Multi-source dynamic collaborative geographic information mutual resolution method and device, equipment and medium

Abstract

The application relates to the field of geographic information mutual analysis, and discloses a multi-source dynamic cooperative geographic information mutual analysis method, device, equipment and medium. The method collects time-synchronized static geographic information, dynamic GPS positioning data and multiple pairs of unstructured text data sets of starting points and ending points, and constructs a dynamic space-time data pool; after double abnormality checking and Kalman filtering cleaning and correction, high-quality space-time data is obtained; a latitude and longitude mapping table of an address is optimized in combination with a three-level administrative range and a dynamic high-frequency word; a non-standard address is analyzed into a structured text through a BERT-CRF model, and accurate coordinates and standardized text are obtained through a hierarchical dynamic matching algorithm. The method realizes meter-level positioning, greatly improves the non-standard address analysis accuracy, reduces the mis-matching rate of addresses with the same name, adapts to dynamic scenes, takes into account the analysis efficiency and reliability, and provides efficient technical support for bidirectional accurate mutual analysis of geographic coordinates and text addresses.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geographic information mutual analysis, in particular to a multi-source dynamic collaborative geographic information mutual analysis method, device, equipment and medium. BACKGROUND

[0002] As a comprehensive technical platform integrating spatio-temporal data collection, location mapping and semantic analysis, the geographic information intelligent system has a core value in the standardized expression of physical space. In the process of spatial data processing, the accurate mapping of location coordinates is particularly critical, and the system needs to perform path planning, regional analysis, location services and other core functions according to the geographic coding results. Therefore, in the field of spatial information science, the mutual analysis of geographic coordinates and text addresses is a basic technical link.

[0003] In the prior art, a static matching method of a single data source (such as a GPS coordinate point or an OSM road network) is mainly used for geographic coding processing. Reverse geographic coding (coordinate conversion address) is realized by OSM road network matching (projecting GPS coordinates to the nearest road) or latitude and longitude grid attribution method (pre-storing grid center address), typical such as Gaode Map API; Forward geographic coding (address conversion coordinates) is realized by regular expression parsing address components (such as province-city-district level rules, querying geographic database) or retrieving crowdsourcing address library (such as OpenAddresses) to directly match text for coordinate positioning, and Google Geocoding service is a main application representative.

[0004] However, in the actual geographic coordinate and text address mutual analysis operation, researchers at home and abroad also face many difficulties. First, the singleness of the data source leads to OSM road network update delay, newly added communities and POIs cannot be analyzed, and the missed detection rate is increased. In complex scenes such as overpasses and tunnels, the matching positioning error of GPS drift points is greater than 200m, causing serious dynamic imbalance. Second, the component analysis based on text matching is prone to semantic ambiguity, and the mis-matching rate of the address with the same name is as high as 22.7%, such as the address with the same name “Zhongshan Road” (there are more than 218 roads with the same name in the country). Third, non-standardized addresses bring great errors to the analysis operation, such as village-level addresses (such as “Wangjiatun East Head Three Forks”) lack standardized expression, and the failure rate of analysis is as high as 38.5%. SUMMARY

[0005] Therefore, it is necessary to provide a multi-source dynamic collaborative geographic information mutual analysis method, device, equipment and medium which can significantly reduce the positioning error in complex scenes and improve the analysis accuracy of non-standard addresses.

[0006] A multi-source dynamic collaborative geographic information mutual analysis method, the method comprising:

[0007] acquire static geographic information synchronized in time, dynamic geographic information, and a non-structured text data set containing unstructured text data for describing pairs of start and end geographic locations, wherein the dynamic geographic information includes positioning data corresponding to each pair of start and end locations in the non-structured text data set;

[0008] based on the static, dynamic geographic information, and unstructured geographic text data, construct a dynamic spatio-temporal data pool, and determine abnormal data in the non-structured text data set;

[0009] based on a Kalman filter fusion algorithm, taking positioning data as a reference observation value and static geographic information as an auxiliary observation value, dynamically adjusting the weight of each data through an iteration mechanism to eliminate random errors of a single data, while performing semantic completion and error correction on the abnormal data, to obtain high-quality spatio-temporal data;

[0010] according to a three-level administrative range binding relationship, combined with dynamic high-frequency words extracted from the non-structured text data set corrected from the high-quality spatio-temporal data, perform triple-level weight distribution to generate an address latitude and longitude mapping table containing confidence, and optimize to obtain a standard address latitude and longitude mapping table;

[0011] construct a hybrid model, the hybrid model including a BERT layer and a CRF layer, wherein the BERT layer embeds a village-level geographic name knowledge graph;

[0012] acquire non-standard address text, use the hybrid model to obtain structured annotated text from the non-standard address text, based on the structured annotated text, through three-level address resolution confidence scoring and dynamic switching matching algorithm, combined with the standard address latitude and longitude mapping table and the precise association features of high-quality spatio-temporal data, obtain the precise association results of non-standard address and geographic coordinates and standardized address text.

[0013] in one embodiment, the static geographic information includes OSM administrative division data, OSM road network data, POI data, and AOI data;

[0014] the dynamic geographic information is GPS positioning data corresponding to pairs of start and end points of a trip automatically collected by a network car operation system;

[0015] the static, dynamic geographic information, and unstructured geographic text data are constructed into a unified spatio-temporal framework with time stamp and latitude and longitude as indexes to obtain the dynamic spatio-temporal data pool;

[0016] according to the positioning data and static geographic information, perform double abnormality checking on the non-structured text data set to determine the abnormal data that exist text information deviation from the standard administrative region.

[0017] In one embodiment, the double anomaly check includes:

[0018] Matching the GPS positioning data with the OSM administrative division vector data boundary coordinates to determine the corresponding standard administrative region information;

[0019] Calculating the string similarity between the unstructured text and the standard administrative region information by the edit distance algorithm, and verifying whether the text mentioned place falls within the OSM administrative region corresponding to the GPS positioning data;

[0020] When the similarity is lower than the preset threshold or the place exceeds the administrative region range, it is determined as abnormal data.

[0021] In one embodiment, based on the Kalman filter fusion algorithm, taking the GPS positioning data as the core reference observation value and the static geographic information as the auxiliary observation value, the data weight is dynamically adjusted through the iteration mechanism to eliminate the random error of single data, including: based on the longitude and latitude reference value of the GPS positioning data, predicting the theoretical error range of the geographic information, and calculating the theoretical error covariance, and then combining the actual deviation of each auxiliary observation value to dynamically adjust the weight.

[0022] In one embodiment, the address longitude and latitude mapping table is optimized to obtain a standard address longitude and latitude mapping table, including:

[0023] The structured address and unstructured address in the address longitude and latitude mapping table are processed by a double regularization engine, wherein the structured address containing road name, house number, and intersection is parsed by using the road network knowledge graph, and the unstructured address with direction description is recognized by using the colloquial parsing engine;

[0024] The semantic similarity of the two types of parsing results is calculated by the cosine similarity algorithm, and when the similarity meets the preset standard, it is determined as a homologous address and a preliminary optimized address longitude and latitude mapping table is obtained by fusion;

[0025] Based on the preliminary optimized address longitude and latitude mapping table, the standard address longitude and latitude mapping table is obtained according to the hybrid matching mode formed by the OSM administrative division data and dynamic high-frequency words.

[0026] In one embodiment, the structured annotated text is obtained by using the hybrid model, including:

[0027] The village-level place name knowledge graph embedded in the BERT layer contains village name, village landmarks, and adjacent village relationships, and the non-standard address text is converted into a semantic vector with multiple layers of semantic information by using a semantic vector generation tool;

[0028] After receiving the semantic vector, the CRF layer performs structured annotation on the compound address structure and outputs a structured annotated text with clear component division.

[0029] In one embodiment, based on the structured annotated text, the precise association result of non-standard address and geographic coordinates and the standardized address text are obtained by three-level address resolution confidence score and dynamic switching matching algorithm, combined with the standard address latitude and longitude mapping table and the precise association characteristics of high-quality space-time data, including:

[0030] A three-level address resolution confidence score model is used to calculate the resolution confidence scores of city, district and village levels for the village name, landmark and direction information in the structured annotated text.

[0031] According to the city-level confidence score result, the city-level trunk road positioning algorithm, the district-level regional POI clustering algorithm or the village-level micro-direction positioning algorithm is used according to the switching mechanism, combined with the pre-set matching algorithm, to match the standard address latitude and longitude mapping table and the structured annotated text, and simultaneously refer to the precise association characteristics of the high-quality space-time data to correct the matching result, and obtain the geographic coordinate data corresponding to the structured annotated text.

[0032] According to the standardization address expression rule, the precise association result of non-standard address and geographic coordinates and the standardized address text in standard format are generated based on the geographic coordinate data.

[0033] The application also provides a multi-source dynamic collaborative geographic information mutual resolution device, which comprises:

[0034] A multi-source geographic information acquisition module is used to acquire time-synchronized static geographic information, dynamic geographic information and a non-structured text data set containing descriptions of multiple pairs of starting and ending geographic positions, wherein the dynamic geographic information includes positioning data corresponding to each pair of starting and ending positions in the non-structured text data set.

[0035] A dynamic space-time data pool construction and abnormal data determination module is used to construct a dynamic space-time data pool based on the static, dynamic geographic information and non-structured geographic text data, and determine abnormal data in the non-structured text data set.

[0036] A high-quality space-time data obtaining module is used to obtain high-quality space-time data by iteratively adjusting the weights of each data to eliminate random errors of a single data based on a Kalman filter fusion algorithm, taking the positioning data as the reference observation value and the static geographic information as the auxiliary observation value, and simultaneously performing semantic completion and error correction on the abnormal data.

[0037] The standard address longitude and latitude mapping table construction module is configured to perform three-level weight distribution in combination with dynamic high-frequency words extracted from the non-structured text data set corrected from the high-quality space-time data according to a three-level administrative range binding relationship, generate an address longitude and latitude mapping table containing a confidence degree, and optimize the address longitude and latitude mapping table to obtain a standard address longitude and latitude mapping table.

[0038] The hybrid model construction module is configured to construct a hybrid model including a BERT layer and a CRF layer, wherein the village-level geographic name knowledge graph is embedded in the BERT layer.

[0039] The non-standard address text analysis module is configured to obtain a non-standard address text, obtain a structured annotation text from the non-standard address text by using the hybrid model, obtain a precise association result of the non-standard address and the geographic coordinates and a standardized address text based on the structured annotation text by means of three-level address analysis confidence scoring and dynamic switching of a matching algorithm in combination with precise association features of the high-quality space-time data and the standard address longitude and latitude mapping table.

[0040] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps in the multi-source dynamic collaborative geographic information mutual analysis method when executing the computer program.

[0041] A computer readable storage medium stores a computer program, and the computer program implements the steps in the multi-source dynamic collaborative geographic information mutual analysis method when executed by a processor.

[0042] The multi-source dynamic collaborative geographic information mutual analysis method, device, equipment and medium described above, by acquiring time-synchronized static geographic information, dynamic geographic information and a set of unstructured text data containing descriptions of multiple pairs of starting point and ending point geographic locations, wherein the dynamic geographic information includes GPS positioning data corresponding to each pair of starting point and ending point locations in the set of unstructured text data, constructing a unified space-time framework with time stamps and latitude and longitude as indexes based on the static geographic information, dynamic geographic information and starting and ending point unstructured geographic text data, obtaining a dynamic space-time data pool, performing double anomaly checking on the set of unstructured text data based on GPS positioning data and static geographic information to determine abnormal data with text information deviating from the standard administrative region, based on the Kalman filter fusion algorithm, taking GPS positioning data as the core reference observation value and static geographic information as the auxiliary observation value, dynamically adjusting the weight of each data to eliminate random errors of a single data through an iterative mechanism, while performing semantic completion and error correction on the abnormal data to obtain high-quality space-time data, performing triple-level weight allocation based on the binding relationship of the three-level administrative range in combination with the dynamic high-frequency words extracted from the unstructured text data set corrected from the high-quality space-time data, generating an address latitude and longitude mapping table containing confidence, and optimizing the address latitude and longitude mapping table to obtain a standard address latitude and longitude mapping table, constructing a BERT-CRF hybrid model including sequentially connected BERT layers and CRF layers, using the BERT-CRF hybrid model to obtain structured annotated text based on non-standard address text, based on the structured annotated text, through three-level address analysis confidence scoring and dynamic switching matching algorithm, in combination with the precise association features of the standard address latitude and longitude mapping table and the high-quality space-time data, obtaining the precise association results of non-standard addresses and geographic coordinates and standardized address text. By using this method, through multi-source space-time data fusion, Kalman filter adaptive weight adjustment, double regularization engine optimization and BERT-CRF model combined with village-level geographic name knowledge graph, millimeter-level precise positioning is realized, the accuracy of non-standard address analysis is improved, the problem of non-standard address expression in villages is successfully solved, the mis-matching rate of addresses with the same name is effectively reduced, through the fusion of static geographic data and dynamic high-frequency words to adapt to new scenes, the dynamic adaptability is improved, the data reliability is ensured through double anomaly checking and data cleaning, and the analysis efficiency and accuracy are considered in combination with the hierarchical adaptive matching strategy, providing efficient and reliable technical support for the bidirectional precise mutual analysis of geographic coordinates and text addresses. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of a multi-source dynamic collaborative geographic information mutual analysis method in one embodiment;

[0044] Figure 2 A multi-source fusion architecture diagram in one embodiment;

[0045] Figure 3A flowchart of a process for converting non-standard address text into accurate geographic location in an embodiment;

[0046] Figure 4 A structural block diagram of a multi-source dynamic collaborative geographic information mutual analysis device in an embodiment;

[0047] Figure 5 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0048] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0049] In order to solve the problems of large positioning error, high semantic analysis failure rate and poor dynamic scene adaptability caused by single-source dependence of data in the prior art, as shown in Figure 1 , a multi-source dynamic collaborative geographic information mutual analysis method is provided, which specifically includes the following steps:

[0050] Step S100, acquiring time-synchronized static geographic information, dynamic geographic information and a non-structured text data set containing descriptions of multiple pairs of starting and ending geographic locations, wherein the dynamic geographic information includes positioning data corresponding to each pair of starting and ending locations in the non-structured text data set.

[0051] Step S110, based on the static and dynamic geographic information and the non-structured geographic text data, constructing a dynamic spatio-temporal data pool and determining abnormal data in the non-structured text data set.

[0052] Step S120, based on the Kalman filter fusion algorithm, taking the positioning data as the reference observation value and the static geographic information as the auxiliary observation value, dynamically adjusting the weights of each data through an iterative mechanism to eliminate random errors of single data, while performing semantic completion and error correction on the abnormal data, to obtain high-quality spatio-temporal data.

[0053] Step S130, according to the three-level administrative range binding relationship, combining the dynamic high-frequency words extracted from the non-structured text data set corrected from the high-quality spatio-temporal data to perform three-level weight distribution, generating an address latitude and longitude mapping table containing confidence, and optimizing to obtain a standard address latitude and longitude mapping table.

[0054] Step S140, constructing a hybrid model, the hybrid model including a BERT layer and a CRF layer, wherein a village-level geographic name knowledge graph is embedded in the BERT layer.

[0055] In step S150, the non-standard address text is obtained, the structured address text is obtained from the non-standard address text by using the hybrid model, the structured address text is used to obtain the precise association result of the non-standard address and the geographic coordinate and the standardized address text by means of the three-level address resolution confidence score and the dynamic switching matching algorithm, the standard address latitude and longitude mapping table, and the precise association characteristics of high-quality space-time data.

[0056] In the method, the core includes three parts: the first part is the collection and fusion cleaning part of the multi-source space-time data shown in steps S100 and S110, in which the static geographic information, dynamic geographic information, online car operation data, and unstructured text data set are synchronously obtained, and a dynamic space-time data pool covering the static geographic information and dynamic scene data is constructed. The second part is the hybrid matching and double regularization address mapping part shown in steps S110 and S130, in which the mapping relationship between the standard address and the latitude and longitude is constructed based on the weight distribution of the administrative range binding and high-frequency statistics, and the double regularization engine and multi-layer similarity decision are adopted to realize the automatic error correction mechanism of the mapping table, to combine the static data of the administrative division of the OSM open source data set and the dynamic high-frequency words in the start and end point address data of the online car operation system, and to form a hybrid matching mode of the static geographic skeleton and the dynamic language characteristics. Compared with the traditional method which only depends on the OSM or POI database, the method is more suitable for the complexity of the Chinese local non-standard address. The double regularization engine combining the road network knowledge graph and the colloquial analysis is adopted to simultaneously process the structured road name (such as “Guangqu Road No. 12”) and the unstructured description (such as “the right side of the east gate of Joy City”), and the cross-modal alignment is realized by the cosine similarity. The third part is the dynamic threshold resolution and hierarchical adaptive matching shown in steps S140 and S150, the BERT-CRF hybrid model and the multi-layer similarity decision are adopted to realize the dynamic threshold resolution of the non-standard address, the BERT layer embeds the village-level place name knowledge graph, the CRF layer optimizes the address structure recognition of “the east side of the village committee”, the matching strategy is dynamically switched according to the scores of the city, district, and village levels, the city-level matching emphasizes the trunk road, the district-level focuses on the regional feature matching, and the village-level emphasizes the micro-orientation, to avoid the performance bottleneck of a single algorithm in processing addresses of different levels.

[0057] In step S100, the static geographic information includes OSM administrative division data, OSM road network data, POI data, and AOI data. The OSM administrative division data covers four-level administrative boundary vector coordinates, administrative codes, and hierarchical correlation relationships of provinces, cities, and districts (villages), and provides a standard administrative range benchmark for address resolution. The OSM road network data includes centerline coordinates, lane numbers, road levels, intersection topological relationships, and other structured information of urban trunk roads, secondary roads, branch roads, and other roads, and supports road-level positioning and path correlation verification. The POI (point of interest) data covers accurate coordinates, names, and attribute labels of various entities such as commercial facilities (shopping malls, supermarkets), public service agencies (schools, hospitals), transportation hubs (bus stations, subway stations), and village-level landmarks (small stores, village committees), and provides a core reference for address semantic matching. The AOI (area of interest) data defines the spatial range of commercial districts, industrial parks, residential complexes, and natural scenic areas in the form of polygon boundary coordinates, and realizes the attribution determination of regional addresses.

[0058] In this embodiment, the dynamic geographic information is GPS positioning data corresponding to the starting points and ending points of each trip automatically collected by the online car-hailing operation system. The GPS positioning data corresponds to the starting points and ending points of each trip of the online car-hailing, and specifically includes trip ID, timestamp, latitude and longitude coordinates (positioning accuracy up to meters), number of positioning satellites, signal strength, and other key attributes. The collection process is time-synchronized with the unstructured geographic text data input by the user, and is uniquely associated through the trip ID, ensuring accurate matching of coordinate data and text description for the same trip, and providing core benchmark observation values for subsequent abnormal data verification, multi-source data fusion, and semantic error correction.

[0059] In this embodiment, the unstructured text data set used to describe multiple pairs of starting point and ending point geographic locations is derived from the real-time input information of each trip in the online car-hailing operation system, and is a dynamic text collection covering a large number of real travel scenarios. Each trip corresponds to a set of starting point text and ending point text, forming a text data accumulation of multiple pairs of starting and ending points, rather than a single isolated description.

[0060] Specifically, the core feature of the unstructured text dataset is the non-standardized expression, i.e., no fixed format constraint, which completely conforms to the user's daily colloquial and scenario-based address description habits, and specifically includes two types of core content: one is the semi-standardized description containing partially structured elements, such as "Guangqu Road No. 12", "XX District Science and Technology Park nearby", and "XX subway station B exit", which contains clear geographical identifiers such as road names, administrative regions, and transportation hubs; the other is completely unstructured and ambiguous description, such as "the east head of Wangjiatun three-way intersection", "50 meters to the right of the east gate of Joy City", "the east road intersection of the village committee", and "20 meters west of the small shop at the south head", which relies on village-level landmarks, direction words ("east head", "right hand side", "west side"), and relative position relationships to achieve address pointing, and lacks a unified standard expression.

[0061] In step S110, as shown in Figure 2 The static geographic information, dynamic geographic information, and start and end point unstructured geographic text data are indexed by time stamp and longitude and latitude to construct a unified space-time framework to obtain a dynamic space-time data pool and form a structured data storage system.

[0062] In step S110, the double anomaly check includes: matching the GPS positioning data with the OSM administrative division vector data boundary coordinates to determine the corresponding standard administrative region information, calculating the string similarity of the unstructured text and the standard administrative region information by an edit distance algorithm, and verifying whether the text mentioned place falls within the OSM administrative region corresponding to the GPS positioning data, and determining as abnormal data when the similarity is lower than a preset threshold or the place exceeds the administrative region range.

[0063] In this embodiment, the GPS positioning data is associated and matched with the OSM administrative division vector data boundary coordinates to determine the standard administrative region information (such as XX province XX city XX district XX street XX community) corresponding to the longitude and latitude. Then, the standard information is compared with the start and end point unstructured geographic text information filled by the user, i.e., the unstructured description of the start and end point in the unstructured text dataset, such as "XX community gate", "XX road nearby", and other possible biased descriptions, to determine whether the text mentioned place falls within the OSM administrative region corresponding to the longitude and latitude by string similarity matching, edit distance algorithm, and POI association verification, and to preliminarily filter out abnormal data with obvious deviation between the text information and the standard administrative region, such as "Chaoyang District" corresponding to the longitude and latitude but "Haidian District" filled in the text.

[0064] Specifically, in the edit distance algorithm, the core keywords of the standard administrative region are extracted: the hierarchical keyword set {“Beijing City” “Chaoyang District” “Jiangwai Street” “Guanghua Li Community”} is extracted from “Beijing City Chaoyang District Jiangwai Street Guanghua Li Community”. The administrative related keywords in the text are extracted: the user text “Haidian District Guanghua Li Community Gate” is segmented to get {“Haidian District” “Guanghua Li” “Community” “Gate”}, and the administrative keywords {“Haidian District” “Guanghua Li”} are screened. The edit distance of the text keywords and the standard keywords is compared one by one. For example, the edit distance between “Haidian District” and “Chaoyang District” is 2 (replace “Hai” with “Chaoyang”), and the similarity = 1- (2 / 3) ≈ 0.33; the edit distance between “Guanghua Li” and “Guanghua Li Community” is 2 (add “Community”), and the similarity = 1- (2 / 5) = 0.6. Take the highest similarity 0.6, which is lower than the threshold 0.7, trigger the abnormal warning. Then, the POI entity in the text is extracted: the POI “Guanghua Li Community” is extracted from “Haidian District Guanghua Li Community Gate”. Query the standard administrative attribution of the POI: call the POI database to get the administrative region of “Guanghua Li Community” as “Beijing City Chaoyang District Jiangwai Street Guanghua Li Community” (consistent with the GPS matching result). The text clearly marks “Haidian District”, which conflicts with the actual attribution of “Chaoyang District” of the POI, further confirming that the text is abnormal data.

[0065] Further, through the further data cleaning and text correction in step S120, in this embodiment, based on the Kalman filter fusion algorithm, taking the GPS positioning data as the core reference observation value and the static geographic information as the auxiliary observation value, the random error of single data is eliminated by dynamically adjusting the weight of each data through the iteration mechanism, including: based on the latitude and longitude reference value of the GPS positioning data, predicting the theoretical error range of the geographic information, and calculating the theoretical error covariance, and then combining the actual deviation of each auxiliary observation value to dynamically adjust the weight of each data source.

[0066] In this embodiment, the Kalman filter fusion algorithm is used to correct the abnormal data, taking the accurate taxi starting and ending point GPS data as the core reference observation value, combining the road network in the OSM map, the boundary coordinates of the OSM administrative division vector data and the surrounding POI, AOI distribution coordinates and other auxiliary observation values, through the prediction and update iteration mechanism of the Kalman filter, first predicting the theoretical error range of the geographic information based on the latitude and longitude reference value and calculating the theoretical error covariance, and then combining the actual deviation of the auxiliary observation value to dynamically adjust the weight of each data source, eliminating the random error of single data.

[0067] Specifically, in the Kalman filter fusion algorithm for error correction of abnormal data, first calculate the actual distance deviation of each type of auxiliary data from the GPS reference coordinates, such as a POI being 3 meters away from GPS, OSM road network being 14 meters away from GPS, and administrative center being 7 meters away from GPS. The second step is to adjust the weight according to the deviation: the smaller the deviation, the higher the weight, such as the POI of 3 meters being increased from 10% to 20%, the OSM road network being adjusted to 15.2%, and the weight being greatly reduced when the deviation is large; After adjustment, ensure that the weight of all data is 100%, appropriately reduce the core weight of GPS (such as to 39.8%), and let the reliable auxiliary data play a greater role. The third step is to fuse all data: according to the adjusted weight and the error of each type of data, calculate the reliability of each type of data, and then fuse GPS and all auxiliary data according to the reliability to obtain more accurate coordinates (such as 116.4800°E, 39.9182°N), and the positioning accuracy is improved to ±2.8 meters.

[0068] Further, based on the accurate correction coordinates obtained after the Kalman filter fusion algorithm, the explicit geographic area with a significantly reduced error radius is obtained, and the optimized spatial range is used to perform semantic completion and error correction on the determined abnormal data in the non-structured geographic text filled by the user, such as the longitude and latitude positioning in the surrounding area of "No. 88, Jianguo Road, Chaoyang District", and the user text fills in "No. 8, Jianguo Road, Haidian District", which is corrected to "XX building, No. 88, Jianguo Road, Chaoyang District" combined with POI data; If the user only fills in "XX Building" and the building has a name conflict in the current administrative area, match the name of the target building closest to the longitude and latitude. Finally, output the associated data pairs of accurate longitude and latitude and corrected geographic text to form high-quality location trajectory basic information, which provides reliable input with both coordinate accuracy and text accuracy for subsequent address resolution and spatial indexing processing. Through multiple accurate longitude and latitude and corrected geographic text associated data pairs, high-quality space-time data is constructed.

[0069] In step S130, based on the province-city-district (village) three-level administrative range binding and the dynamic high-frequency words (such as "Joy City" and "Science and Technology Park" repeatedly appearing place names) extracted from the online car text data, a weight coefficient of the address-longitude and latitude mapping table is generated through a preset weight calculation model (such as province-level administrative range weight 30%, city-level 35%, village-level 35%, and high-frequency word additional 10% weight), and the confidence of each mapping relationship is labeled, such as "Beijing Chaoyang District (116.48°E, 39.92°N)", weight 95%, to obtain the address-longitude and latitude mapping table, and form a weight adjustment rule library to record the influence logic of high-frequency words on weight.

[0070] Furthermore, the address latitude and longitude mapping table is optimized to obtain a standard address latitude and longitude mapping table. This includes processing structured and unstructured addresses in the address latitude and longitude mapping table separately using a dual regularization engine. Specifically, structured addresses containing road names, house numbers, and intersections are parsed using a road network knowledge graph, while unstructured addresses with directional descriptions are identified using a colloquial parsing engine. The semantic similarity between the two types of parsing results is calculated using a cosine similarity algorithm. When the similarity meets a preset standard, the addresses are determined to be from the same source and are fused to obtain a preliminary optimized address latitude and longitude mapping table. Based on the preliminary optimized address latitude and longitude mapping table, a standard address latitude and longitude mapping table is obtained according to a hybrid matching pattern formed by OSM static administrative division data and dynamic high-frequency words.

[0071] Specifically, based on the road network knowledge graph (including structured information such as road names, house numbers, and intersections), structured addresses such as "No. 12 Guangqu Road" in the address latitude and longitude mapping table are parsed. Simultaneously, an oral parsing engine processes unstructured addresses such as "50 meters east of Joy City" in the address latitude and longitude mapping table. The parsing results for the structured addresses are output along with their corresponding latitude, longitude, and administrative level, and then matched with the parsing results for the unstructured addresses to obtain specific POI coordinates and semantic tags. The oral parsing engine is a pre-trained Chinese geographic text understanding model that can recognize unstructured descriptions such as "right-hand side" and "opposite side."

[0072] Furthermore, the semantic similarity between structured and unstructured address parsing results is calculated using the cosine similarity algorithm (preferably, a similarity ≥ 0.8 is set to determine the same source address), and the two types of outputs are merged to form a unified address mapping result (e.g., when structured and unstructured parsing are from the same source, they are merged into XX Building No. 12 Guangqu Road + precise coordinates), and the alignment confidence score is labeled.

[0073] Next, the OSM static administrative division data (which provides a fixed geographic skeleton) is deeply integrated with the ride-hailing dynamic high-frequency word database (popular place names updated in real time, such as newly built business districts), and incorporated into the Chinese local address feature database (including characteristic address suffix rules such as "village", "hamlet", and "lane") to further optimize the address latitude and longitude mapping table and obtain the standard latitude and longitude mapping table.

[0074] In step S140, the aim is to construct a BERT-CRF hybrid model adapted to non-standard address parsing. Through a collaborative architecture of semantic understanding and structural recognition, it overcomes the technical challenges of non-standard village-level address representation and ambiguous component division. The model as a whole adopts a concatenated structure of BERT layer semantic encoding and CRF layer sequence annotation.

[0075] Specifically, when constructing the BERT-CRF hybrid model, based on the general BERT-CRF architecture (encoding layer → projection layer → decoding layer), customization optimization is made for the characteristics of "village address component ambiguity (such as 'Wangjiatun East Head'), strong semantic dependence (such as 'Village Committee West Intersection')". The overall structure is divided into 4 layers in series, among which the key is the deep integration of BERT layer and village-level place name knowledge graph. The BERT-CRF hybrid model includes input preprocessing layer, BERT semantic encoding layer, linear projection layer and CRF decoding layer.

[0076] Specifically, in the input preprocessing layer, the WordPiece tokenizer of BERT is used for segmentation adaptation, but for the colloquial expressions in village addresses (such as "San Chakou" and "Xiaomaiduo"), a "village address exclusive word list" (generated based on entities in the village-level place name knowledge graph) is supplemented to avoid splitting "Wangjiatun" into "Wang", "Jia", and "Tun", and to ensure the integrity of core entities such as village names and landmarks. Special tokens are added at the beginning and end of the input text, and the annotation format is added, <cls>(sequence start), <sep>(End of sequence), the annotation format uses the BIO system, focusing on the core entity types of village-level addresses: B-VILL (start of village name, such as "Wang" in "Wang-Jia-Tun"), I-VILL (internal village name, such as "Jia" and "Tun"), B-LAND (start of landmark, such as "San" in "San-Cha-Kou" and "Cunwei"), I-LAND (internal landmark), B-DIR (start of direction, such as "Dong" in "Dong-Tou" and "Xi-Ci"), I-DIR (internal direction), and O (non-entity, such as "de" and "nearby"). Example: "Wang-Jia-Tun-Dong-Tou-San-Cha-Kou" is annotated as [B-VILL, I-VILL, I-VILL, B-DIR, I-DIR, B-LAND, I-LAND, I-LAND].

[0077] Specifically, in the BERT semantic encoding layer, the underlying architecture uses a 12-layer bidirectional Transformer encoder with a hidden layer dimension of 768 and 12 attention heads. Compared with the general BERT, an "address semantic feature embedding" is additionally added - the "direction words (east / west / south / north)" and "landmark suffixes (Tun / Zhuang / Ao / Cha Kou)" that frequently appear in village-level addresses are converted into fixed vectors, which are concatenated with the BERT token embedding and position embedding to strengthen the semantic differentiation of "Dong-Tou" and "Wang-Jia-Tun" type exclusive expressions.

[0078] Specifically, in the linear projection layer, the 768-dimensional semantic vector output by the BERT encoding layer is mapped to the "village-level address entity label space" (in this scenario, the number of labels is 7: B-VILL / I-VILL / B-LAND / I-LAND / B-DIR / I-DIR / O) through a fully connected layer, generating "emission scores" for each token corresponding to each label (for example, the score of "Dong-Tou" corresponding to B-DIR is 1.8, and the score corresponding to O is -0.5). To avoid overfitting caused by the small sample size of village-level addresses, a Dropout layer (dropout rate = 0.1) is added after the fully connected layer, and L2 regularization (λ = 1e-5) is applied to the weights to ensure the model's ability to distinguish between similar addresses such as "Li-Jia-Ao" and "Li-Jia-Cun".

[0079] Specifically, the CRF decoding layer is used to solve the problem of general BERT "label prediction unconstrained". According to the rule of "entity order fixed" in the village address (such as "village name → direction → landmark", instead of "direction → village name"), the dependency relationship between labels is modeled (such as "B-DIR" can only be followed by "B-VILL / I-VILL", and cannot be followed by "O"). In this layer, a "village address label transition matrix" is constructed, and is initialized based on annotation data statistics (such as the transition probability of "B-VILL→I-VILL" is set to 0.95, the transition probability of "B-VILL→B-DIR" is set to 0.04, and the transition probability of "O→B-DIR" is set to 0.01), and the label sequence with the optimal "emission score + transition probability" is selected by the Viterbi algorithm, to ensure the rationality of address component division (such as avoiding mislabeling "Wang Jia Tun Dong Tou" as "B-DIR→I-DIR→B-VILL").

[0080] In step S150, the non-standard address text is converted using the high-quality spatio-temporal data, the standard address latitude and longitude mapping table, and the BERT-CRF hybrid model obtained in the above steps, and the process is as shown in Figure 3

[0081] In this embodiment, the structured annotation text is obtained by using the BERT-CRF hybrid model first. The village-level place name knowledge graph embedded by the BERT layer includes village names, village landmarks, and adjacent village relationships. The non-standard address text is converted into a semantic vector with multiple layers of semantic information by using a semantic vector generation tool. Then, the CRF layer receives the semantic vector and performs structured annotation on the compound address structure, and outputs the structured annotation text with clear component division.

[0082] Specifically, the village-level place name knowledge graph (including village names, village landmarks, and adjacent village relationships) is loaded in the BERT layer of the pre-trained BERT model. The core content of the village-level place name knowledge graph is a "triple set", which includes three key relationships: { (village name, contains, landmark); (village name, adjacent, village name); (landmark, direction, village name)}. The fusion process refers to the "knowledge injection logic" of K-BERT, but is simplified according to the characteristics of short village address text and high entity density, and is implemented in three steps including knowledge query (K-Query), knowledge injection (K-Inject), and semantic adaptation (Soft-Position + visible matrix).

[0083] ​Specifically, knowledge query (K-Query): = accurate matching of entities in address text and graph triples. Input the original village-level non-standard address text, village-level place name knowledge graph K; based on rule matching, the village name and landmark in the knowledge graph are pre-screened for entity recognition, and candidate entities are extracted from the text: village name "Wangjatun", landmark "Dongtou Sanjiaokou"; query the triples associated with the candidate entities from the knowledge graph K, and get: T1= (Wangjatun, contains, Dongtou Sanjiaokou), T2= (Dongtou Sanjiaokou, located, Wangjatun east side); finally output the filtered triple set E strongly related to the current address text.

[0084] Specifically, knowledge injection (K-Inject), construct "address-knowledge" fused sentence tree. Embed the knowledge in the triples into the original address text to form a sentence tree with the original text as the main chain and the knowledge supplement as the branch, solving the problem of incomplete village address expression.

[0085] Specifically, semantic adaptation (Soft-Position + visible matrix) is used to eliminate knowledge noise and retain address semantics. Soft-Position Embedding: Assign the original text token a main position number (e.g., "Wangjatun" = 1, "Dongtou" = 2, "Sanjiaokou" = 3), and the knowledge supplement token is assigned a position number and offset associated with the main token (e.g., <k>contains the association "Wangjiatun", position number = 1.1; <k>Located in the association "East Sanjiaokou", location number = 3.1), to ensure that BERT can identify the relationship that the knowledge supplement token depends on the original token. Visible matrix Seeing Layer: construct a 2D visible matrix, knowledge supplement tokens (such as, <k>(contains) only be able to associate with the original token (Wangjiatun) and the same branch of knowledge token (such as, <k>The East Three Forks) produces attention interaction and cannot interact with other original tokens (e.g., "East Head").

[0086] Further, the non-standard address semantic vector output by the BERT layer is input into a CRF (Conditional Random Field) model, i.e., a pre-trained address structure labeler, to accurately analyze the complex address structure of "village committee east intersection", "school north gate bus station", etc., and output a structured labeling result (e.g., the address "Li Gezhuang Village South Head Small Shop 20 meters west" is output as:

Village Name: Li Gezhuang

Landmark: South Head Small Shop

Direction: West 20 meters

[0087] In the present embodiment, based on the structured labeled text, through three-level address resolution confidence scoring and dynamic switching of matching algorithms, combined with the standard address latitude and longitude mapping table and the precise association characteristics of high-quality spatio-temporal data, the precise association result of non-standard address and geographic coordinates and the standardized address text are obtained, including: using a three-level address resolution confidence scoring model, the village name, landmark, and direction information in the structured labeled text are calculated for city, district, and village level resolution confidence scores, according to the city-level confidence score result, the city-level trunk road positioning algorithm, or the district-level regional POI clustering algorithm, or the village-level micro-orientation positioning algorithm is used according to the switching mechanism, combined with the pre-set matching algorithm, the standard address latitude and longitude mapping table and the structured labeled text are matched, while referring to the precise association characteristics of high-quality spatio-temporal data, the matching result is error corrected to obtain the geographic coordinate data corresponding to the structured labeled text, finally, according to the standardized address expression rules, based on the geographic coordinate data, the precise association result of non-standard address and geographic coordinates and the standardized address text in the standard format are generated

[0088] Specifically, the city / district / village three-level resolution confidence is calculated by a three-level address resolution confidence scoring model. The city-level score looks at whether the BERT-CRF model resolves a clear city name, and then compares and matches the standard administrative area (with higher weight). For example, "Ligaocun Village, 20 meters west of the small shop at the south end" does not directly mention the city name, but through the village name it is found that it belongs to "Tangshan City", but the matching degree is 75 points, and the final city-level score is 45 points. The district-level score takes into account whether the district name is resolved and whether the binding of the city-level administrative area is consistent (for example, "Fengnan District" must belong to "Tangshan City"), as well as the matching coverage rate of POIs in the area. The above address does not mention the district name, but it is found that "Fengnan District" is consistent and there is a matching POI in the area, so the final district-level score is 65 points. The village-level score focuses on whether the village name is complete and whether the binding with the district-level is valid, as well as the matching degree of micro-orientation (such as "20 meters west") and landmarks (such as "small shop at the south end"). The address has clear village name, landmark, and orientation, with a village-level score of 92 points. In addition, if a level does not directly resolve an entity (such as city, district), the upper limit of the score will be limited by the reverse query of the next level; if the resolution result conflicts with the reverse query, it will also be penalized.

[0089] Set the policy switching threshold (such as "city-level confidence < 60% switch to district-level strategy"), and dynamically select the matching algorithm from the hierarchical matching algorithm library based on the scoring results: city-level address matching focuses on trunk road positioning algorithm, district-level focuses on regional POI clustering algorithm, and village-level emphasizes micro-orientation positioning algorithm, finally outputting the precise non-standard address latitude and longitude association results and standardized address text. If the city-level score is ≥ 60 points, use the city trunk road as the anchor point, combined with the road name and orientation in the text to quickly lock the position (such as "Beijing Guangqu Road, 500 meters east", directly match the Guangqu Road coordinates and then offset). If the city-level score is < 60 points but the district-level is ≥ 70 points, focus on the clustering core of POIs (such as shopping malls, bus stations, etc.) in the district, combined with the POI type matching in the text (such as "Shanghai Pudong New Area, Lujiazui South", find the clustering center of Lujiazui commercial POIs and then correct). If the city-level and district-level scores are not enough (such as 45 points and 65 points above), find the precise coordinates of the village-level landmark (such as "small shop at the south end" coordinates 116.5200°E, 39.8805°N) from the standard mapping table, then calculate the offset coordinates according to the orientation (west) and distance (20 meters), and finally fine-tune by referring to high-quality data such as GPS trajectories of online car-hailing services to ensure an error of ±3 meters.

[0090] In the above multi-source dynamic collaborative geographic information mutual analysis method, through real-time cross verification of GPS positioning data, OSM multi-dimensional geographic data and online car-hailing dynamic data, and self-adaptive weight adjustment of Kalman filter fusion algorithm, a multi-source dynamic fusion system is constructed, which effectively offsets the random error of single data and the influence of GPS drift in complex road network scenes, and compresses the positioning error of the traditional scheme to 500 meters to 10 meters, providing reliable meter-level spatial reference for automatic driving intelligent transportation and other scenes. At the same time, relying on the mixed model architecture of BERT layer embedded village-level name knowledge graph and CRF layer optimized composite address structure recognition, combined with double regularization engine and cross-modal alignment technology, and combined with city, county and village three-level dynamic switching matching strategy, the industry problem of fuzzy division of non-standard address expression is accurately solved, and the accuracy of non-standard address analysis is improved from 39.2% of the traditional scheme to 98.7%, and the semantic failure problem is completely solved. In addition, based on the improved clustering algorithm of road topological constraint, combined with the multi-source features in the dynamic spatio-temporal data pool, the stop point recognition accuracy reaches 93.8%, successfully mines 41% of the implicit high-frequency stop points missed by the traditional method, and through the mixed matching mode of static geographic skeleton and dynamic high-frequency words, accurately associates various POI entities such as commercial district traffic hub and village-level landmarks, and comprehensively improves the correctness and practicality of geographic information mutual analysis, and provides strong data support for business site selection and city management scenes.

[0091] It should be understood that, although Figure 1 The steps in the flowchart of the method are shown in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the method can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0092] In one embodiment, as shown in Figure 4 A multi-source dynamic collaborative geographic information mutual analysis device is provided, comprising: a multi-source geographic information acquisition module 200, a dynamic spatio-temporal data pool construction and abnormal data determination module 210, a high-quality spatio-temporal data obtaining module 220, a standard address latitude and longitude mapping table construction module 230, a mixed model construction module 240 and a non-standard address text analysis module 250, wherein:

[0093] The multi-source geographic information acquisition module 200 is configured to acquire time-synchronized static geographic information, dynamic geographic information, and a non-structured text data set containing unstructured text data for describing pairs of starting point and ending point geographic positions, wherein the dynamic geographic information includes positioning data corresponding to each pair of starting point and ending point positions in the non-structured text data set.

[0094] The dynamic spatio-temporal data pool construction module 210 is configured to construct a dynamic spatio-temporal data pool based on the static, dynamic geographic information, and the non-structured geographic text data, and determine abnormal data in the non-structured text data set; the high-quality spatio-temporal data obtaining module 220 is configured to, based on a Kalman filter fusion algorithm, take the positioning data as a reference observation value and the static geographic information as an auxiliary observation value, dynamically adjust the weight of each data to eliminate random errors of a single data through an iteration mechanism, and simultaneously perform semantic completion and error correction on the abnormal data to obtain high-quality spatio-temporal data.

[0095] The standard address longitude and latitude mapping table construction module 230 is configured to perform three-level weight distribution in combination with dynamic high-frequency words extracted from the non-structured text data set corrected from the high-quality spatio-temporal data according to a three-level administrative range binding relationship, generate an address longitude and latitude mapping table containing a confidence degree, and optimize the address longitude and latitude mapping table to obtain a standard address longitude and latitude mapping table.

[0096] The hybrid model construction module 240 is configured to construct a hybrid model, the hybrid model including a BERT layer and a CRF layer, wherein the village-level geographic name knowledge graph is embedded in the BERT layer.

[0097] The non-standard address text analysis module 250 is configured to acquire a non-standard address text, obtain a structured annotated text from the non-standard address text by using the hybrid model, and obtain a precise association result of the non-standard address and the geographic coordinates and a standardized address text based on the structured annotated text, through three-level address analysis confidence scoring and dynamic switching of a matching algorithm, in combination with precise association features of the standard address longitude and latitude mapping table and the high-quality spatio-temporal data.

[0098] The specific limitations of the multi-source dynamic collaborative geographic information mutual analysis device can be referred to the limitations of the multi-source dynamic collaborative geographic information mutual analysis method in the foregoing, which will not be described herein. Each module in the multi-source dynamic collaborative geographic information mutual analysis device described above can be realized by software, hardware, or a combination thereof, in whole or in part. Each module described above can be embedded in or independent of a processor in a computer device in a hardware form, or can be stored in a memory in a computer device in a software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0099] In one embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram of the computer device can be as shown in Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a multi-source dynamic collaborative geographic information mutual analysis method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0100] Those skilled in the art can understand that, Figure 5 The skilled in the art can understand that,

[0101] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the following steps:

[0102] Obtaining time-synchronized static geographic information, dynamic geographic information and a set of unstructured text data containing descriptions of multiple pairs of start and end geographic locations, wherein the dynamic geographic information includes positioning data corresponding to each pair of start and end locations in the set of unstructured text data;

[0103] Based on the static, dynamic geographic information and unstructured geographic text data, a dynamic spatio-temporal data pool is constructed, and abnormal data in the set of unstructured text data is determined;

[0104] Based on the Kalman filter fusion algorithm, taking the positioning data as the reference observation value and the static geographic information as the auxiliary observation value, the data weight is dynamically adjusted through the iteration mechanism to eliminate the random error of single data, and the abnormal data is semantically completed and corrected to obtain high-quality spatio-temporal data;

[0105] According to the three-level administrative range binding relationship, combined with the dynamic high-frequency words extracted from the set of unstructured text data corrected from the high-quality spatio-temporal data, a three-level weight distribution is performed to generate an address latitude and longitude mapping table containing confidence, and the standard address latitude and longitude mapping table is obtained after optimization;

[0106] construct a hybrid model, the hybrid model comprising a BERT layer and a CRF layer, wherein a village-level geographical name knowledge graph is embedded in the BERT layer;

[0107] obtain non-standard address text, obtain structured annotation text from the non-standard address text by using the hybrid model, and obtain a precise association result of the non-standard address and geographical coordinates and a standardized address text based on the structured annotation text, through three-level address resolution confidence scoring and dynamic switching of a matching algorithm, in combination with the standard address latitude and longitude mapping table and precise association features of high-quality space-time data.

[0108] In one embodiment, a computer-readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0109] obtain time-synchronized static geographical information, dynamic geographical information, and a non-structured text data set containing unstructured text data for describing pairs of start and end geographical positions, wherein the dynamic geographical information comprises positioning data corresponding to each pair of start and end positions in the non-structured text data set;

[0110] based on the static and dynamic geographical information and the non-structured geographical text data, construct a dynamic space-time data pool, and determine abnormal data in the non-structured text data set;

[0111] based on a Kalman filter fusion algorithm, taking the positioning data as a reference observation value and the static geographical information as an auxiliary observation value, dynamically adjusting the weight of each data through an iteration mechanism to eliminate random errors of a single data, while performing semantic completion and error correction on the abnormal data, to obtain high-quality space-time data;

[0112] based on a three-level administrative range binding relationship, performing three-level weight distribution in combination with dynamic high-frequency words extracted from the non-structured text data set corrected from the high-quality space-time data, generating an address latitude and longitude mapping table containing confidence, and optimizing to obtain a standard address latitude and longitude mapping table;

[0113] construct a hybrid model, the hybrid model comprising a BERT layer and a CRF layer, wherein a village-level geographical name knowledge graph is embedded in the BERT layer;

[0114] obtain non-standard address text, obtain structured annotation text from the non-standard address text by using the hybrid model, and obtain a precise association result of the non-standard address and geographical coordinates and a standardized address text based on the structured annotation text, through three-level address resolution confidence scoring and dynamic switching of a matching algorithm, in combination with the standard address latitude and longitude mapping table and precise association features of high-quality space-time data.

[0115] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0116] The technical features of the above embodiments can be combined in any way. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0117] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.< / k> < / k> < / k> < / k> < / sep> < / cls>

Claims

1. A multi-source dynamic collaborative geographic information mutual resolution method, characterized in that, The method includes: Acquire time-synchronized static geographic information, dynamic geographic information, and an unstructured text dataset containing multiple pairs of origin and destination geographic locations, wherein the dynamic geographic information includes GPS positioning data corresponding to each pair of origin and destination locations in the unstructured text dataset, and the static geographic information includes OSM administrative division data. Based on the static and dynamic geographic information and unstructured geographic text data, a dynamic spatiotemporal data pool is constructed. Based on the GPS positioning data and static geographic information, the unstructured text dataset undergoes dual anomaly verification to identify abnormal data where the text information deviates from the standard administrative region. The dual anomaly verification includes: associating and matching the GPS positioning data with the boundary coordinates of the OSM administrative division vector data to determine the corresponding standard administrative region information; calculating the string similarity between the unstructured text and the standard administrative region information using an edit distance algorithm; and verifying whether the location mentioned in the text falls within the OSM administrative region corresponding to the GPS positioning data. When the similarity is lower than a preset threshold or the location exceeds the administrative region range, it is determined to be abnormal data. Based on the Kalman filter fusion algorithm, with location data as the baseline observation and static geographic information as the auxiliary observation, the weight of each data is dynamically adjusted through an iterative mechanism to eliminate random errors in single data. At the same time, semantic completion and error correction are performed on the abnormal data to obtain high-quality spatiotemporal data. Based on the binding relationship of the three-level administrative scope, and combined with the dynamic high-frequency words extracted from the unstructured text dataset after correction from the high-quality spatiotemporal data, a three-level weight allocation is performed to generate an address latitude and longitude mapping table containing confidence, and then optimized to obtain a standard address latitude and longitude mapping table. A hybrid model is constructed, which includes a BERT layer and a CRF layer, wherein a village-level place name knowledge graph is embedded in the BERT layer; Obtain non-standard address text, use the hybrid model to obtain structured labeled text based on the non-standard address text, and based on the structured labeled text, use a three-level address resolution confidence score and a dynamic switching matching algorithm, combined with the standard address latitude and longitude mapping table and the accurate association features of high-quality spatiotemporal data, to obtain the accurate association results between non-standard addresses and geographic coordinates and standardized address text.

2. The multi-source dynamic collaborative geographic information mutual resolution method according to claim 1, characterized in that, The static geographic information also includes OSM road network data, POI data, and AOI data; The dynamic geographic information refers to the GPS positioning data corresponding to the origin and destination of the trip, which are automatically collected by the ride-hailing operation system. The static and dynamic geographic information, as well as the unstructured geographic text data, are used as indexes to construct a unified spatiotemporal framework, resulting in the dynamic spatiotemporal data pool.

3. The multi-source dynamic collaborative geographic information mutual resolution method according to claim 2, characterized in that, Based on the Kalman filter fusion algorithm, GPS positioning data is used as the core benchmark observation and static geographic information is used as the auxiliary observation. The weights of each data are dynamically adjusted through an iterative mechanism to eliminate random errors in single data. This includes: based on the latitude and longitude benchmark value of GPS positioning data, predicting the theoretical error range of geographic information and calculating the theoretical error covariance, and then dynamically adjusting the weights in combination with the actual deviations of each auxiliary observation.

4. The multi-source dynamic collaborative geographic information mutual resolution method according to claim 2, characterized in that, The address latitude and longitude mapping table is optimized to obtain a standard address latitude and longitude mapping table, which includes: The structured and unstructured addresses in the address latitude and longitude mapping table are processed by a dual regularization engine. Specifically, the structured addresses containing road names, house numbers, and intersections are parsed using a road network knowledge graph, while the unstructured addresses with directional descriptions are identified using a colloquial parsing engine. The semantic similarity between the two types of parsing results is calculated using the cosine similarity algorithm. When the similarity meets the preset standard, they are determined to be from the same source and are merged to obtain a preliminary optimized address latitude and longitude mapping table. Based on the preliminarily optimized address latitude and longitude mapping table, and according to the mixed matching pattern formed by the OSM administrative division data and dynamic high-frequency words, the standard address latitude and longitude mapping table is obtained.

5. The multi-source dynamic collaborative geographic information mutual resolution method according to claim 2, characterized in that, The structured labeled text obtained using the hybrid model includes: The village-level place name knowledge graph embedded in the BERT layer includes village names, village landmarks and relationships with neighboring villages. The non-standard address text is converted into a semantic vector with multi-layer semantic information through a semantic vector generation tool. After receiving the semantic vector, the CRF layer performs structured annotation on the composite address structure and outputs structured annotation text with clear component division.

6. The multi-source dynamic collaborative geographic information mutual resolution method according to claim 2, characterized in that, Based on the structured labeled text, through a three-level address resolution confidence score and a dynamic switching matching algorithm, combined with the standard address latitude and longitude mapping table and the accurate association features of high-quality spatiotemporal data, the accurate association results between non-standard addresses and geographic coordinates and standardized address text are obtained, including: A three-level address resolution confidence scoring model is adopted to calculate the resolution confidence scores at the city, district, and village levels for village names, landmarks, and location information in the structured labeled text. Based on the city-level confidence score, and according to the switching mechanism, the city-level main road positioning algorithm, or the district-level regional POI clustering algorithm, or the village-level micro-location positioning algorithm are used, combined with the preset matching algorithm, to match the standard address latitude and longitude mapping table with the structured annotation text. At the same time, the accurate correlation characteristics of the high-quality spatiotemporal data are referenced to correct the error of the matching result, so as to obtain the geographic coordinate data corresponding to the structured annotation text. Based on the standardized address representation rules and the geographic coordinate data, a precise association result between non-standard addresses and geographic coordinates, as well as standardized address text in a standardized format, are generated.

7. A multi-source dynamic collaborative geographic information mutual resolution device, characterized in that, The device provides a real-time geographic information mutual resolution method according to any one of claims 1-6, wherein the device comprises: A multi-source geographic information acquisition module is used to acquire time-synchronized static geographic information, dynamic geographic information, and an unstructured text dataset containing multiple pairs of start and end point geographic locations, wherein the dynamic geographic information includes location data corresponding to each pair of start and end point locations in the unstructured text dataset; The module for constructing a dynamic spatiotemporal data pool and identifying abnormal data is used to construct a dynamic spatiotemporal data pool based on the static and dynamic geographic information and unstructured geographic text data, and to identify abnormal data in the unstructured text dataset. The high-quality spatiotemporal data acquisition module is used to obtain high-quality spatiotemporal data based on the Kalman filter fusion algorithm, using location data as the baseline observation and static geographic information as the auxiliary observation. The module dynamically adjusts the weights of each data through an iterative mechanism to eliminate random errors in single data. At the same time, it performs semantic completion and error correction on the abnormal data. The standard address latitude and longitude mapping table construction module is used to perform a three-level weight allocation based on the binding relationship of the three-level administrative scope, combined with the dynamic high-frequency words extracted from the unstructured text dataset after correction from the high-quality spatiotemporal data, to generate an address latitude and longitude mapping table containing confidence, and optimize it to obtain the standard address latitude and longitude mapping table. A hybrid model building module is used to build a hybrid model, which includes a BERT layer and a CRF layer, wherein the BERT layer embeds a village-level place name knowledge graph; The non-standard address text parsing module is used to obtain non-standard address text, and to obtain structured labeled text based on the non-standard address text using the hybrid model. Based on the structured labeled text, through a three-level address resolution confidence score and a dynamic switching matching algorithm, combined with the standard address latitude and longitude mapping table and the accurate correlation features of high-quality spatiotemporal data, the module obtains the accurate correlation results between non-standard addresses and geographic coordinates and standardized address text.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images