Map search method and apparatus, and electronic device, storage medium and program product

Abstract

The embodiments of the present application can be applied to the field of maps. Disclosed in the embodiments of the present application are a map search method and apparatus, and an electronic device, a storage medium and a program product. The method comprises: by means of a spatial semantic model, acquiring a first spatial semantic vector corresponding to a search statement, wherein the spatial semantic model is obtained by means of training previous map search data, and the previous map search data includes text information and spatial information; by means of a spatial semantic index library, acquiring from a location point-of-interest vector library a location point-of-interest vector which matches the first spatial semantic vector, and using same as a target location point-of-interest vector corresponding to the first spatial semantic vector, wherein the location point-of-interest vectors in the location point-of-interest vector library are acquired by means of the spatial semantic model; and on the basis of the target location point-of-interest vector, determining a target location point of interest corresponding to the search statement. The present application can realize more accurate map searching.

Description

Map search method and device, electronic device, storage medium, and program product

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on February 8, 2024, application number 202410176284.3, and application name “Map search method and device, electronic device, storage medium, program product”, all contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of map technology, and in particular to a map search technology. Background Art

[0003] With the rapid development of computer and communication technologies, electronic maps have become widely used in people's daily travel. People can use electronic maps to search for and locate points of interest (POIs). As you can see, a POI is an area with a specific function, such as a residence or company.

[0004] In map search engines, recall is a crucial step in the search process, enabling users to retrieve relevant POIs from a vast pool of POIs based on a user's search query. Therefore, improving the accuracy of POI search is a topic that those skilled in the art must continuously research. Summary of the Invention

[0005] To solve the above technical problems, embodiments of the present application provide a map search method and device, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] One aspect of an embodiment of the present application provides a map search method, which is executed by an electronic device, and includes: obtaining a search statement, and obtaining a first spatial semantic vector corresponding to the search statement through a spatial semantic model; wherein the spatial semantic model is obtained by training prior map search data, and the prior map search data contains text information and spatial information; obtaining a location interest point vector matching the first spatial semantic vector from a location interest point vector library through a spatial semantic index library as a target location interest point vector corresponding to the first spatial semantic vector; wherein the location interest point vectors in the location interest point vector library are obtained through the spatial semantic model; and determining the target location interest point corresponding to the search statement based on the target location interest point vector.

[0007] Another aspect of an embodiment of the present application provides a map search device, which is deployed on an electronic device and includes: a vector acquisition module, configured to acquire a search statement and acquire a first spatial semantic vector corresponding to the search statement through a spatial semantic model; wherein the spatial semantic model is obtained by training prior map search data, and the prior map search data contains text information and spatial information; a vector recall module, configured to acquire a location interest point vector matching the first spatial semantic vector from a location interest point vector library through a spatial semantic index library as a target location interest point vector corresponding to the first spatial semantic vector; wherein the location interest point vectors in the location interest point vector library are acquired through the spatial semantic model; and a result acquisition module, configured to determine the target location interest point corresponding to the search statement based on the target location interest point vector.

[0008] Another aspect of an embodiment of the present application provides an electronic device, comprising: one or more processors; and a memory for storing one or more computer programs, wherein when the one or more computer programs are executed by the one or more processors, the electronic device implements the map search method as described above.

[0009] Another aspect of an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor of an electronic device, the electronic device executes the map search method as described above.

[0010] Another aspect of an embodiment of the present application provides a computer program product, including a computer program, which implements the map search method described above when executed by a processor.

[0011] In the technical solution provided in the embodiments of the present application, on the one hand, the first spatial semantic vector corresponding to the search statement is obtained through the spatial semantic model. Since the spatial semantic model is obtained by training the prior map search data containing text information and spatial information, the spatial semantic vector corresponding to the obtained search statement has better spatial characterization ability. The subsequent acquisition of the target location interest point based on the first spatial semantic vector corresponding to the search statement can help improve the search accuracy; on the other hand, the location interest point vector matching the first spatial semantic vector is obtained from the location interest point vector library through the spatial semantic search library to obtain the target location interest point vector of the first spatial semantic vector. Since the vector search method does not require a complete match like the text inverted index method, but only requires similarity in the vector space, it is more suitable for complex map searches, and the location interest point vector in the location interest point vector library is also obtained through the spatial semantic model, so that the target location interest point vector obtained in the embodiment of the present application has higher accuracy, which further helps to improve the search accuracy.

[0012] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] FIG1 is a schematic diagram of an exemplary implementation environment of the present application;

[0014] FIG2 is a flow chart of a map search method proposed by an exemplary embodiment of the present application;

[0015] FIG3 is a flow chart of a map search method proposed by another exemplary embodiment of the present application;

[0016] FIG4 is a schematic diagram of an exemplary process of obtaining a target location point of interest corresponding to a search statement;

[0017] FIG5 is a flowchart of a map search method proposed in another exemplary embodiment of the present application;

[0018] FIG6 is a schematic diagram of an exemplary training architecture of a spatial semantic model;

[0019] FIG7 is a flowchart of a map search method proposed in another exemplary embodiment of the present application;

[0020] FIG8 is an exemplary architecture diagram of a spatial semantic index library with limited search distance;

[0021] FIG9 is a schematic diagram illustrating an exemplary overall map search process;

[0022] FIG10 is a schematic diagram of an exemplary offline evaluation effect;

[0023] FIG11 is a schematic diagram of an exemplary online evaluation effect;

[0024] FIG12 is a block diagram of a map search device according to an exemplary embodiment of the present application;

[0025] FIG13 shows a schematic structural diagram of a computer system suitable for implementing an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different figures represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0027] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0028] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0029] In this application, "plurality" refers to two or more. "And / or" describes the relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.

[0030] The terms "first," "second," "third," and "fourth," etc., in the specification and claims of this application and the accompanying drawings are used to distinguish different objects, not to describe a specific order. The terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0031] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0032] Please refer to Figure 1, which is a schematic diagram of an exemplary implementation environment of the present application. The implementation environment is a map search system, including a terminal 110 and a server 120, and a wired or wireless communication connection is pre-established between the terminal 110 and the server 120.

[0033] Among them, the terminal 110 can be a smart phone, a computer, an intelligent voice interaction device, a smart home appliance, a vehicle terminal, an aircraft, etc., and this embodiment does not limit it. The terminal 110 is used to run a map application. It can be understood that the map application displays a user interface (UI) so that the user can enter a search statement in the user interface and obtain the target location point of interest through the user interface. For example, in the user interface 101 displayed by the terminal 110 in the example of Figure 1, when the user enters the search statement "Food in Olympic Park" in the search box and clicks the "Search" button, the user interface 102 is jumped to display. The user interface 102 displays map data of multiple restaurants near Olympic Park in City A, such as Restaurant 1 near Olympic Park in City A, Restaurant 2 near Olympic Park in City A, Restaurant 3 near Olympic Park in City A, Restaurant 4 near Olympic Park in City A, Restaurant 5 near Olympic Park in City A, etc. Furthermore, as an exemplary embodiment, a voice input entrance may also be provided in the user interface 101, so that after triggering the entrance, the user can input a search statement by voice, that is, the map application supports the function of performing map search based on the user's voice. This embodiment does not limit the specific method of how the user inputs the search statement.

[0034] The server 120 is used to provide data services for the map application running on the terminal 110. The server 120 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. This embodiment does not limit the specific product form of the server 120.

[0035] Exemplarily, a search engine is running on server 120. After receiving a search statement uploaded by terminal 110, the search engine first obtains a first spatial semantic vector corresponding to the search statement using a spatial semantic model. Then, using a spatial semantic index library, the search engine retrieves a location interest point vector that matches the first spatial semantic vector from a location interest point vector library as the target location interest point vector for the first spatial semantic vector. Finally, the target location interest point corresponding to the search statement is determined based on the obtained target location interest point vector. The target location interest point typically includes ranking information for multiple search location information. For example, after sorting multiple search location information from greatest to least according to their degree of matching with the search statement, the top K ranked search location information from the resulting ranking sequence are selected as the target location interest points. Server 120 sends the target location interest points to terminal 110, causing terminal 110 to display the target location interest points, thereby allowing the user to learn about the target location interest points corresponding to the search statement. It should be noted that the detailed process of how the search engine obtains the target location interest points corresponding to the search statement can be found in the subsequent embodiments and will not be elaborated on in this embodiment.

[0036] In a map POI search scenario, it is necessary to retrieve POIs related to the search query from a large number of POIs based on the search query input by the user. The implementation environment of the example in Figure 1 can also be implemented as a map POI search scenario. Specifically, the server 120 obtains target location points of interest corresponding to the search query, and each search location information corresponds to a POI.

[0037] To improve the accuracy of map POI searches, considering that map searches are sensitive to spatial information in addition to text, users may also require spoken searches, making search requirements more complex, and that vector searches do not require exact matches like text inverted indexes, but only require similarity in the vector space, thus offering a natural advantage in processing complex spoken or mistyped requests. Therefore, embodiments of the present application propose a map search method designed to obtain more accurate target locations and points of interest.

[0038] Please refer to Figure 2, which is a flowchart of a map search method proposed in an exemplary embodiment of the present application. This map search method can be applied to the implementation environment shown in Figure 1, and can be specifically executed by terminal 110 or server 120 in the implementation environment shown in Figure 1. Of course, this map search method can also be specifically executed by other terminals or servers that have map search application requirements, and this embodiment is not limited to this.

[0039] As shown in FIG2 , in an exemplary embodiment, the map search method includes S210 to S230 , which are described in detail as follows:

[0040] S210: Obtain a search statement, and obtain a first spatial semantic vector corresponding to the search statement through a spatial semantic model.

[0041] To search for a point of interest (POI), users can enter a search query on the map. A search query is a keyword or phrase that users enter to find one or more POIs (e.g., restaurants, hospitals, schools, etc.) on the map. The search query can include key information such as the name, type, and location of the POI they are searching for.

[0042] It should be noted that currently mainstream semantic vectors are typically trained based on large amounts of text corpora and can effectively characterize the semantic similarity between texts. However, in map search scenarios, due to the high sensitivity to spatial information, it is necessary to build a semantic model that can capture both textual and spatial information.

[0043] The present embodiment thus provides a spatial semantic model. Since the spatial semantic model is obtained by training with prior map search data, and the prior map search data contains text information as well as spatial information, the spatial semantic model has both text semantic characterization capabilities and spatial semantic characterization capabilities. Spatial information can be, for example, cities, longitude and latitude information, landmark locations, and other content that can describe spatially related information, which is not limited here. Therefore, the first spatial semantic vector corresponding to the search statement obtained by the spatial semantic model in the present embodiment contains not only text semantic information, but also spatial semantic information, so that in the subsequent recall process of the first spatial semantic vector, a more accurate recall result can be obtained based on the text semantic information and the spatial semantic information, which can then be used to improve the accuracy of the target location point of interest.

[0044] Compared with the semantic model provided by the related technology, the spatial semantic model provided by the embodiment of the present application adds a spatial feature extraction network and a feature fusion network, so that semantic information is extracted through the original text feature extraction network, spatial information is extracted through the newly added spatial feature extraction network, and the first spatial semantic vector is obtained by fusion through the feature fusion network, so that the first spatial semantic vector reflects both semantic information and spatial information.

[0045] It should also be noted that the search statement mentioned in this embodiment can be text directly entered by the user, such as text entered by the user in the search box included in the user interface, or text obtained by recognizing the voice input by the user, or text entered by other means, such as a search statement received from another terminal. This embodiment does not limit this.

[0046] In another exemplary embodiment, in order to further improve the spatial characterization accuracy of the search statement by the spatial semantic model, the search statement can also be parsed to obtain the search content and the first search position, and then the search content and the first search position are input into the spatial semantic model to obtain the first spatial semantic vector corresponding to the search statement output by the spatial semantic model. That is to say, it is understood that a semantic search of the search content is performed at the spatial position corresponding to the first search position. For example, if the search statement is "food in place B" and the user's current location is place A, by parsing the search statement, it can be obtained that the search content is "food" and the first search position is "place B". This can avoid the spatial semantic model from mistakenly characterizing the user's current location as the first search position when extracting the spatial statement vector of the search statement, thereby improving the semantic spatial search accuracy in a priori manner, thereby obtaining a more accurate first spatial semantic vector.

[0047] In addition, as an exemplary implementation, the first search position and the concatenated string of the search content and the first search position can also be input into the spatial semantic model, so that the subsequent execution of the vector recall stage based on the obtained first spatial semantic vector can avoid the under-recall caused by a single request, thereby further improving the recall accuracy.

[0048] S220 , obtaining a location interest point vector that matches the first spatial semantic vector from a location interest point vector library through a spatial semantic index library as a target location interest point vector corresponding to the first spatial semantic vector.

[0049] The location point of interest vector library is a vector library composed of all location point of interest vectors. Location point of interest vectors are vectors of location points of interest and are used to represent the characteristics of location points of interest. To determine the target location point of interest that the user wishes to search for, after obtaining the first spatial semantic vector, the first spatial semantic vector can be compared with the location point of interest vectors in the location point of interest vector library to determine the correlation between the search statement and different location points of interest, thereby obtaining the final target location point of interest.

[0050] When determining the target location point of interest vector, the correlation between the first spatial semantic vector and each location point of interest vector in the location point of interest vector library can be calculated. If the correlation between the location point of interest vector and the first spatial semantic vector meets certain conditions, the location point of interest vector is considered to match the first spatial semantic vector, and the location point of interest vector whose correlation meets certain conditions is then determined as the target location point of interest vector. In one possible implementation, the correlation between the location point of interest vector and the first spatial semantic vector meeting certain conditions may mean that the correlation between the location point of interest vector and the first spatial semantic vector reaches a preset threshold, or the correlation between each location point of interest vector and the first spatial semantic vector is sorted in descending order, and if the correlation is in the top N digits, it is determined that the certain condition is met.

[0051] In one possible implementation, a map search may include a recall phase and a sorting phase. The recall phase may involve obtaining a small number of target location POI vectors that match the first spatial semantic vector from a large number of location POI vectors. This means preliminarily screening a batch of target location POI vectors that may be relevant to the user query from the large number of location POI vectors, thereby narrowing the search scope in the subsequent sorting phase. The sorting phase may involve sorting the recalled target location POI vectors to select the target location POI.

[0052] In the above-mentioned map search scenario, this step involves the process acting on the recall stage. Specifically, a vector search method is used to perform vector recall on the first spatial semantic vector obtained in S210. This can adapt to complex spoken search requests or mistyped search requests, thereby improving the recall accuracy and thus improving the accuracy of the target location points of interest.

[0053] The spatial semantic index library mentioned in this embodiment is specifically used to recall POI vectors that match an input vector (e.g., the first spatial semantic vector) from the POI vector library. It should be noted that the POI vectors in the POI vector library are also obtained through the spatial semantic model, which ensures that the recalled vectors also contain both textual semantic information and spatial semantic information.

[0054] S230: Determine the target location interest point corresponding to the search statement according to the target location interest point vector.

[0055] This embodiment can achieve a higher accuracy of the target location interest point vector corresponding to the search statement obtained through the aforementioned steps, so the accuracy of determining the target location interest point corresponding to the search statement based on the target location interest point vector corresponding to the search statement is also correspondingly higher. It can be understood that as an exemplary implementation, the correlation between the first spatial semantic vector corresponding to the search statement and each target location interest point vector can be calculated, and then the multiple target location interest point vectors are sorted in descending order according to the correlation, and the topK target location interest point vectors are selected as the target vectors, and the POI information corresponding to each target vector is used as the final target location interest point. Among them, the topK rankings are understood to be the K target location interest point vectors with the highest correlation. The specific value of K is not restricted here, and can be set according to actual application requirements.

[0056] Therefore, in the technical solution provided in this embodiment, the first spatial semantic vector corresponding to the search statement is obtained through the spatial semantic model. Since the spatial semantic model is obtained by training the prior map search data containing text information and spatial information, the spatial semantic vector corresponding to the obtained search statement has better spatial characterization ability. Subsequently, obtaining the target location interest point based on the first spatial semantic vector corresponding to the search statement can help improve the search accuracy. On the other hand, the target location interest point vector of the first spatial semantic vector is obtained by obtaining the location interest point vector matching the first spatial semantic vector from the location interest point vector library through the spatial semantic search library. Since the vector search method does not require a complete match like the text inverted index method, but only requires similarity in the vector space, it is more suitable for complex map searches, and the location interest point vectors in the location interest point vector library are also obtained through the spatial semantic model, so that the target location interest point vector obtained in the embodiment of the present application has higher accuracy, which further helps to improve the search accuracy, thereby obtaining target location interest points with higher accuracy, so that users can get a better search experience through map applications.

[0057] In another exemplary embodiment, considering that the overall search effect can be improved by simultaneously building a text inverted index and a vector index, another map search method as shown in FIG3 is also provided.

[0058] As shown in FIG3 , the exemplary map search method further includes S310 to S320 based on the embodiment shown in FIG2 . On this basis, the specific method of S230 may be as shown in S330 , which is described in detail as follows:

[0059] S310: Convert the location points of interest in the location points of interest library into shard data through the text inverted index library, and obtain the shard data matching the search statement based on the shard data as the target shard data corresponding to the search statement.

[0060] In this embodiment, the text inverted index library is used to convert POIs in the POI library into shard data and obtain shard data that matches the input text based on the converted distribution data. It is understandable that the input text involved in this embodiment is also a search statement. In the text inverted index library, the corresponding longitude and latitude information of the POIs in the POI library is converted into shard data according to a preset sharding method, and the shard is used as the keyword (Key) and all POI data on the shard is used as the value (Value) to construct an inverted index. The text inverted index corresponding to the search statement is intersected to obtain the recall result, i.e., the target shard data.

[0061] It should be understood that the preset sharding methods include, for example, GeoHash (an algorithm for encoding longitude and latitude) sharding, Mercator coordinate conversion (a map projection coordinate system) sharding, etc. The text postings chain corresponding to the search statement can be a postings chain constructed based on the text semantic vector corresponding to the search statement. The text semantic vector corresponding to the search statement can be, for example, a BERT (Bidirectional Encoder Representations from Transformers) model or other similar models with a Transformer (encoder-decoder structure, a neural network model based on a self-attention mechanism that aims to capture global dependencies in an input sequence) architecture or an LSTM (Long Short-Term Memory, a time recursive neural network) architecture. This embodiment does not limit this.

[0062] S320: Obtain a second spatial semantic vector of the location point of interest corresponding to the target segment data through a spatial semantic model.

[0063] The target segmentation data obtained above is the result related to the search statement obtained through the text inverted index method. Therefore, the POI can be obtained accordingly based on the target segmentation data. This POI is also understood as the POI obtained through the text inverted index method, and the spatial semantic vector of this POI can be obtained accordingly through the spatial semantic model.

[0064] S330, calculating the similarity between the first spatial semantic vector and the target location interest point vector, and calculating the similarity between the first spatial semantic vector and the second spatial semantic vector, so as to determine the target location interest point corresponding to the search statement according to the calculated similarities.

[0065] This embodiment not only calculates the similarity between the first spatial semantic vector corresponding to the search statement and the target location point of interest vector obtained through the spatial semantic index library, but also calculates the similarity between the first spatial semantic vector corresponding to the search statement and the second spatial semantic vector corresponding to the target shard data, and then summarizes all the similarities and sorts them from large to small. Finally, the POI associated with the first spatial statement vector corresponding to the topK similarities is selected as the target location point of interest corresponding to the search statement.

[0066] The technical solution provided by this embodiment can also be expressed as a processing flow as shown in Figure 4. It can be seen that for the search statement, not only the target position interest point vector is obtained through the spatial semantic model and the spatial semantic index library, but also the text reverse sorting recall result is obtained. The correlation calculation refers to calculating the similarity between the spatial semantic vector corresponding to the search statement obtained through the spatial semantic model and the target position interest point vector and each vector contained in the text reverse sorting recall result, and finally the target position interest point corresponding to the search statement is output according to the similarity sorting.

[0067] Therefore, the target location interest points finally output by this embodiment are POI information with a high correlation with the search statement, which is jointly determined by building a text inverted index and a spatial semantic vector index. It can be understood that the correlation between the target location interest point vector obtained by the text inverted index and the search statement may be better than the correlation between the target location interest point vector obtained by the spatial semantic vector index and the search statement, so that the target location interest points finally output include the K POIs with the highest correlation with the search statement obtained by combining the two indexing methods, thereby improving the overall accuracy of POI search.

[0068] In another exemplary embodiment, the spatial semantic model includes a text feature extraction network, a spatial feature extraction network, and a feature fusion network, and the output signals of the text feature extraction network and the spatial feature extraction network are used as input signals for the feature fusion network. The text feature extraction network is used to extract text feature information from the input signal, the spatial feature extraction network is used to extract spatial feature information from the input signal, and the feature fusion network is used to perform feature fusion processing on the input text feature information and spatial feature information, and output the fused feature vector accordingly.

[0069] As shown in FIG5 , in an exemplary embodiment, the map search method further includes a process of training a spatial semantic model, specifically including S510 - S530 , which are described in detail as follows:

[0070] S510: Obtain a training sample set of the current batch.

[0071] First, let's introduce the important parameters involved in the machine learning model training process: batch and batch size. A batch represents one iteration, and each iteration updates the parameters of the model network structure. The batch size represents the number of training samples used in one iteration. Therefore, each time the model network structure parameters are adjusted, it is necessary to obtain the current batch of training samples to train the spatial semantic model.

[0072] In this embodiment, the training samples used to train the spatial semantic model are previous map search data. Each piece of previous map search data includes a previous search request and corresponding POI click information. In other words, the training samples come from real historical search and click data within the map application. For example, a piece of previous map search data primarily includes a previous search request (such as the search statement, the city of the second search location, the latitude and longitude information of the second search location, etc.) and POI click information (such as the POI name, POI address, POI latitude and longitude, POI tag, etc.).

[0073] S520, for each training sample, extract first text information and first spatial information from the previous search request, input the extracted first text information into the text feature extraction network, and input the extracted first spatial information into the spatial feature extraction network, and obtain the sentence feature vector output by the feature fusion network, and extract second text information and second spatial information from the location interest point click information, input the extracted second text information into the text feature extraction network, and input the extracted second spatial information into the spatial feature extraction network, and obtain the location interest point feature vector output by the feature fusion network.

[0074] Based on the network structure of the spatial semantic model, this embodiment requires, for each training sample, extracting first text information and first spatial information from the previous search request contained therein, and then inputting the extracted first text information into the text feature extraction network so that the text feature extraction network outputs text feature information accordingly, and inputting the extracted first spatial information into the spatial feature extraction network so that the spatial feature extraction network outputs spatial feature information accordingly, and the feature fusion network performs feature fusion processing on the input text feature information and spatial feature information, and outputs a sentence feature vector accordingly; and, extracting second text information and second spatial information from the POI click information contained therein, and then inputting the extracted second text information into the text feature extraction network so that the text feature extraction network outputs text feature information accordingly, and inputting the extracted second spatial information into the spatial feature extraction network so that the spatial feature extraction network outputs spatial feature information accordingly, and the feature fusion network performs feature fusion processing on the input text feature information and spatial feature information, and outputs a POI feature vector accordingly.

[0075] As an exemplary embodiment, the process of extracting first text information from a previous search request included in previous map search data may include: extracting a previous search statement from the previous search request and using the previous search statement as the first text information. The process of extracting first spatial information from a previous search request included in previous map search data may include: extracting at least one first spatial information from the previous search request, including the city where the second search location is located, the latitude and longitude information of the second search location, and landmarks related to the second search location. The process of extracting second text information from POI click information included in previous map search data may include: extracting at least one second text information from the description text and structural information of the historical POI from the POI click information. The process of extracting second spatial information from POI click information included in previous map search data may include: extracting at least one second spatial information from the POI click information, including the city where the historical POI is located, the latitude and longitude information of the historical POI, and landmarks related to the historical POI.

[0076] It should be noted that the text feature extraction network can include a feature extraction network layer and a pooling layer connected in sequence. The feature extraction network layer can adopt the BERT model or other models with similar Transformer architecture or LSTM architecture, without limitation. When the feature extraction network layer adopts the BERT model, the average output signal of the last layer can be used as the text semantic feature representation of the text information to obtain text semantic features containing richer information content.

[0077] The process of extracting landmarks related to the second search location from the previous search request or extracting landmarks related to the historical POI from the POI click information may include: obtaining a first landmark set located within the map range where the second search location is located based on the latitude and longitude information of the second search location, and obtaining a second landmark set located within the map range where the historical POI is located based on the latitude and longitude information of the historical POI, and then sorting the landmarks in the first landmark set in descending order of click popularity, and selecting landmarks with a specified ranking as landmarks related to the second search location, and sorting the landmarks in the second landmark set in descending order of click popularity, and selecting landmarks with a specified ranking as landmarks related to the historical POI. Among them, the selected landmarks with a specified ranking may be one or more, and this embodiment does not limit this. It can be understood that the source of the landmarks may be manually identified landmarks in the map data, such as business districts, office buildings, communities, etc. Landmark data also corresponds to longitude and latitude information. Therefore, the top N most popular landmarks at the current location (such as the target search location or POI location) can be obtained through longitude and latitude information and user click data. The obtained landmark information is vectorized as an input feature, thereby establishing a coarse-to-fine spatial description method based on the three maintenances of city, longitude and latitude, and landmarks.

[0078] In other exemplary embodiments, landmarks related to the second search location or landmarks related to the POI may also be predicted using a pre-trained model, which is not limited in this embodiment.

[0079] In addition, since the longitude and latitude information are continuous variables, this embodiment also has a special design for the vectorization of the longitude and latitude information. For example, the spatial feature extraction network needs to discretize the input longitude and latitude information at different scales, and then extract spatial information features from the obtained discretized information. Multi-scale discretization can be generally understood as discretizing and encoding the longitude and latitude information according to different regional division scales. Thus, the same longitude and latitude information can obtain multi-scale discretized information, and the spatial features corresponding to the longitude and latitude information are extracted based on the multi-scale discretized information, which can avoid the problem of effective boundary coverage and insufficient accuracy caused by a single scale. Among them, GeoHash and other methods can be used to achieve the discretization of longitude and latitude information.

[0080] POI descriptions include, for example, the POI's name, alias, and address. Search queries often include the name or address, so this textual information is represented separately in the spatial semantic model. Furthermore, structural information about the POI, such as its tag, type, and brand, is also represented as a feature, effectively capturing correlations.

[0081] The feature fusion network contains a self-attention network layer, which fuses the input multi-channel feature vector representations (including text vector representations and spatial vector representations) through the self-attention network layer. During the training process, the spatial semantic model can effectively learn the weight information of each feature vector representation in the final model through the distribution of POI click data, thereby achieving better feature fusion effects.

[0082] S530 , calculating corresponding training loss values ​​according to the sentence feature vectors and location interest point feature vectors corresponding to each training sample, and adjusting parameters of the spatial semantic model based on the training loss values.

[0083] After the feature fusion network performs feature fusion processing on each feature representation according to the weights, the sentence feature vector corresponding to the previous search request and the POI feature vector corresponding to the POI click information corresponding to the previous search request can be obtained. By introducing the contrast loss function, the corresponding training loss value is calculated according to the sentence feature vector and location point of interest feature vector corresponding to each training sample, and the parameters of the spatial semantic model are adjusted based on the training loss value, finally achieving iterative tuning of the spatial semantic model.

[0084] As an exemplary embodiment, the process of calculating corresponding training loss values ​​according to the sentence feature vectors and location interest point feature vectors corresponding to each training sample and adjusting parameters of the spatial semantic model based on the training loss values ​​may include the following steps S531-S532:

[0085] S531, taking the training sample containing the POI click information as a positive sample, and extracting multiple random negative samples and multiple hard negative samples corresponding to the positive sample from the training sample set of the current batch;

[0086] S532, respectively calculating the similarities between the sentence feature vectors and the POI feature vectors corresponding to the positive sample, each random negative sample, and each difficult negative sample, and determining the corresponding training loss value based on the calculated similarities.

[0087] In the above process, the positive sample is the training sample containing POI click information, that is, the training sample that actually generates POI click behavior. Extracting multiple random negative samples corresponding to the positive sample from the current batch of training sample sets refers to randomly selecting multiple training samples from the training samples other than the positive sample in the current batch of training sample sets. Extracting multiple difficult negative samples corresponding to the positive sample from the current batch of training sample sets refers to calculating the similarity between each training sample other than the positive sample in the current batch of training sample sets and the positive sample, and selecting multiple training samples with the highest similarity.

[0088] By calculating the similarity between the sentence feature vectors and the POI feature vectors corresponding to the positive samples, each random negative sample, and each hard negative sample, and determining the corresponding training loss value based on the calculated similarity, the spatial semantic model can be overall adjusted based on the determined training loss value.

[0089] An exemplary contrast loss function can be expressed as follows:

[0090] Among them, q i represents the sentence feature vector corresponding to the previous search request, p i represents the POI feature vector corresponding to the POI click information corresponding to the previous search request, cos(q i ,p i ) represents the cosine similarity between the sentence feature vector and the POI feature vector, S represents the number of random negative samples, H represents the number of hard negative samples, τ represents an adjustable parameter, and loss i Represents the training loss value.

[0091] It should also be noted that this embodiment can also use other contrast loss functions to calculate the training loss value, such as the Margin Rank Loss function, etc., and the specific form of the loss function is not limited here.

[0092] Figure 6 is a schematic diagram of an exemplary training architecture of a spatial semantic model. As can be seen from Figure 6, this embodiment further proposes a method of spatial semantic fusion modeling. By introducing the city, longitude and latitude information, and high-quality landmarks near the longitude and latitude requested by the user, as well as the city of the POI clicked by the user, the longitude and latitude of the POI, and high-quality landmarks near the POI and other spatial information into the training process of the spatial semantic model, this information is fused with the requested text and the structural information of the POI for modeling, thereby improving the spatial characterization capability from the model level and obtaining a spatial semantic model suitable for map search scenarios.

[0093] In another exemplary embodiment, considering that in a map search scenario, users have a strong demand for close distances, especially in general-intent requests, such as food, hotels, etc., if there are distant results, it will bring a bad experience to users. And in a map search scenario, due to the huge amount of POI data, usually close to 100 million, a simple similarity matching calculation does not meet the application requirements. To solve these problems, the embodiment of the present application specifically constructs the spatial semantic index library as a spatial semantic index library with a limited search distance. Specifically, based on a preset search distance limitation condition, a spatial semantic index library with a limited search distance is constructed, so as to obtain the target location point of interest vector corresponding to the spatial semantic vector through the constructed spatial semantic index library with a limited search distance.

[0094] As shown in FIG7 , in an exemplary embodiment, based on a preset search distance limiting condition, the process of constructing a spatial semantic index library with a limited search distance includes steps S710 to S730 , which are described in detail as follows:

[0095] S710: Construct an inverted index for the location interest point vectors in the location interest point vector library in a clustering manner.

[0096] This embodiment takes into account the performance accuracy during search requests and the size of the index file, and adopts the parameter mode of IVFPQ to construct a spatial semantic index library with a limited search distance. It can be understood that IVFPQ is an indexing method for data retrieval that combines the technology of inverted file (IVF) and product quantization (PQ). The inverted index is a data structure used to speed up searches. For each feature vector, the inverted index stores a list of data containing the feature vector, which makes it possible to quickly locate data containing similar features when querying. Product quantization is a dimensionality reduction and quantization technology. In data retrieval, very high-dimensional feature vectors are usually used to describe data. Product quantization reduces the complexity of storage and calculation by decomposing these high-dimensional vectors into smaller sub-vectors and independently quantizing each sub-vector, which helps to speed up retrieval.

[0097] In the inverted index stage, an inverted index will be constructed for the POI vectors in the POI vector library in a clustering manner. When performing vector search, a target number of POI vectors that are most similar to the input vector will be searched from the POI vector library.

[0098] S720 , performing product quantization indexing after inverted indexing based on the search distance restriction condition.

[0099] The product quantization process is used to accelerate the calculation by bit quantization after obtaining the target number of inverted POI vectors through the inverted index.

[0100] Conventional product quantization processing cannot introduce a distance metric. To achieve a search distance limit, it is necessary to recall N results in reverse order, traverse and calculate the distance of all results, and then retain the results with the limited distance. However, this approach has some disadvantages. For example, although the set of N recalled results is large, there is no guarantee that all close results are in the set. Moreover, the larger the set of N recalled results, the more time-consuming it is, and the more time-consuming it is to calculate the distance a posteriori. To solve this problem, this embodiment introduces a distance metric in the product quantization processing stage, approximating and optimizing the distance limitation problem in a more optimal and faster way.

[0101] Furthermore, in some exemplary embodiments, the search distance constraint is typically pre-set, but may also be customizable. For example, upon receiving an instruction to update the search distance constraint, the search distance constraint is updated in response to the instruction, and the spatial semantic index library that limits the search distance is updated based on the updated search distance constraint. The updated spatial semantic index library then updates the proximity condition used in the vector search process accordingly.

[0102] S730: A spatial semantic index library with a limited search distance is constructed by using an inverted index and a product quantization index.

[0103] In this embodiment, the spatial semantic index library that limits the search distance is understood as introducing pre-calculated spatial information during the product quantization stage. Compared to an approach that does not introduce pre-calculated spatial information during the product quantization stage, this embodiment can perform distance calculations on a larger data scale.

[0104] Please refer to Figure 8, which is an exemplary architecture diagram of a spatial semantic index library with limited search distance. It can be seen that in the process of providing vector index search by the spatial semantic index library, by introducing distance measurement in the product quantization index stage, that is, the search distance limitation condition, the result distance can be accurately limited in the process of vector approximation search. Compared with the existing method of performing distance filtering after recalling all results, it can improve the accuracy of close-range recall.

[0105] To facilitate a holistic understanding of the technical solutions provided by this application, Figure 9 illustrates an exemplary overall map search process. Specifically, the map search process is divided into an online portion and an offline portion. The online portion primarily implements the process from obtaining a search statement to outputting the corresponding ranking results, while the offline portion primarily implements the construction of a spatial semantic model and a spatial semantic index library.

[0106] As shown in Figure 9, the online part mainly proposes a technical means of sentence parsing + semantic search. By performing "wherewhat" parsing on the search sentence requested by the user, the search content (what part) and the first search position (where part) are obtained. By performing a semantic search for the what part at the spatial position corresponding to the where part, the accuracy of the semantic spatial search is improved in an a priori way, and the first search position that the user wants to search can be effectively obtained.

[0107] In the offline phase, a spatial semantic fusion modeling technique is proposed. By integrating spatial information such as the city, the user's requested latitude and longitude, and high-quality landmarks near the requested location with text and POI structured information during spatial semantic model training, the model learns the spatial relationship between requests and results, improving the model's spatial representation capabilities. As a result, the online spatial semantic model can obtain spatial semantic vectors that depict richer spatial information, ensuring the accuracy of subsequent spatial searches and effectively facilitating the acquisition of the user's desired target location.

[0108] The offline part also proposes a technical means of semantic search with a specified search range. By introducing a distance metric in the process of vector index search and accurately limiting the result distance in the process of vector approximation search, semantic results within the limited distance can be accurately recalled.

[0109] It is also necessary to understand that the spatial semantic model shown in FIG9 is connected to the spatial semantic vector and the POI vector library respectively, which is used to indicate that the spatial semantic vector is obtained through the spatial semantic model, and the POI vectors in the POI vector library are also obtained through the spatial semantic model. The full amount of POI data is connected to the spatial semantic index library, which is used to indicate that the spatial semantic index library needs to use the POI data in the full amount of POI data when performing a recall search for the spatial semantic vector. Similarly, the POI vector library is connected to the correlation calculation module to indicate that the POI vectors in the POI vector library need to be used when performing correlation calculations. The spatial semantic vector is connected to the correlation calculation module to indicate that the correlation calculation is a correlation calculation performed on the spatial semantic model and the target location point of interest vector and the vector in the text reverse sorting recall result. The detailed processing procedures have been described in detail in the aforementioned embodiments and will not be repeated here.

[0110] Since each of the above-mentioned technical means can be used to improve the accuracy of map search, the integration of the above three technical means in the overall map search process can largely achieve effective modeling of spatial semantics and recall of close-range semantic results, which can meet the spatial information needs of map search scenarios.

[0111] The following also provides the results of the performance evaluation of the overall map search process shown in Figure 9. The performance evaluation mainly includes offline and online aspects.

[0112] The primary purpose of offline performance evaluation is to assess the optimization results before going online. This is done to determine whether the expected results have been achieved and whether the requirements for going online have been met. It is also used to identify deficiencies in existing optimization solutions and to iterate on them. It should be noted that the optimized solution refers to the overall map search process shown in Figure 9, while the unoptimized solution refers to a map search solution that does not utilize the three technical approaches proposed in this application.

[0113] The offline evaluation mainly uses the win rate evaluation method. During each evaluation, 5,000 random search statements are randomly selected from the online version to compare the results of the optimized version before and after, and the winning and losing of the top 10 results before and after are evaluated. The win rate is calculated as follows:

[0114] The offline evaluation results are shown in Figure 10. As shown in Figure 10, after using the optimized solution, both recall and ranking relevance have been significantly improved.

[0115] In terms of online performance evaluation, considering that conversion rate is the most important objective indicator for map search online and can intuitively reflect the optimization effects of various search-related aspects, an A / B performance evaluation of conversion rate was conducted on the map suggestion side. The conversion rate calculation formula is as follows:

[0116] It should be understood that in the map field, suggestion can be understood as a specific interface for implementing search suggestions, which can display suggested search items in the form of a drop-down list in the user interface. For example, when you enter "food", some restaurant information searched will be displayed in the drop-down list of the user interface. AB testing can be understood as a testing method in product, marketing, or website design, also known as control testing. This testing method compares two or more versions of a product to determine which version is better or more effective.

[0117] When the optimization results are released online, an experimental bucket is selected for the new version's effectiveness test, and a control bucket is selected for the old version's effectiveness evaluation. The difference between the two is used as the final A / B conversion rate performance. The results of the two optimization schemes on A / B conversion rates are shown in Figure 11. As can be seen from Figure 11, after adopting the optimized scheme, the recall and ranking relevance rates both increased, and the overall full-link optimization resulted in a 0.55% increase in conversion rate.

[0118] It can be proved from the above that this application can effectively integrate the relationship between space and semantics through modeling, and can realize the search of semantic results close to the specified space. In the recall stage of the map search scenario, for complex requests initiated by users, mistyped requests or spoken language requests, semantic recognition can be effectively performed, and close results of the user-specified space can be returned. In the sorting stage of the search scenario, by calculating the correlation between the semantics of the search statement and the semantics of the POI, the effect of the sorting model can be significantly improved.

[0119] Figure 12 is a block diagram of a map search device shown in an exemplary embodiment of the present application. The map search device can be configured on the terminal 110 or server 120 in the implementation environment shown in Figure 1, or on other terminals or servers. This embodiment does not limit this.

[0120] As shown in FIG12 , the exemplary map search device includes:

[0121] A vector acquisition module 1010 is configured to acquire a search statement and obtain a first spatial semantic vector corresponding to the search statement using a spatial semantic model; wherein the spatial semantic model is trained using prior map search data, and the prior map search data includes text information and spatial information;

[0122] The vector recall module 1020 is configured to obtain, from a location interest point vector library using a spatial semantic index library, a location interest point vector that matches the first spatial semantic vector as a target location interest point vector corresponding to the first spatial semantic vector; wherein the location interest point vector in the location interest point vector library is obtained using a spatial semantic model;

[0123] The result acquisition module 1030 is configured to determine the target location interest point corresponding to the search statement according to the target location interest point vector.

[0124] In another exemplary embodiment, the vector acquisition module 1010 is further configured to:

[0125] Parse the search statement to obtain the search content and the first search position;

[0126] The search content and the first search position are input into the spatial semantic model to obtain a first spatial semantic vector corresponding to the search statement output by the spatial semantic model.

[0127] In another exemplary embodiment, the exemplary map search device further includes:

[0128] an inverted index module configured to convert the location points of interest in the location points of interest library into shard data through the text inverted index library, obtain the shard data matching the search statement as the target shard data corresponding to the search statement based on the shard data, and obtain the second spatial semantic vector of the location point of interest corresponding to the target shard data through the spatial semantic model;

[0129] The correlation calculation module is configured to calculate the similarity between the first spatial semantic vector and the target location interest point vector, and to calculate the similarity between the first spatial semantic vector and the second spatial semantic vector, so as to determine the target location interest point corresponding to the search statement based on the calculated similarity.

[0130] In another exemplary embodiment, the spatial semantic model includes a text feature extraction network, a spatial feature extraction network, and a feature fusion network, and the output signals of the text feature extraction network and the spatial feature extraction network are both used as input signals of the feature fusion network; the exemplary map search device also includes:

[0131] A training sample acquisition module is configured to acquire a training sample set of a current batch; wherein each previous map search data is used as a training sample, and the previous map search data includes a previous search request and corresponding location point of interest click information;

[0132] a training sample processing module configured to, for each training sample, extract first text information and first spatial information from a previous search request, input the extracted first text information into a text feature extraction network, input the extracted first spatial information into a spatial feature extraction network, and obtain a sentence feature vector output by a feature fusion network; and extract second text information and second spatial information from location point of interest click information, input the extracted second text information into the text feature extraction network, input the extracted second spatial information into the spatial feature extraction network, and obtain a location point of interest feature vector output by the feature fusion network;

[0133] The loss adjustment module is configured to calculate the corresponding training loss value according to the sentence feature vector and the location interest point feature vector corresponding to each training sample, and adjust the parameters of the spatial semantic model based on the training loss value.

[0134] In another exemplary embodiment, the loss adjustment module is further configured to:

[0135] The training samples containing the location point of interest click information are used as positive samples, and multiple random negative samples and multiple hard negative samples corresponding to the positive samples are extracted from the current batch of training sample sets;

[0136] The similarities between the sentence feature vectors and the location interest point feature vectors corresponding to the positive samples, each random negative sample, and each difficult negative sample are calculated respectively, and the corresponding training loss value is determined based on the calculated similarities.

[0137] In another exemplary embodiment, the multiple hard negative samples are multiple training samples with the highest similarity selected based on the similarity between the positive samples and the other training samples except the positive samples in the current batch of training samples.

[0138] In another exemplary embodiment, the previous search request includes a second search location in the previous search request and a landmark related to the second search location in the previous search request; and the training sample processing module is further configured to:

[0139] A previous search statement is extracted from the previous search request and used as the first text information, and at least one first spatial information is extracted, including the city where the second search location is located, the latitude and longitude information of the second search location, and landmarks related to the second search location.

[0140] In another exemplary embodiment, the location point of interest click information includes a previously clicked location point of interest and a landmark related to the previously clicked location point of interest; the training sample processing module is further configured to:

[0141] At least one second text information is extracted from the location point of interest click information, including the description text and structural information of the previous location point of interest, and at least one second spatial information is extracted, including the city where the previous location point of interest is located, the latitude and longitude information of the previous location point of interest, and a landmark related to the previous location point of interest.

[0142] In another exemplary embodiment, the training sample processing module is further configured to: discretize the input latitude and longitude information at different scales in the spatial feature extraction network, and then extract spatial information features from the obtained discretized information.

[0143] In another exemplary embodiment, the training sample processing module is further configured to:

[0144] Obtaining, based on the latitude and longitude information of the second search location, a first set of landmarks within the map range where the second search location is located, and obtaining, based on the latitude and longitude information of the prior location point of interest, a second set of landmarks within the map range where the prior location point of interest is located;

[0145] The landmarks in the first landmark set are sorted in descending order of click popularity, and the landmarks with the specified ranking are selected as the landmarks related to the second search location; and the landmarks in the second landmark set are sorted in descending order of click popularity, and the landmarks with the specified ranking are selected as the landmarks related to the prior location point of interest.

[0146] In another exemplary embodiment, the spatial semantic index library is a spatial semantic index library that limits the search distance; the exemplary map search device further includes:

[0147] The index library construction module is configured to construct a spatial semantic index library with limited search distance based on a preset search distance limitation condition.

[0148] In another exemplary embodiment, the index library construction module is further configured to:

[0149] Constructing an inverted index for the location interest point vectors in the location interest point vector library in a clustering manner;

[0150] Constructing a product quantized index after the inverted index based on the search distance constraint;

[0151] The spatial semantic index library with limited search distance is composed of inverted index and product quantization index.

[0152] In another exemplary embodiment, the index library construction module is further configured to:

[0153] When an instruction to update the search distance limiting condition is obtained, an update process of the search distance limiting condition is executed in response to the instruction;

[0154] The spatial semantic index library with limited search distance is updated based on the updated search distance limitation condition.

[0155] It should be noted that the map search device provided in the above embodiment and the map search method provided in the above embodiment are based on the same concept. The specific manner in which the various modules and units perform their operations has been described in detail in the method embodiments and will not be repeated here. In actual applications, the map search device provided in the above embodiment can, as needed, allocate the aforementioned functions to different functional modules, i.e., divide the internal structure of the device into different functional modules to perform all or part of the functions described above. This is not a limitation herein.

[0156] An embodiment of the present application also provides an electronic device, comprising: one or more processors; and a memory for storing one or more computer programs. When the one or more computer programs are executed by the one or more processors, the electronic device implements the map search method provided in the above-mentioned embodiments.

[0157] Figure 13 shows a schematic diagram of the structure of a computer system of an electronic device suitable for implementing an embodiment of the present application. It should be noted that the computer system 1100 of the electronic device shown in Figure 13 is only an example and should not bring any limitation to the functions and scope of use of the embodiment of the present application.

[0158] As shown in Figure 13, the computer system 1100 includes a central processing unit (CPU) 1101, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1102 or the program loaded from the storage part 1108 into the random access memory (RAM) 1103, such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM 1103. The CPU 1101, ROM 1102 and RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0159] The following components are connected to the I / O interface 1105: an input section 1106 including a keyboard, a mouse, and the like; an output section 1107 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 1108 including a hard disk; and a communication section 1109 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the I / O interface 1105 as needed. Removable media 1111, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 1110 as needed, so that computer programs read from the removable media can be installed in the storage section 1108 as needed.

[0160] 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 computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1109, and / or installed from a removable medium 1111. When the computer program is executed by the central processing unit (CPU) 1101, the various functions defined in the system of the present application are executed.

[0161] It should be noted that the computer-readable medium shown in the embodiments of the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. More specific examples of computer-readable storage media may 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), a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. The computer program contained in the computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0162] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. Among them, each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from 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 or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0163] The units involved in the embodiments described in this application may be implemented by software or hardware, and the units described may also be set in a processor. In some cases, the names of these units do not constitute limitations on the units themselves.

[0164] Another aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the map search method described above. The computer-readable storage medium may be included in the electronic device described in the above embodiments, or may exist independently and not be incorporated into the electronic device.

[0165] Another aspect of the present application provides a computer program product, comprising a computer program stored in a computer-readable storage medium. A processor of an electronic device reads the computer program from the computer-readable storage medium and executes the computer program, causing the electronic device to perform the map search method provided in each of the above embodiments.

[0166] The above content is only a preferred exemplary embodiment of the present application and is not intended to limit the implementation scheme of the present application. Ordinary technicians in this field can easily make corresponding changes or modifications based on the main ideas and spirit of the present application. Therefore, the scope of protection of the present application shall be based on the scope of protection required by the claims.

[0167] It is understandable that in the specific implementation of this application, related data such as search statements and POIs are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

Claims

1. A map search method, performed by an electronic device, comprising: Obtaining a search statement and obtaining a first spatial semantic vector corresponding to the search statement through a spatial semantic model; wherein the spatial semantic model is trained by prior map search data, and the prior map search data includes text information and spatial information; Obtaining, from a location interest point vector library through a spatial semantic index library, a location interest point vector that matches the first spatial semantic vector as a target location interest point vector corresponding to the first spatial semantic vector; wherein the location interest point vector in the location interest point vector library is obtained through the spatial semantic model; Determine the target location interest point corresponding to the search statement according to the target location interest point vector.

2. The method according to claim 1, wherein obtaining the first spatial semantic vector corresponding to the search statement through the spatial semantic model comprises: Parsing the search statement to obtain search content and a first search position; The search content and the first search position are input into the spatial semantic model to obtain a first spatial semantic vector corresponding to the search statement output by the spatial semantic model.

3. The method according to claim 1 or 2, further comprising: Converting the location points of interest in the location points of interest library into shard data through a text inverted index library, and acquiring, based on the shard data, shard data that matches the search statement as target shard data corresponding to the search statement; Acquire a second spatial semantic vector of the location point of interest corresponding to the target fragment data through the spatial semantic model; The determining, based on the target location point of interest vector, the target location point of interest corresponding to the search statement includes: Calculate the similarity between the first spatial semantic vector and the target location interest point vector, and calculate the similarity between the first spatial semantic vector and the second spatial semantic vector, so as to determine the target location interest point corresponding to the search statement based on the calculated similarities.

4. The method according to any one of claims 1 to 3, wherein the spatial semantic model comprises a text feature extraction network, a spatial feature extraction network, and a feature fusion network, wherein output signals of the text feature extraction network and the spatial feature extraction network serve as input signals of the feature fusion network; and the method further comprises the following step of training the spatial semantic model: Get the training sample set of the current batch; where, Using each previous map search data as a training sample, the previous map search data including a previous search request and corresponding location point of interest click information; For each training sample, extracting first text information and first spatial information from the previous search request, inputting the extracted first text information into the text feature extraction network, inputting the extracted first spatial information into the spatial feature extraction network, and obtaining a sentence feature vector output by the feature fusion network; and extracting second text information and second spatial information from the location point of interest click information, inputting the extracted second text information into the text feature extraction network, inputting the extracted second spatial information into the spatial feature extraction network, and obtaining a location point of interest feature vector output by the feature fusion network; A corresponding training loss value is calculated according to the sentence feature vector and the location interest point feature vector corresponding to each training sample, and parameters of the spatial semantic model are adjusted based on the training loss value.

5. The method according to claim 4, wherein the calculating corresponding training loss values ​​according to the sentence feature vectors and the location point of interest feature vectors corresponding to each training sample, and adjusting parameters of the spatial semantic model based on the training loss values, comprises: Taking the training sample containing the click information of the location point of interest as a positive sample, and extracting a plurality of random negative samples and a plurality of hard negative samples corresponding to the positive sample from the training sample set of the current batch; The similarities between the sentence feature vectors and the location interest point feature vectors corresponding to the positive sample, each of the random negative samples, and each of the difficult negative samples are respectively calculated, and the corresponding training loss value is determined based on the calculated similarities.

6. The method according to claim 5, wherein the multiple hard negative samples are selected based on the similarity between other training samples in the current batch of training samples except the positive sample and the positive sample, and the multiple training samples with the highest similarity are selected.

7. The method according to any one of claims 4 to 6, wherein the previous search request includes a second search location of the previous search request and a landmark related to the second search location of the previous search request; and extracting the first text information and the first spatial information from the previous search request comprises: A previous search statement is extracted from the previous search request, and the previous search statement is used as the first text information, and at least one of the first spatial information including the city where the second search location is located, the latitude and longitude information of the second search location, and landmarks related to the second search location is extracted.

8. The method according to any one of claims 4 to 7, wherein the location point of interest click information includes a previously clicked location point of interest and a landmark related to the previously clicked location point of interest; and extracting the second text information and the second spatial information from the location point of interest click information comprises: Extract at least one second text information from the descriptive text and structural information of the previous location point of interest from the location point of interest click information, and extract at least one second spatial information from the city where the previous location point of interest is located, the latitude and longitude information of the previous location point of interest, and a landmark related to the previous location point of interest.

9. The method according to claim 7 or 8, wherein the spatial feature extraction network is used to discretize the input longitude and latitude information at different scales and then extract spatial information features from the obtained discretized information.

10. The method according to claim 7 or 8, wherein the process of extracting landmarks related to the second search location or the previous location point of interest comprises: Acquire a first set of landmarks within a map range where the second search location is located based on the latitude and longitude information of the second search location, and acquire a second set of landmarks within a map range where the prior point of interest is located based on the latitude and longitude information of the prior point of interest; The landmarks in the first landmark set are sorted in descending order of click popularity, and the landmarks with a specified ranking are selected as the landmarks related to the second search location, and the landmarks in the second landmark set are sorted in descending order of click popularity, and the landmarks with a specified ranking are selected as the landmarks related to the prior location point of interest.

11. The method according to any one of claims 1 to 10, wherein the spatial semantic index library is a spatial semantic index library with a limited search distance; the method further comprising: Based on a preset search distance limiting condition, a spatial semantic index library with limited search distance is constructed.

12. The method according to claim 11, wherein the step of constructing the spatial semantic index library with the limited search distance based on the preset search distance limiting condition comprises: Constructing an inverted index for the location interest point vectors in the location interest point vector library in a clustering manner; Constructing a product quantization index after the inverted index based on the search distance restriction condition; The spatial semantic index library for limiting the search distance is formed by the inverted index and the product quantization index.

13. The method according to claim 11, further comprising: When an instruction to update the search distance limiting condition is obtained, an update process of the search distance limiting condition is executed in response to the instruction; The spatial semantic index library with limited search distance is updated based on the updated search distance limiting condition.

14. A map search device, the device being deployed on an electronic device, the device comprising: a vector acquisition module configured to acquire a search statement and acquire a first spatial semantic vector corresponding to the search statement using a spatial semantic model; wherein the spatial semantic model is trained using prior map search data, the prior map search data including text information and spatial information; a vector recall module configured to obtain, from a location interest point vector library through a spatial semantic index library, a location interest point vector that matches the first spatial semantic vector as a target location interest point vector corresponding to the first spatial semantic vector; wherein the location interest point vectors in the location interest point vector library are obtained through the spatial semantic model; The result acquisition module is configured to determine the target location interest point corresponding to the search statement according to the target location interest point vector.

15. An electronic device comprising: one or more processors; A memory for storing one or more computer programs, which, when executed by the one or more processors, enables the electronic device to implement the map search method according to any one of claims 1 to 13.

16. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor of an electronic device, causes the electronic device to execute the map search method according to any one of claims 1 to 13.

17. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the map search method according to any one of claims 1 to 13 is implemented.

Images