate position, d = 128 is the encoded latitude, and i is the dimension index. The core advantage of sine-cosine encoding is that it preserves spatial continuity, ensuring that similar geographic locations have similar representations in the embedding space. This is crucial for geographic clustering of buildings and reasoning about spatial relationships. Furthermore, this encoding method is scale-invariant, adapting to location representation requirements at varying geographic granularities.

[0102] The model pre-training phase used a large-scale cross-departmental building-related text corpus, which included building attribute information (such as structural type and functional description), address information (such as province, city, district, street, and house number), building maintenance records, and elevator installation and renovation data. The pre-training tasks included:

[0103] Masked Language Model (MLM): randomly masks 15% of the tokens and predicts the original vocabulary. The loss function is cross entropy:

[0104]

[0105] Address component prediction: predict the address component label corresponding to the token, enhance geographic knowledge modeling, and use the same loss function as above. Use the Adam optimizer (initial learning rate 2×10 -5 , batchsize=32), the training cycle is 10 epochs, and the model parameters are optimized to adapt to the semantic scene of urban buildings.

[0106] The input text is converted into a token sequence through a tokenizer (based on Byte-Pair Encoding). Each token is mapped to a 768-dimensional word embedding vector, which is combined with address tag embedding and geographic coordinate encoding to form a comprehensive input representation. The embedding vector calculation formula is:

[0107] x i =e word (w i )+e tag (c i )+e geo (lati, loni)

[0108] where e word 、e tag 、e geoThey are word embedding, label embedding, and coordinate encoding. This multimodal fusion method integrates semantic information and geographic information in the vector space, providing a rich feature foundation for subsequent attention calculations.

[0109] The attention calculation process is implemented through a cascade of 12 layers of Transformer encoders, with each encoder layer calculating context-dependent token representations based on a self-attention mechanism. This multi-layered encoding architecture progressively extracts and integrates the semantic and geographic dependencies in building descriptions, building a hierarchical representation learning system from surface lexical features to deeper semantic associations. Within each Transformer layer, a multi-head attention mechanism decomposes the attention operation into multiple parallel attention heads, each focusing on a different feature subspace, further enhancing the model's expressive power. The output of each layer is a 768-dimensional vector, preserving rich contextual information. This vector serves as the input to the next layer, forming a deeply cascaded feature extraction network.

[0110] GeoRoBERTa finally outputs the contextualized representation of each token (dimension 768) as the input feature of the address parsing and matching tasks. The output vector sequence is:

[0111] H=h1,h2,...,h n ,

[0112] In summary, the encoding network, through the infusion of geographic knowledge from GeoRoBERTa and its multi-layered Transformer architecture, achieves precise semantic modeling of multi-source heterogeneous data. By integrating semantic and geographic features, this module establishes a deep connection between language and space, providing high-quality contextualized feature representations for subsequent address parsing and matching tasks. This significantly improves the accuracy and generalization of semantic associations between urban buildings.

[0113] In some embodiments, the address resolution network includes an expanded convolutional neural network, a long short-term memory network, and a conditional random field layer; the expanded convolutional neural network is used to extract local multi-scale features from the text-geography fusion vector output by the encoding network; the long short-term memory network is used to extract dependencies in the text-geography fusion vector; the input of the conditional random field layer is a first fusion vector, which is obtained by splicing the outputs of the expanded convolutional neural network and the long short-term memory network; the output of the conditional random field layer is an address tag sequence.

[0114] like Figure 3As shown in the figure, the address parsing branch is built on the semantic representation generated by the encoding network (such as GeoRoBERTa), and achieves high-precision semantic deconstruction and entity recognition of building address text through a multi-level feature extraction and sequence labeling architecture, namely the fusion of expanded dilated convolutional neural network (IDCNN), bidirectional long short-term memory network (BiLSTM) and conditional random field (CRF). Figure 3 As shown in the figure, this branch converts unstructured address text into structured semantic representation through a hierarchical feature extraction and reasoning process, providing basic support for cross-departmental building data integration.

[0115] For example, the Dilated Convolutional Neural Network (IDCNN) is a neural network designed specifically for sequence modeling. It uses dilated convolutions to expand the receptive field of the convolution kernel, efficiently capturing local contextual information without stacking multiple layers. Therefore, by introducing spacing (dilation rate) in the convolution kernel, dilated convolutions enable the model to cover longer sequence ranges, making it particularly suitable for parsing the hierarchical structure of address text (such as street and house numbers).

[0116] As an example, IDCNN is configured with 3 layers of dilated convolution, with dilation rates of 1, 2, and 4, a convolution kernel size of 3×3, and 256 output channels. The convolution operation is defined as:

[0117] h t =Conv(x t-d：t+d , W)+b

[0118] Among them, x t-d：t+d is the input feature window, W is the convolution kernel weight, b is the bias, and d is determined by the dilation rate. Increasing the dilation rate causes the receptive field of each layer to grow exponentially, covering short, medium, and long range context information respectively, generating a 256-dimensional local feature vector.

[0119] For example, the core advantages of bidirectional long short-term memory networks include their ability to effectively model long-distance dependencies in sequence data. By combining two LSTM units, one forward and one backward, BiLSTM processes sequence information bidirectionally, both forward and backward, making it particularly well-suited for capturing long-range semantic associations that may exist in address text (such as the subordinate relationship between building names and administrative divisions, and the distinction between primary address components and auxiliary information). The uniqueness of LSTM units lies in their carefully designed gating mechanisms—including a forget gate, an input gate, and an output gate. These gating structures intelligently control the flow of information, determining which historical information to retain, which new information to integrate, and which internal states to output, effectively alleviating the vanishing gradient problem in traditional recurrent neural networks.

[0120] As an example, a BiLSTM can be configured with 2 layers and a hidden layer dimension of 256. Its output is calculated as:

[0121]

[0122] in, represents the hidden state of the forward LSTM at position t, which depends on the current input x t and the hidden state at the previous position Indicates that the hidden state of the reverse LSTM at position t depends on the current input x t and the hidden state of the next position By concatenating the hidden states in both the forward and reverse directions, BiLSTM generates a 512-dimensional feature representation for each token (a combination of 256-dimensional forward LSTM features and 256-dimensional reverse LSTM features). This bidirectional processing mechanism of BiLSTM significantly enhances the model's ability to model complex contextual dependencies in address sequences, enabling the system to understand long-range association patterns and semantic integrity in address text.

[0123] The IDCNN and BiLSTM network structures each excel at capturing features of different scales and properties: IDCNN efficiently extracts local multi-scale features, while BiLSTM excels at modeling long-term dependencies in sequences. To leverage the complementary strengths of these two structures, this system integrates their outputs through a feature fusion mechanism. Specifically, for each position t in the sequence, the output features of the IDCNN and BiLSTM are fused into a single, combined feature vector through a concatenation operation:

[0124] f t =[h IDCNN,t ;h BiLSTM,t ]

[0125] The concatenated feature dimension is 768 (256-dimensional IDCNN + 512-dimensional BiLSTM). To reduce computational complexity, the dimension is reduced to 512 through a linear layer:

[0126] z t =W f f t +b f ,

[0127] in, The features after dimensionality reduction retain key semantic information and are suitable for subsequent sequence labeling.

[0128] Here, the Conditional Random Field (CRF) is a probabilistic graphical model used for sequence labeling tasks. It ensures global consistency in sequence prediction by modeling the transition probability between labels. CRF optimizes the label sequence by maximizing the log-likelihood probability of the sequence. It is particularly suitable for tasks that require label dependency in address parsing (for example, in the Chinese address format, "province" should usually be followed by "city" rather than "district", and "street" usually comes after "district", etc.). The probability distribution of CRF is defined as:

[0129]

[0130] Among them, s t (y t ) is the tag score of the tth token, Ty t -1,y t Score label transfer. The BIOES (Begin-Inside-Outside-End-Single) labeling scheme is used. The label set includes Province (province), City (city), District (county), Street (street / township), Community (community / residential area), Building (building name), Number (house number), POI (point of interest), and Other (other information). CRF decodes the data using a dynamic programming algorithm (such as Viterbi) to output the optimal label sequence.

[0131] The training process uses the negative log-likelihood loss function:

[0132] L=-log P(y|x)

[0133] The Adam optimizer is used, where y is the true label sequence, x is the input feature sequence, and P(y|x) is the conditional probability calculated by CRF. The initial learning rate is 1×10 -4 , the weight decay is 1×10-5 , with a batch size of 16. During training, the accuracy on the validation set was monitored. When accuracy stopped improving, a learning rate decay strategy (reduced to 1 / 10 of the original learning rate) was implemented to optimize model convergence. Joint training ensured the coordinated optimization of the IDCNN, BiLSTM, and CRF modules, improving the recognition accuracy of address entities (such as building names and house numbers) in cross-departmental data.

[0134] Throughout the training process, the three core modules, IDCNN, BiLSTM, and CRF, are jointly optimized in an end-to-end manner, ensuring the coordinated development of feature extraction and sequence labeling. This holistic optimization strategy enables the system to automatically adjust the parameters of each module, improving the recognition accuracy of address entities in heterogeneous cross-departmental data, especially the accurate extraction of key entities such as building names and house numbers.

[0135] Through this multi-layered architecture, the address resolution network integrates the multi-scale local feature extraction capabilities of IDCNN, the long-range dependency modeling capabilities of BiLSTM, and the sequential consistency guarantee mechanism of CRF to achieve accurate semantic parsing of building address text. The system not only handles standardized address formats but also demonstrates robustness to unstructured addresses, those with missing components, or those with diverse expressions. This provides key technical support for the semantic alignment and integration of heterogeneous, multi-source data across departments.

[0136] In some embodiments, the input of the address matching network is the paired text-geography fusion vector output by the encoding network; the address matching network includes a pooling layer, a feature interaction layer and a fully connected network; the pooling layer is used to pool the input paired text-geography fusion vector to obtain a first pooling result vector and a second pooling result vector; the feature interaction layer is used to fuse the first pooling result vector and the second pooling result vector to obtain a second fusion vector; the fully connected network is used to map the second fusion vector to the address category matching space.

[0137] In some embodiments, the pooling layer performs average pooling and maximum pooling, and fuses the results of average pooling and maximum pooling; the feature interaction layer performs: calculating the difference vector between the first pooling result vector and the second pooling result vector, calculating the dot product vector between the first pooling result vector and the second pooling result vector; splicing the first pooling result vector, the second pooling result vector, the difference vector and the dot product vector to obtain a second fused vector.

[0138] like Figure 4As shown in Figure 2, the address matching network can be built on the semantic representation generated by the encoding network (such as GeoRoBERTa) to process the similarity evaluation of address pairs and determine whether two address records point to the same building entity in the real world. This function is crucial for building identity authentication and semantic alignment in cross-departmental data integration. Figure 4 As shown in the figure, the matching mechanism achieves high-precision address entity matching through a series of steps including deep feature extraction, multi-strategy feature pooling, multi-dimensional feature interaction, and hierarchical classification decision-making.

[0139] The address matching network is based on the output features of the encoding network (i.e., the text-geo fusion vector). As an example, for two input address texts A and B, the encoding network generates token-level context representations.

[0140] Optionally, the encoding network uses an improved Transformer architecture that not only captures the linguistic semantic information in the address text, but also extracts geospatial relevance through a special geographic knowledge enhancement mechanism, thereby generating an information-rich 768-dimensional feature vector sequence for each token:

[0141] H A =h A，1 , h A，2 ,...,h A，m , H B =h B，1 , h B，2 ,...,h B，n ,

[0142] Here, m and n are the number of tokens of addresses A and B, respectively. This step utilizes the geographic knowledge injection of the encoding network (such as address label embedding and coordinate encoding) to ensure that the features contain semantic information of building attributes, addresses, and business information.

[0143] Optionally, a complementary feature pooling strategy can be adopted, combining average pooling and maximum pooling operations to effectively aggregate token-level local features into a global representation that can represent the entire address entity.

[0144] Here, average pooling effectively captures the overall semantic distribution and general characteristics of the address by calculating the mean of all token features.

[0145] Here, max pooling extracts the most significant features in each dimension to highlight the key identifying elements in the address. The mathematical definitions of these two pooling operations are:

[0146]

[0147] Where N is the number of tokens, h i is the token feature. Average pooling and maximum pooling are calculated for addresses A and B respectively, resulting in two 768-dimensional vectors. In order to preserve the complementary information captured by different pooling strategies, the results of average pooling and maximum pooling are further concatenated to generate a richer 1536-dimensional address representation vector:

[0148] v A =[h A，avg ;h A，max ],v B =[h B，avg ;h B，max ],

[0149] This concatenation operation not only preserves the global semantics and local salient features of the address, but also enhances the robustness of the representation. It can adapt to address descriptions in different formats and expressions and has strong adaptability to non-standardized address input.

[0150] To capture the semantic relationship between two addresses, the address representation vector v is calculated A and v B These interaction features mainly include difference vectors and dot product vectors.

[0151] The difference vector reflects the degree of difference between two addresses in each dimension through element-level subtraction operation.

[0152] The dot product vector quantifies the similarity strength of two addresses in each dimension through element-wise multiplication operations:

[0153] v diff =v A -v B , v dot =v A ⊙v B

[0154] Among them, ⊙ represents element-by-element multiplication. These two interactive features depict the semantic relationship between address pairs from different perspectives.

[0155] Furthermore, the original address representation vector (i.e., the pooling result vector processed by the pooling layer), the difference vector, and the dot product vector can be concatenated to form the final high-dimensional feature representation:

[0156] v final =[v A ;v B ;v diff ;v dot ],

[0157] This 6144-dimensional fused feature vector (composed of four 1536-dimensional vectors) comprehensively integrates the original semantics, relative differences, and interactive similarity information of the address pairs, providing an information-rich and multi-dimensional input representation for subsequent refined classification.

[0158] Optionally, the matching and classification phase maps the high-dimensional feature space to the address matching category space through a carefully designed fully connected network. The fully connected network gradually extracts complex relationships and abstract patterns between features through multiple layers of nonlinear transformations.

[0159] Optionally, the fully connected network can adopt a three-layer hidden layer network architecture, with the dimensions of each layer set to 1024, 512, and 256 respectively, to achieve a progressive mapping from the high-dimensional feature space to the low-dimensional decision space. The ReLU activation function is used between each layer to introduce nonlinear transformations and enhance the network's expressive power:

[0160] h l+1 =ReLU(W l h l +b l ), ReLU(x)=max(0,x)

[0161] Among them, W l 、b l are the weights and biases of the lth layer. The last layer of the network outputs the matching probability through the Softmax function:

[0162] P(y)=Soffmax(W out h last +b out ),

[0163] Match categories include complete match (addresses point to the same building), partial match (addresses partially overlap, such as the same street but different house numbers), and mismatch (addresses are unrelated). Softmax ensures that the sum of probabilities is 1, enhancing classification reliability.

[0164] The training uses the cross entropy loss function to measure the difference between the predicted probability and the true label:

[0165]

[0166] Among them, y i is the true label, is the predicted probability. Using a gradient descent algorithm with an adaptive learning rate (such as the Adam optimizer), the parameter update step size can be automatically adjusted. The initial learning rate is 5×10 -5 , the batch size is 24. In order to promote the model to converge to the optimal solution more stably, the learning rate will decay to 0.9 times the original value after every 5 full epochs of training to promote convergence:

[0167] lr t+5 =0.9*lr t

[0168] This learning rate decay strategy maintains a large parameter update step size in the early stages of training to quickly approach the optimal solution, while using a smaller step size for fine-tuning in the later stages of training. This effectively avoids overfitting and improves the model's generalization capabilities. The training process optimizes the parameters of the fully connected network to ensure accurate matching of building addresses, attributes, and business information in cross-departmental data.

[0169] It should be noted that the Address Matching Network, through its architectural design of hierarchical feature extraction, multi-strategy pooling, multi-dimensional interaction, and deep classification, achieves precise semantic matching of address pairs, providing reliable technical support for semantic alignment of cross-departmental data and modeling of building spatial and temporal associations. This Address Matching Network not only handles standardized address formats but also exhibits strong adaptability to unstructured, partially missing, or diversely expressed address descriptions, effectively addressing the heterogeneous nature of address data in real-world applications.

[0170] In some embodiments, the loss function of the address resolution model is a conditional random field loss function L parse The loss function of the address matching model adopts the multi-class cross entropy loss function L match The joint training process of the address resolution model and the address matching model includes: constructing a joint loss function; wherein the joint loss function is

[0171] L=αL parse +βL match +λ\|θ\| 2

[0172] Among them, α and β are task weights, which are initially set to 0.5, and λ is the regularization coefficient, which is set to 1×10 -5 , θ is the model parameter.

[0173] By collaboratively optimizing address resolution and address matching, two closely related tasks with their own characteristics, and conducting joint training, the model's semantic parsing and association capabilities for cross-departmental building data can be significantly improved. The feature representation sharing mechanism and task-specific branch architecture, combined with adaptive dynamic weight adjustment technology, achieve efficient knowledge transfer and fusion learning in a heterogeneous data environment. In response to the multi-dimensional complexity of building data - including physical attribute information (such as building structure, area, and floor height), multi-format address expressions (standardized and non-standardized coexisting), and multi-source business data (such as house maintenance records, elevator installation projects, and historical renovation and expansion information) - the multi-task learning framework effectively enhances the generalization and robustness of the model through a joint training mechanism, and establishes a unified semantic understanding standard for cross-departmental building data.

[0174] Defining specific losses for each task and weighting them together can coordinate the optimization process of address parsing and matching tasks. The address parsing task uses the conditional random field (CRF) loss to measure the log-likelihood of the predicted probability of the sequence label and the true label:

[0175] L parse = -logP(y|x)

[0176] Where y is the true label sequence and x is the input sequence. P(y|x) is calculated using the CRF probability distribution and represents the probability of label sequence y given the label sequence given the input sequence x. The CRF loss function fully considers the contextual dependencies and transition constraints between address components, capturing the implicit grammatical rules in the city hierarchy and address format specifications, and improving the global consistency of sequence labeling.

[0177] The address matching task is constructed as a three-class classification problem, and a multi-class cross entropy loss function is used to evaluate the information difference between the classification prediction and the true class label:

[0178]

[0179] Among them, y i is the one-hot encoding of the true category, p i To predict the probability, the matching categories are finely divided into three semantic relationships: complete match (two addresses point to exactly the same building entity), partial match (there is a containment or overlap relationship between the addresses, such as the same block but pointing to different building units), and mismatch (addresses point to completely different building entities).

[0180] The joint loss function balances task contributions and prevents overfitting by weighted summation and introducing L2 regularization:

[0181] L=αL parse +βL match+λ\|θ\| 2

[0182] Among them, α and β are task weights, which are initially set to 0.5, and λ is the regularization coefficient, which is set to 1×10 -5 , θ is a model parameter. The joint loss ensures the coordinated optimization of the two tasks and improves the processing ability of building attributes, addresses, and business information.

[0183] In some embodiments, dynamic weight adjustment optimizes α and β by monitoring task performance to solve the problem of task convergence imbalance.

[0184] In some embodiments, during the joint training process, a performance change rate may be determined; based on the performance change rate, the weight of the first task and the weight of the second task may be adjusted;

[0185] Among them, the weight update formula is:

[0186] α t+1 =α t ·(1+γ(Δperf parse -Δperf match ))

[0187] β t+1 =1-α t+1

[0188] Here, γ is the adjustment rate, which is set to 0.1.

[0189] Optionally, based on the relative change in task performance, the weights can be dynamically adjusted to prioritize tasks with lagging performance and enhance the overall model generalization ability. The weight adjustment is based on the performance change rate, defined as:

[0190]

[0191] Among them, perf task，t is the performance indicator of the tth round (the parsing task uses accuracy, and the matching task uses F1 score).

[0192] When the parsing task performance stagnates (Δperf parse <Δperf match ), increasing α strengthens the parsing task, while decreasing it strengthens the matching task. This adaptive adjustment mechanism based on performance differences not only balances the learning progress between tasks but also dynamically adjusts the optimization strategy based on data characteristics and model status, significantly improving the model's adaptability in complex heterogeneous data environments.

[0193] Optionally, a multi-task learning framework can set up a parameter sharing mechanism to strategically reuse some parameters in the neural network hierarchy, thereby reducing computing resource overhead while enhancing implicit knowledge transfer between tasks.

[0194] Parameter sharing in the embodiments of this application can include the basic feature extraction layer, i.e., the encoding network. For example, the basic structure can use the GeoRoBERTa encoder. GeoRoBERTa is a geographically enhanced pre-trained language model based on the Transformer architecture. By sharing all parameters of its 12-layer encoder, it provides a unified feature representation that integrates semantic and geographic information for address resolution and matching tasks:

[0195] H = GeoRoBERTa(x),

[0196] Here, x represents the input sequence, and H represents the token-level features. Task-specific branches (1DCNN-BiLSTM-CRF for parsing and fully connected networks for matching) use independent parameters to ensure the extraction of task-specific features. This strategy facilitates cross-task learning by sharing the encoder while maintaining branch flexibility.

[0197] Training uses an alternating training strategy to enhance collaborative optimization between tasks. In each mini-batch, a task is randomly selected for forward propagation and backpropagation, and the corresponding loss is calculated and the parameters are updated:

[0198]

[0199] Among them, η is the learning rate, L task The loss for the selected task. A warm-up period of 2 epochs was set initially to optimize only the GeoRoBERTa encoder parameters and stabilize shared feature learning. The total number of training rounds was 20 epochs, and the best model was selected based on the comprehensive performance (parsing accuracy and matching F1 score) on the validation set. Training was performed using the Adam optimizer with an initial learning rate of 5×10 -5 , the batch size is 16 to ensure adaptability to multi-source heterogeneous data.

[0200] It is important to note that the multi-task learning framework successfully achieves efficient collaborative optimization of address resolution and address matching tasks through one or more of the following: a joint loss function, strategic parameter sharing, an alternating training mechanism, and an innovative dynamic weight adjustment technique. This provides powerful technical support for the spatiotemporal semantic association of cross-departmental building data. This framework not only improves the model's ability to handle multi-source heterogeneous data, but also enhances the system's robustness to varying address representations and building attribute variations, laying a solid foundation for data integration in smart city construction.

[0201] Please refer to Figure 5 , Figure 5 A method for creating a building spatiotemporal object based on the model framework provided in any embodiment of the present application is provided. Figure 5 As shown, the exemplary process of creating a building spatiotemporal object based on the model is as follows:

[0202] Acquire heterogeneous data from different departments and text data containing addresses.

[0203] For text data, structured address information is extracted through the address parsing model to generate standard address data.

[0204] As an example, the text data may be long text data, such as text data with a character count greater than a preset threshold value. The threshold value may be set according to actual application scenarios, such as 200.

[0205] Input standard address data into the address matching model along with heterogeneous data from different departments.

[0206] The address matching model matches the received standard address data with heterogeneous data from different departments.

[0207] If the match is successful, the matching address pairs are extracted, and based on the matching address pairs and the heterogeneous data corresponding to the address pairs, the building spatiotemporal objects are constructed, the spatiotemporal attributes are further integrated, and the generated spatiotemporal objects are stored.

[0208] If the match fails, the unmatched data will be recorded and fed back for manual verification.

[0209] In this embodiment, the address matching model and / or the address resolution model is determined based on an embodiment of any model framework in this application.

[0210] Through this method, the effective integration and spatiotemporal objectification of building-related information from different sources is achieved.

[0211] Reference below Figure 6 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The terminal device in the embodiments of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., as well as fixed terminals such as digital TVs and desktop computers, etc. Figure 6 The electronic device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0212] like Figure 6As shown, the electronic device may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. Various programs and data required for the operation of the electronic device 600 are also stored in the RAM 603. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0213] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead.

[0214] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication device 609, or installed from the storage device 608, or installed from the ROM 602. When the computer program is executed by the processing device 601, the above-mentioned functions defined in the method of the embodiment of the present application are performed.

[0215] It should be noted that the computer-readable medium mentioned above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this application, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, device, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. This propagated data signal can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0216] In some embodiments, the client and server can communicate using any currently known or future developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), internetworks (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future developed networks.

[0217] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0218] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including, but not limited to, object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0219] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0220] The units described in the embodiments of this application may be implemented in software or hardware. In some cases, the name of a unit does not limit the unit itself. For example, a selection unit may also be described as a "unit for selecting pixels of the first type."

[0221] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0222] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0223] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned disclosed concepts. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

[0224] In addition, although adopting specific order to describe each operation, this should not be interpreted as requiring these operations to be executed in the specific order shown or in sequential order.Under certain environment, multitasking and parallel processing may be advantageous.Similarly, although comprising some specific implementation details in the above discussion, these should not be interpreted as limiting the scope of the application.Some features described in the context of separate embodiment can also be implemented in a single embodiment in combination.On the contrary, the various features described in the context of a single embodiment also can be implemented in multiple embodiments individually or in the mode of any suitable subcombination.

[0225] Although the subject matter has been described in language specific to structural features and / or methodological logical acts, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.

Claims

1. A method for creating a building spatiotemporal object, characterized in that: include: Acquire heterogeneous data from different departments and text data containing addresses; For text data, the address parsing model is used to extract structured address information and generate standard address data. Input standard address data into the address matching model along with heterogeneous data from different departments; The address matching model matches the received standard address data with heterogeneous data from different departments; If the match is successful, the matching address pair is extracted, and based on the matching address pair and the heterogeneous data corresponding to the address pair, the building spatiotemporal object is constructed, the spatiotemporal attributes are further integrated, and the generated spatiotemporal object is stored; If the match fails, the unmatched data will be recorded and fed back for manual verification; The address matching model and / or the address resolution model are determined based on the following model framework: The model framework includes: an address resolution model and an address matching model, wherein the address resolution model and the address matching model share an encoding network, the input of the encoding network is a text address, and the output of the encoding network is a text-geographic fusion vector; the address resolution model and the address matching model are trained using a multi-task joint training method; The address resolution model includes an encoding network and an address resolution network; the input of the address resolution model is address text, and the output of the address resolution model is an address tag sequence; the address resolution model is used to convert the address text in the building multi-source data into an address tag sequence; The address matching model includes an encoding network and an address matching network; the input of the address matching model is a pair of address texts, and the output of the address matching model is a relationship category between the pair of address texts; the address matching model is used to identify address texts indicating the same building in multi-source building data; a building spatiotemporal object of a target building, comprising building information and an address tag sequence extracted from building multi-source data indicating the target building; The loss function of the address resolution model is the conditional random field loss function L parse , The loss function of the address matching model adopts the multi-class cross entropy loss function L match ; The joint training process of the address resolution model and the address matching model includes: constructing a joint loss function; wherein the joint loss function is L=αL parse +βL match +λ||θ|| 2 Among them, α and β are task weights, which are initially set to 0.5, and λ is the regularization coefficient, which is set to 1×10 -5 , θ is the model parameter; The joint training process of the address resolution model and the address matching model includes: Determine the rate of change in performance during joint training; Adjusting the weight of the first task and the weight of the second task based on the performance change rate; Among them, the performance change rate is defined as: Among them, perf task，t is the performance indicator of round t, where the accuracy is used for parsing tasks and the F1 score is used for matching tasks; Among them, the weight update formula is: a t+1 =a t ·(1+γ(Δperf parse -Δperf match )) β t+1 ＝1-α t+1 Here, γ is the adjustment rate, which is set to 0.

1.

2. The method according to claim 1, characterized in that The encoding network includes a word segmenter, an address tag embedding module, a geographic coordinate encoding module and a feature extraction layer; The word segmenter is used to convert the input address text into an address text vector; The address label embedding module is used to convert the address elements in the address text into address label vectors; The geographic coordinate encoding module is used to convert the address elements in the address text into geographic coordinate vectors; The input of the feature extraction layer is a text-geometry splicing vector generated based on the address text vector, the address label vector and the geographic coordinate vector, and the output of the feature extraction layer is a text-geometry fusion vector.

3. The method according to claim 2, characterized in that The geographic coordinate encoding adopts a spatial continuity preservation technology to convert the latitude and longitude coordinates into a multi-dimensional continuous vector; The feature extraction layer adopts a 12-layer Transformer encoder structure, each layer has a 768-dimensional hidden state and 12 attention heads.

4. The method according to claim 1, wherein Building spatiotemporal objects include: Building identification, which is the unique identifiable identification of a building object; The building spatiotemporal reference is the spatiotemporal reference of the building object; spatial location of buildings; Building spatial form; Building property characteristics; the behavioral capabilities of buildings, their response to changes in the natural environment and to societal planning measures; A varying collection of buildings.

5. The method according to claim 1, wherein The address resolution network includes a dilated convolutional neural network, a bidirectional long short-term memory network and a conditional random field layer; The dilated convolutional neural network is used to extract local multi-scale features from the text-geography fusion vector output by the encoding network; The bidirectional long short-term memory network is used to extract the dependency relationship in the text-geography fusion vector; The input of the conditional random field layer is a first fused vector, which is obtained by concatenating the outputs of the dilated convolutional neural network and the bidirectional long short-term memory network; The output of the conditional random field layer is a sequence of address labels.

6. The method according to claim 1, characterized in that The input of the address matching network is the paired text-geography fusion vector output by the encoding network; The address matching network includes a pooling layer, a feature interaction layer and a fully connected network; The pooling layer is used to pool the input paired text-geography fusion vectors to obtain a first pooling result vector and a second pooling result vector; The feature interaction layer is used to fuse the first pooling result vector and the second pooling result vector to obtain a second fused vector; The fully connected network is used to map the second fusion vector to the address category matching space.

7. The method according to claim 6, characterized in that The pooling layer performs average pooling and maximum pooling, and fuses the results of average pooling and maximum pooling; The feature interaction layer performs: calculating the difference vector between the first pooling result vector and the second pooling result vector, calculating the dot product vector between the first pooling result vector and the second pooling result vector; splicing the first pooling result vector, the second pooling result vector, the difference vector and the dot product vector to obtain a second fusion vector.

Images