Search intent determination and model training method and apparatus, and storage medium

Abstract

This disclosure provides a method, apparatus, and storage medium for determining search intent and training a model. The method includes: acquiring current scene features and individual historical search sequences of the navigated object; acquiring a group historical search sequence based on the start and end points of the target navigation route; inputting the individual historical search sequences into a self-attention mechanism processing module included in the model to obtain the preference features of the navigated object; inputting the preference features, current scene features, and individual historical search sequences into a first target attention mechanism processing module included in the model to obtain a first recommendation feature; inputting the preference features, current scene features, and group historical search sequences into a second target attention mechanism processing module included in the model to obtain a second recommendation feature; and inputting the first and second recommendation features into an intent prediction module included in the model to obtain the search intent of the navigated object to search along the target navigation route.

Description

Technical Field

[0001] This disclosure relates to the field of machine learning technology, specifically to a method, apparatus, and storage medium for determining search intent and training a model. Background Technology

[0002] Because map navigation services can provide users with efficient, low-cost travel routes and accurate route guidance, more and more applications are integrating map navigation-related service capabilities, such as travel applications, ride-hailing applications, and lifestyle service applications.

[0003] When a user plans to travel along the route provided by the map navigation service or has already traveled along the aforementioned route, the application software generally also supports the user to search for points of interest such as gas stations, charging stations, restaurants, subway stations, and rest areas along the travel route. This function can be called searching along the way or searching along the route.

[0004] Since complex search operations are performed by users while traveling along their routes, there are certain safety risks, especially when users are driving along the routes. Therefore, in order to enable users to use the "search along the way" function more safely and conveniently, existing technologies predict users' search intentions and proactively provide users with search keywords corresponding to those intentions, thereby reducing the complexity of user operations.

[0005] Therefore, accurately predicting users' search intent is a problem that those skilled in the art must solve during the implementation of the "search along the way" or "search along the route" function. Summary of the Invention

[0006] This disclosure provides a method, apparatus, and storage medium for determining search intent and training a model.

[0007] In a first aspect, embodiments of this disclosure provide a method for determining search intent, wherein the method is applied during the process of a navigated object traveling along a target navigation route, and the search intent is determined by a trained search intent prediction model, including: Obtain the current scene features and the pre-recorded individual historical search sequences generated by the navigated object during the historical navigation process; Based on the starting point and ending point of the target navigation route, a group historical search sequence is obtained from the historical search records corresponding to the historical navigation records of the historically navigated object; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical scene features and historical search tags; The individual's historical search sequence is input into the self-attention mechanism processing module included in the model to obtain the preference features of the navigated object; The preference features, the current scene features, and the individual's historical search sequence are input into the first target attention mechanism processing module included in the model to obtain the first recommendation features; The preference features, the current scene features, and the group's historical search sequence are input into the second target attention mechanism processing module included in the model to obtain the second recommendation features; The first recommendation feature and the second recommendation feature are input into the intent prediction module included in the model to obtain the search intent of the navigated object to search along the target navigation route.

[0008] Secondly, embodiments of the present invention provide a model training method, comprising: Obtain the target historical scene features and target historical search tag features corresponding to the target historical navigation route of the navigated object, as well as the individual historical search sequence generated by the navigated object during the historical navigation process before the target historical navigation route; Based on the starting point and ending point of the target's historical navigation route, obtain the group's historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object; The individual historical search sequence and the group historical search sequence of the navigated object are used as training data and input into the search intent prediction model; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical search label and historical scene feature; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The model includes a first target attention mechanism processing module that obtains a first recommendation feature based on the preference features, the target historical scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module that obtains a second recommendation feature based on the preference features, the target historical scene features, and the group's historical search sequence; The model includes an intent prediction module that obtains the search intent corresponding to the target historical scene features based on the first recommendation feature and the second recommendation feature. The loss function adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module simultaneously based on the search intent output by the intent prediction module and the target historical search tag features used as sample tags.

[0009] Thirdly, embodiments of the present invention provide a search intent determination device, used to determine the search intent of a navigated object during its journey along a target navigation route using a trained search intent prediction model, including: The individual feature acquisition module is configured to acquire the current scene features and the pre-recorded individual historical search sequence generated by the navigated object during the historical navigation process; The group feature acquisition module is configured to acquire a group historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object, based on the start and end points of the target navigation route; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical scene features and historical search tags; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the input individual's historical search sequence. The model includes a first target attention mechanism processing module, which obtains a first recommendation feature based on the input preference features, the current scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module, which obtains a second recommendation feature based on the input preference features, the current scene features, and the group's historical search sequence; The model includes an intent prediction module that, based on the input first recommendation features and second recommendation features, obtains the search intent of the navigated object to search along the target navigation route.

[0010] Fourthly, embodiments of the present invention provide a model training apparatus, comprising: The individual training data acquisition module is configured to acquire the target historical scene features and target historical search tag features corresponding to the target historical navigation route of the navigated object, as well as the individual historical search sequence generated by the navigated object during the navigation process before the target historical navigation route. The group training data acquisition module is configured to acquire the group historical search sequence from the historical search records corresponding to the historical navigation records of the historical navigated object, based on the starting point and ending point of the target's historical navigation route; The training data input module is configured to input the individual historical search sequence and the group historical search sequence of the navigated object as training data into the search intent prediction model; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical search tags and historical scene features; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The model includes a first target attention mechanism processing module, which obtains a first recommendation feature based on the preference features, the target historical scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module, which obtains a second recommendation feature based on the preference features, the target historical scene features, and the group's historical search sequence; The model includes an intent prediction module that, based on the first recommendation feature and the second recommendation feature, obtains the search intent corresponding to the target historical scene feature; The loss function, based on the search intent output by the intent prediction module and the target historical search tag features used as sample tags, simultaneously adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module included in the model.

[0011] The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above function.

[0012] In one possible design, the above-described device includes a memory and a processor. The memory stores one or more computer instructions that support the device in performing the corresponding methods described above, and the processor is configured to execute the computer instructions stored in the memory. The device may also include a communication interface for communicating with other devices or communication networks.

[0013] Fifthly, embodiments of this disclosure provide an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described in any of the preceding aspects.

[0014] In a sixth aspect, embodiments of this disclosure provide a computer-readable storage medium for storing computer instructions used by any of the above-described devices, which, when executed by a processor, are used to implement the methods described in any of the above aspects.

[0015] In a seventh aspect, embodiments of this disclosure provide a computer program product comprising computer instructions which, when executed by a processor, are used to implement the methods described in any of the preceding aspects.

[0016] The technical solutions provided in this disclosure may have the following beneficial effects: In this embodiment of the disclosure, when training the search intent prediction model, individual historical search sequences and group historical search sequences are used as input data. These data represent the search records generated by a single navigated object and its group during the historical navigation process, respectively. First, the preference features of the navigated object are extracted from the individual's historical search behavior through a self-attention mechanism processing module. Then, this information is further processed using two different target attention mechanism processing modules:

[0017] The first target attention mechanism module combines the preference features of the navigated object, the relevant scene features of the target navigation route, and the individual's historical search records to generate the first recommendation feature.

[0018] The second target attention mechanism module generates a second recommendation feature based on the same preference features and scene features, but interacts with the group-level historical search records.

[0019] Finally, these two recommendation features are fused to predict the search intent of the navigated object on a specific future navigation route. During training, the model parameters are continuously adjusted by constructing a loss function and comparing the model's predictions with the actual search behavior labels. This allows the model to learn which historical search activities and scenario factors are most critical for intent prediction in the current context. This approach not only considers individual preferences but also incorporates the analysis of group behavior, thus significantly improving prediction accuracy. This multi-dimensional attention mechanism-based method effectively improves the accuracy and reliability of search intent prediction while capturing individual preferences and group behavior. In this way, the model can better understand and predict the search behavior of the navigated object in different scenarios, providing a more personalized search experience.

[0020] In predicting the search intent of a navigated object, this embodiment uses the individual historical search sequence of the navigated object during the historical navigation process, and the group historical search sequence of other navigated objects during the historical navigation process, as input data for the model. The self-attention mechanism processing module included in the model calculates the preference features of the navigated object based on the individual historical search sequence of the navigated object. Then, the first target attention mechanism processing module included in the model calculates a first recommendation feature representing the similarity between the preference features of the navigated object, the current scene features of the target navigation route, and each element in the individual historical search sequence of the navigated object. The second target attention mechanism processing module included in the model calculates a second recommendation feature representing the similarity between the preference features of the navigated object, the current scene features of the target navigation route, and each element in the group historical search sequence. The first recommendation feature and the second recommendation feature are further combined to predict the search intent of the navigated object on the target historical navigation route. In this prediction method, the first recommendation feature represents the similarity between the search scenario corresponding to the target navigation route and the search scenario of the navigated object in the historical navigation process. That is, the first recommendation feature represents the individual's search preference for the current scenario. The second recommendation feature represents the similarity between the scenario corresponding to the target navigation route and the search scenarios of other navigation objects in the historical navigation process. That is, the second recommendation feature represents the group's search preference for the current scenario. The model combines the first and second recommendation features to predict the search user's search intent, thereby improving the model's prediction accuracy.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0022] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0023] Figure 1 A flowchart illustrating a search intent determination method according to an embodiment of the present disclosure is shown.

[0024] Figure 2 A flowchart illustrating a model training method according to an embodiment of the present disclosure is shown.

[0025] Figure 3 A schematic diagram of the training structure of a search intent prediction model according to an embodiment of the present disclosure is shown.

[0026] Figure 4 A structural block diagram of a search intent determination apparatus according to an embodiment of the present disclosure is shown.

[0027] Figure 5A structural block diagram of a model training apparatus according to an embodiment of the present disclosure is shown.

[0028] Figure 6 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0029] Figure 7 This is a schematic diagram of the structure of a computer system suitable for implementing a search intent determination method and / or model training method according to an embodiment of the present disclosure. Detailed Implementation

[0030] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of the exemplary embodiments have been omitted from the drawings.

[0031] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and do not preclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0032] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] The user information (including but not limited to user device information such as location information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points shall be provided for users to choose to authorize or refuse.

[0034] Existing search functions typically use rule-based strategies to determine a user's search intent. Taking a map navigation application as an example, when the user is traveling along the route provided by the application, if they enter a "search along the way" page, this page proactively displays search keywords that the user might want to enter. This allows the user to select the desired keywords from the displayed options and directly perform the search, avoiding the inconvenience of manually entering search keywords. In existing technologies, the process of determining the search keywords displayed on the "search along the way" page—that is, determining the user's search intent—includes the following:

[0035] The system uses rules and strategies to identify historical route search records that are the same as or similar to the current travel scenario of the navigated object from its historical route search data. If no historical route search records that are the same as or similar to the current travel scenario of the navigated object are identified, or if the navigated object has no historical route search data, then the system uses rules and strategies to identify historical route search records that are the same as or similar to the current travel scenario of the navigated object from the historical route search data of other navigated objects. Search keywords recorded in the same or similar historical search history will be proactively displayed on the search page for the navigated user to choose from. If the navigated user selects a search keyword displayed on the search page, it indicates that the prediction of the user's search intent was accurate. Conversely, if the navigated user does not select a search keyword displayed on the search page but instead enters a new search keyword, it indicates that the prediction of the user's search intent was inaccurate.

[0036] The inventors of this disclosure have discovered that existing technologies use rules and strategies to determine historical search records that are the same as or similar to the current travel scenario of the navigated object. However, because these rules and strategies are manually defined, they have limitations and often result in inaccurate prediction of search intent, leading to a deterioration in user experience. Therefore, this disclosure provides a technical solution for determining search intent, thereby improving the problem of inaccurate search intent prediction in existing technologies and enhancing the user experience of searching along routes.

[0037] The details of the embodiments of this disclosure are described in detail below through specific examples.

[0038] Figure 1 A flowchart illustrating a search intent determination method according to an embodiment of this disclosure is shown. Figure 1 As shown, the method includes the following steps:

[0039] In step S101, the current scene features and the pre-recorded individual historical search sequences generated by the navigated object during the historical navigation process are obtained; This solution is applied to the process of a navigated object traveling along a target navigation route, determining the search intent through a trained search intent prediction model. In specific implementation, current scene features refer to the characteristics generated when the navigated object travels along the target navigation route, which may include the origin and destination features of the target navigation route (such as origin and destination coordinates, point of interest type at the destination, etc.), travel time (i.e., the time when the target navigation route was requested), weather conditions during the journey, whether the journey occurred on a weekday or a holiday, etc. The historical navigation process refers to the navigation process before the navigated object travels along the target navigation route. An individual historical search sequence includes a series of sequence elements of the navigated object ordered by travel time; each sequence element corresponds to a historical navigation process, and each sequence element includes historical scene features and the historical search tags corresponding to those historical scene features. It should be noted that the process of the navigated object traveling along the target navigation route includes: when the navigated object receives the target navigation route recommended by the navigation service but has not started navigation, it enters the search page along the route or along the way through the route selection page, thereby triggering the method process of determining the search intent disclosed herein; or, after the navigated object receives the target navigation route recommended by the navigation service and selects one of the routes to start navigation, it enters the search page along the route or along the way through the navigation page, thereby triggering the method process of determining the search intent disclosed herein.

[0040] In step S102, based on the starting point and ending point of the target navigation route, a group historical search sequence is obtained from the historical search records corresponding to the historical navigation records of the historically navigated object; In this embodiment, the historical navigation objects may include the navigation object and other navigation objects, or only other navigation objects. Each sequence element in the group historical search sequence corresponds to a historical navigation process, and the sequence elements are also arranged according to the travel time. Each sequence element includes: historical scene features and historical search tags.

[0041] Unless otherwise specified, the historical scene features and historical search tags recorded in this disclosure refer to the historical route search scene features and historical route search tags recorded when all the navigated objects use the route search service.

[0042] The sequence elements included in the individual historical search sequence and the group historical search sequence are preferably sorted in order of travel time from earliest to latest.

[0043] In a preferred embodiment of this disclosure, the historical search tag can be a sequence of length N, where N equals the total number of interest point types. That is, each bit in the sequence corresponds to an interest point of a certain type, and the value of each bit in the historical search tag is used to characterize whether the navigated object / other navigated objects used the search term of this type of interest point in the historical navigation process.

[0044] For example, if there are five types of points of interest (POIs): food, gas station, convenience store, charging station, and restroom, then the historical search tags would be a sequence of length 5, where each element corresponds to one type of POI. Therefore, the sequence of historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 indicates that this type of POI was searched during the corresponding historical navigation process via a "follow-the-path" search, while a value of 0 indicates that this type of POI was not searched. Assuming the historical search tags are specifically [0,1,0,0,1], it means that gas stations and restrooms were searched during the corresponding historical navigation process via a "follow-the-path" search. This example is only for clearer illustration of historical search tags and should not be considered a limitation on the implementation of historical search tags.

[0045] The search intent prediction model in this embodiment includes a trained self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module. The following details the processing procedure after inputting individual historical search sequences and group historical search sequences into the model. The specific processing procedure of the model includes: In step S103, the individual's historical search sequence is input into the self-attention mechanism processing module included in the model to obtain the preference features of the navigated object; The self-attention mechanism module employs a self-attention model structure. When processing the input individual historical search sequence, it considers the relationship between each element in the sequence and all other elements. This allows for a more accurate understanding of the contextual information within the input sequence, ensuring that the obtained preference features of the navigated object accurately represent its personal preferences when using the "follow-the-path" search. For example, the navigated object might prefer searching for food in one travel scenario, or gas stations in another. In some embodiments, the self-attention mechanism module calculates the correlation (or weight) between each element in the input individual historical search sequence and all other elements.

[0046] In step S104, the preference features, the current scene features, and the individual's historical search sequence are input into the first target attention mechanism processing module included in the model to obtain the first recommendation features; The first target attention mechanism processing module adopts the target-attention model structure. Target-attention is generally used in recommendation tasks, mainly for attention operations on candidate data and used data. Therefore, the first target attention mechanism processing module is used to perform attention operations on preference features, current scene features and individual historical search sequences.

[0047] Furthermore, the input of the target-attention model structure includes two parts: key input (also known as K input) and query input (also known as Q input). In one embodiment of this disclosure, the concatenated feature obtained by concatenating preference features and current scene features can be used as the query input to the first target attention mechanism processing module, and the individual's historical search sequence can be used as the key input to the first target attention mechanism processing module. The first target attention mechanism processing module performs attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene features corresponding to each element of the individual's historical search sequence and the query, and using the calculated similarity value as a weight to weight the historical search label of that element in the individual's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the first recommendation feature.

[0048] Therefore, in step S104, the preference features, the current scene features, and the individual's historical search sequence are input into the first target attention mechanism processing module included in the model to obtain the first recommendation features, including: The preference features and the current scene features are concatenated to obtain the query object; The first target attention mechanism processing module, trained with the individual's historical search sequence as a key and the query object as input, obtains a first recommendation feature. The first recommendation feature is a weighted sum of the similarity value of the historical scene features in the individual's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

[0049] The aforementioned similarity value is the similarity value between the historical scene features, preference features, and current scene features corresponding to each element in the individual's historical search sequence. The weighted sum representation is the weighted summation of the historical search tags in the individual's historical search sequence, with the similarity value corresponding to each element in the historical scene features as the weight.

[0050] In step S105, the preference features, the current scene features, and the group's historical search sequence are input into the second target attention mechanism processing model included in the model to obtain the second recommendation features; Similarly, the second target attention mechanism processing module also adopts the target-attention model structure, which is used to perform attention operations on preference features, current scene features and group historical search sequences, thereby obtaining the second recommendation features based on the similarity between the current situation represented by preference features and current scene features and the historical situation related to the historical navigation object represented by the group historical search sequence.

[0051] Similar to the first target attention mechanism processing module, in this disclosure, the concatenated feature obtained by concatenating the preference feature and the current scene feature is used as the query input, and the group's historical search sequence is used as the key input to the second target attention mechanism processing module. This module performs an attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene feature corresponding to each element in the group's historical search sequence and the query, and using the calculated similarity value as a weight to weight the historical search tag corresponding to that element in the group's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the second recommendation feature. Therefore, step S105 involves inputting the preference feature, the current scene feature, and the group's historical search sequence into the second target attention mechanism processing module included in the model to obtain the second recommendation feature, including:

[0052] The group's historical search sequence is used as a key and the query object is input into the second target attention mechanism processing module to obtain the second recommendation feature. The second recommendation feature is a weighted sum of the similarity value of the historical scene features in the group's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

[0053] The aforementioned similarity value is the similarity value between the concatenated features of historical scene features, preference features, and current scene features corresponding to each element in the group's historical search sequence. The weighted sum representation is the sum of the weighted similarity values ​​corresponding to each element in the group's historical search sequence and the corresponding historical search tags in the group's historical search sequence.

[0054] The following example illustrates the calculation process of the first and second recommendation features: Assuming an individual's or group's historical search sequence comprises n elements, represented as: {historical scene feature 1, historical search tag 1}, {historical scene feature 2, historical search tag 2}, ..., {historical scene feature n, historical search tag n}, the similarity between historical scene feature 1 and the query object is represented by similarity 1, the similarity between historical scene feature 2 and the query object is represented by similarity 2, and the similarity between historical scene feature n and the query object is represented by similarity n. Then, the first recommended feature or the second recommended feature = similarity 1 × historical search tag 1 + similarity 2 × historical search tag 2 + ... + similarity n × historical search tag n.

[0055] In step S106, the first recommendation feature and the second recommendation feature are input into the intent prediction module included in the model to obtain the search intent of the navigated object to search along the target navigation route.

[0056] In some embodiments, the intent prediction module may employ a neural network model, such as an MLP model, or other models; this disclosure does not impose specific limitations on this. The first recommendation feature and the second recommendation feature are concatenated and input into the intent prediction module to obtain the search intent.

[0057] When the historical search tags are a sequence of N, the search intent output by the intent prediction module is a scoring sequence of length N+1. This scoring sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether the navigated object will search for that type of interest point on the target navigation route. If the score is greater than or equal to the recommendation threshold, it means that the navigated object is predicted to search for that type of interest point on the target navigation route; if the score is less than the recommendation threshold, it means that the navigated object is predicted not to search for that type of interest point on the target navigation route.

[0058] Using the example above, if the point of interest type is [food, gas station, convenience store, charging station, toilet], the search intent output by the intent prediction module is [5, 7, 3, 6, 4, 0]. The first score is the recommendation threshold, and the second to sixth scores correspond to the five point of interest types mentioned above. The point of interest types with scores higher than the threshold are food and convenience stores. Therefore, food and convenience stores are the point of interest types to be recommended to the navigation target. Finally, the search box on the "follow-the-path" page will actively provide search keywords corresponding to food and convenience stores.

[0059] In the search intent determination method proposed in the above embodiments, step S102, namely, obtaining the group historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object based on the starting point and ending point of the target navigation route, can be implemented in the following manner: Filter historical navigation records to find those with origin and destination that match the target navigation route; wherein, the historical navigation records can be the historical navigation records of the navigated object and other navigation objects besides the navigated object, or the historical navigation records of other navigation objects besides the navigated object; the travel time of these historical navigation records is earlier than the travel time corresponding to the target navigation route of the navigated object; Based on the historical search tags and historical scene features in the target's historical navigation records, the group's historical search sequence is constructed.

[0060] The above describes a search intent determination method proposed in this disclosure. The search intent prediction model used in this method is an offline-trained model. During offline training of the search intent prediction model, the loss function adjusts the parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module included in the search intent prediction model simultaneously based on the scoring sequence and sample labels output by the search intent prediction model. For an understanding of the sample labels and the offline training process of the model, please refer to the model training method provided in this disclosure.

[0061] Figure 2 A flowchart illustrating a model training method according to an embodiment of this disclosure is shown. Figure 2 As shown, the model training method includes the following steps:

[0062] In step S201, the target historical scene features and target historical search tags corresponding to the target historical navigation route of the navigated object are obtained, as well as the individual historical search sequence generated by the navigated object in the historical navigation process before the target historical navigation route. In practical implementation, target historical scene features refer to the features generated when the navigated object travels along the target historical navigation route. These features may include the origin and destination features of the target historical navigation route (such as origin and destination coordinates, point of interest type at the destination, etc.), travel time (i.e., the time when the target navigation route was requested), weather conditions during the journey, whether the journey occurred on a weekday or a holiday, etc. The historical navigation process prior to the target historical navigation route refers to the navigation process before the navigated object travels along the target historical navigation route. An individual historical search sequence includes a series of sequence elements of the navigated object, ordered by travel time. Each sequence element corresponds to a historical navigation process, and each sequence element includes historical scene features and the historical search tags corresponding to those features.

[0063] In a preferred embodiment of this disclosure, the target historical search tag can be a sequence of length N, where N equals the total number of types of points of interest. That is, each bit in the sequence corresponds to a type of point of interest, and the value of each bit in the target historical search tag is used to characterize whether the navigated object has used the search term of this type of point of interest to search along the target historical navigation route.

[0064] For example, if there are five types of points of interest (POIs): food, gas station, convenience store, charging station, and restroom, then the target historical search tags would be a sequence of length 5. Each element in the sequence corresponds to one type of POI, so the sequence of target historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 in the sequence indicates that POIs of that type were searched along the target historical navigation route using the "follow-the-path" search, while a value of 0 indicates that POIs of that type were not searched. Assuming the historical search tags are specifically [0,1,0,0,1], this means that gas stations and restrooms were searched along the target historical navigation route using the "follow-the-path" search. This example is only for more clearly illustrating target historical search tags and should not be considered a limitation on the implementation of target historical search tags.

[0065] In step S202, based on the starting point and ending point of the target historical navigation route, a group historical search sequence is obtained from the historical search records corresponding to the historical navigation records of the historically navigated object; In this embodiment, the historical navigation objects may include the navigation object and other navigation objects, or only other navigation objects. Each sequence element in the group historical search sequence corresponds to a historical navigation process, and the sequence elements are also arranged according to the travel time. Each sequence element includes: historical scene features and historical search tags.

[0066] Unless otherwise specified, the historical scene features and historical search tags recorded in this disclosure refer to the historical route search scene features and historical route search tags recorded when all the navigated objects use the route search service.

[0067] The sequence elements included in the individual historical search sequence and the group historical search sequence are preferably sorted in order of travel time from earliest to latest.

[0068] In a preferred embodiment of this disclosure, the historical search tag can be a sequence of length N, where N equals the total number of interest point types. That is, each bit in the sequence corresponds to an interest point of a certain type, and the value of each bit in the historical search tag is used to characterize whether the navigated object / other navigated objects used the search term of this type of interest point in the historical navigation process.

[0069] For example, if there are five types of points of interest (POIs): food, gas stations, convenience stores, charging stations, and restrooms, then the historical search tags would be a sequence of length 5, where each element corresponds to one type of POI. Therefore, the sequence of historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 indicates that this type of POI was searched during the corresponding historical navigation process via the "search along the way" function, while a value of 0 indicates that this type of POI was not searched. Assuming the historical search tags are specifically [0,1,0,0,1], it means that gas stations and restrooms were searched during the corresponding historical navigation process via the "search along the way" page. This example is only for more clearly illustrating historical search tags and should not be considered a limitation on the implementation of historical search tags.

[0070] The search intent prediction model in this embodiment includes a self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module. The following will detail the process of training the model using individual historical search sequences and group historical search sequences as training data. The training process of the model specifically includes: In step S203, the individual historical search sequence and the group historical search sequence of the navigated object are used as training data and input into the search intent prediction model; In this embodiment, the model includes a self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module, all of which are trained in the same training process. The first target attention mechanism processing module and the second target attention mechanism processing module are two separate modules that do not share model parameters.

[0071] The training data for the above model includes: target historical scene features, individual historical search sequences, group historical search sequences, and target historical search labels. When collecting samples, a navigation route from one of the historical navigation processes of the navigated object can be selected as the target historical navigation route. For example, the most recent historical navigation route completed by the navigated object can be selected as the target historical navigation route, and individual historical search sequences can be collected from the historical navigation processes prior to this latest historical navigation route. It can be understood that one target historical navigation route corresponds to one set of training data. This training data includes the target historical scene features, individual historical search features, and group historical search features corresponding to the target historical navigation route. The sample labels of this training data are the target historical search labels within that historical navigation route. To improve the prediction accuracy of the search intent prediction model, multiple sets of training data need to be collected. Each set of training data is input into the search intent prediction model to complete one training session. Each training process can be understood as inputting training samples into the search intent prediction model. After processing the training data, the search intent prediction model outputs a prediction result, which is the search intent corresponding to the target historical scene features output by the intent prediction module of the search intent prediction model. Then, by constructing a loss function, the search intent and the target historical search label used as the sample label are brought into the loss function. By optimizing the loss function, the model parameters of the search intent model are adjusted. Adjusting the model parameters of the search intent model includes adjusting the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module.

[0072] In step S204, the self-attention mechanism processing module included in the model obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object; The self-attention mechanism module employs a self-attention model structure. When processing the input individual historical search sequence, it considers the relationship between each element in the sequence and all other elements. This allows for a more accurate understanding of the contextual information within the input sequence, ensuring that the obtained preference features of the navigated object accurately represent its personal preferences when using the "follow-the-path" search. For example, the navigated object might prefer searching for food in one travel scenario, or gas stations in another. In some embodiments, the self-attention mechanism module calculates the correlation (or weight) between each element in the input individual historical search sequence and all other elements.

[0073] In step S205, the first target attention mechanism processing module included in the model obtains the first recommendation feature based on the preference feature, the target historical scene feature, and the individual historical search sequence; The first target attention mechanism processing module adopts the target-attention model structure. Target-attention is generally used in recommendation tasks, mainly for attention operations on candidate data and used data. Therefore, the first target attention mechanism processing module is used to perform attention operations on preference features, target historical scene features and individual historical search sequences.

[0074] Furthermore, the input to the target-attention model structure includes two parts: The input consists of a key value (also known as K-input) and a query object (also known as Q-input). In one embodiment of this disclosure, the concatenated feature obtained by combining preference features and current scene features can be used as the query object input to the first target attention mechanism processing module. The individual's historical search sequence can be used as the key value input to the first target attention mechanism processing module. The first target attention mechanism processing module performs attention operations on the key value and the query object. This attention operation includes calculating the similarity between the historical scene features corresponding to each element of the individual's historical search sequence and the query object, and using the calculated similarity value as a weight to weight the historical search tags of that element in the individual's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the first recommendation feature.

[0075] Therefore, step S205, namely, the first target attention mechanism processing module included in the model, obtains the first recommendation feature based on the preference features, the target historical scene features, and the individual historical search sequence, including: The preference features and the target historical scene features are concatenated to obtain the query object; The individual's historical search sequence is used as a key and the query object is input into the first target attention mechanism processing module to obtain the first recommendation feature. The first recommendation feature is a weighted sum of the similarity value of the historical scene features in the individual's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the target historical scene features.

[0076] The aforementioned similarity value is the similarity value between the historical scene features, preference features, and current scene features corresponding to each element in the individual's historical search sequence. The weighted sum representation is the weighted summation of the historical search tags in the individual's historical search sequence, with the similarity value corresponding to each element in the historical scene features as the weight.

[0077] In step S206, the second target attention mechanism processing module included in the model obtains the second recommendation feature based on the preference feature, the target historical scene feature, and the group historical search sequence; Similarly, the second target attention mechanism processing module also adopts the target-attention model structure, which is used to perform attention operations on preference features, target historical scene features and group historical search sequences, thereby obtaining the second recommendation feature, which is similar to the current situation represented by preference features and target historical scene features and the historical situation related to the historical navigation object represented by the group historical search sequence.

[0078] Similar to the first target attention mechanism processing module, in this disclosure, the concatenated feature obtained by concatenating the preference feature and the target historical scene feature is used as the query input, and the group historical search sequence is used as the key input to the second target attention mechanism processing module. The module performs attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene feature corresponding to each element in the group historical search sequence and the query, and using the calculated similarity value as a weight value to weight the historical search tag corresponding to that element in the group historical search sequence. Finally, the weight values ​​corresponding to all elements are summed to obtain the second recommendation feature.

[0079] Therefore, in step S206, the second target attention mechanism processing module included in the model obtains the second recommendation features based on the preference features, the target historical scene features, and the group historical search sequence, including: The group's historical search sequence is used as a key and the query object is input into the second target attention mechanism processing module to obtain the second recommendation feature. The second recommendation feature is a weighted sum of the similarity value of the historical scene features in the group's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the target historical scene features.

[0080] The aforementioned similarity value is the similarity value between the concatenated features of historical scene features, preference features, and current scene features corresponding to each element in the group's historical search sequence. The weighted sum representation is the sum of the weighted similarity values ​​corresponding to each element in the group's historical search sequence and the corresponding historical search tags in the group's historical search sequence.

[0081] The following example illustrates the calculation process of the first and second recommendation features: Assuming an individual's or group's historical search sequence comprises n elements, represented as: {historical scene feature 1, historical search tag 1}, {historical scene feature 2, historical search tag 2}, ..., {historical scene feature n, historical search tag n}, the similarity between historical scene feature 1 and the query object is represented by similarity 1, the similarity between historical scene feature 2 and the query object is represented by similarity 2, and the similarity between historical scene feature n and the query object is represented by similarity n. Then, the first recommended feature or the second recommended feature = similarity 1 × historical search tag 1 + similarity 2 × historical search tag 2 + ... + similarity n × historical search tag n.

[0082] In step S207, the intent prediction module included in the model obtains the search intent corresponding to the target historical scene features based on the first recommendation feature and the second recommendation feature; The output of the intent prediction module is a scoring sequence of length N+1. This sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether the navigated object will search for that type of interest point on the target historical navigation route. If the score is greater than or equal to the recommendation threshold, it indicates that the interest point of that type is predicted to be searched by the navigated object on the target historical navigation route; if the score is less than the recommendation threshold, it indicates that the interest point of that type is predicted not to be searched by the navigated object on the target historical navigation route.

[0083] In some embodiments, the intent prediction module may employ a neural network model, such as an MLP model, or other models; this disclosure does not impose specific limitations on this. The first recommendation feature and the second recommendation feature are concatenated and input into the intent prediction module to obtain the search intent.

[0084] The search intent output by the intent prediction module is a scoring sequence of length N+1. This sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether that type of interest point will be searched by the navigated object on the target navigation route. If the score is greater than or equal to the recommendation threshold, it indicates that the interest point of that type is predicted to be searched by the navigated object on the target navigation route; if the score is less than the recommendation threshold, it indicates that the interest point of that type is predicted not to be searched by the navigated object on the target navigation route.

[0085] Using the example above, if the point of interest type is [food, gas station, convenience store, charging station, toilet], the search intent output by the intent prediction module is [5, 7, 3, 6, 4, 0]. The first score is the scoring threshold, and the second to sixth scores correspond to the five point of interest types mentioned above. The point of interest types with scores higher than the scoring threshold are food and convenience stores. Therefore, food and convenience stores are the point of interest types to be recommended to the navigation target. Finally, the search box on the "Search Along the Way" page will actively provide search keywords corresponding to food and convenience stores.

[0086] In step S208, the loss function adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intention prediction module, which are included in the model, based on the search intent output by the intent prediction module and the target historical search tags used as sample tags.

[0087] In the training method of the search intent prediction model proposed in the above embodiments, step S202, namely, obtaining the group historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object based on the starting point and ending point of the target's historical navigation route, can be implemented in the following manner: Filter historical navigation records from the historical navigation records to find those with origin and destination that match the target historical navigation route; wherein, the historical navigation records can be historical navigation records of the navigated object and other navigation objects besides the navigated object, or historical navigation records of other navigation objects besides the navigated object; the travel time of these historical navigation records is earlier than the travel time corresponding to the target navigation route of the navigated object; Based on the historical search tags and historical scene features in the target's historical navigation records, the group's historical search sequence is constructed.

[0088] The following section, using the search intent output by the intent prediction module as an example, describes the construction process of the loss function: in, This represents the value of the loss function. Indicates the recommendation threshold. This indicates the score corresponding to the negative label. This indicates the score corresponding to the positive label. A negative label refers to the type of interest point in the sample labels that has not been searched by the navigated object, while a positive label refers to the type of interest point in the sample labels that has been searched by the navigated object.

[0089] Each training session is also a process of optimizing the loss function, with the goal of optimizing the loss function value. It gets smaller and smaller. To make the loss function value... To make si-sj, si-s0, and s0-sj increasingly smaller, if si-sj, si-s0, and s0-sj become negative, then si is less than sj, si is less than s0, and s0 is less than sj. In other words, the score si corresponding to the negative label is less than the score corresponding to the positive label, and the corresponding score si is less than the recommendation threshold s0. The score sj corresponding to the positive label is greater than the recommendation threshold s0.

[0090] Since the scoring and recommendation thresholds mentioned above are all calculated from the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module, the model parameters that minimize the loss function can be obtained when optimizing the loss function. These model parameters are the model parameters adjusted during this training process.

[0091] Figure 3 A schematic diagram illustrating the training structure of a search intent prediction model according to an embodiment of the present disclosure is shown. Figure 3 As shown, sample data is first collected, which includes the latest historical navigation route of the navigated object. This route is used as the target historical navigation route. Then, the target historical scene features and target historical search tags of the target historical navigation route are collected. The target historical scene features include, for example, the origin and destination coordinates of the target historical navigation route, the type of the destination point of interest, the navigation request time, the weather, whether it is a weekday, and whether it is a holiday. The sample data also includes all or part of the historical navigation routes of the navigated object within a certain historical period, excluding the latest historical navigation route. This data related to historical navigation routes can be processed into individual historical search sequences.

[0092] In addition, the sample data also includes all or part of the historical navigation routes within a previous period that share the same start and end points as the target historical navigation route. This data may or may not include the historical navigation routes of the navigated route. The data related to these historical navigation routes can be processed into a historical search sequence within the group.

[0093] The above sample data, according to Figure 3 The data flow shown, after being processed by the self-attention processing mechanism module, the first target attention processing mechanism module, the second target attention processing mechanism module, and the intent prediction module, results in a sequence of data with a length of N+1. The first element in this sequence is the recommendation threshold, and the following N elements are the scores corresponding to each type of interest point.

[0094] During prediction, if the score corresponding to a certain type of interest is higher than the recommendation threshold, then that type of interest can be recommended to the navigating object; otherwise, it will not be recommended. During recommendation, the search keywords corresponding to one or more interest types with the highest scores can be recommended based on the application scenario. If all scores are lower than the recommendation threshold, no recommendations will be made.

[0095] During training, the loss function is optimized using the sequence data and the target history search labels used as sample labels for the target history navigation routes. This adjusts the module parameters of the self-attention processing mechanism module, the first target attention processing mechanism module, the second target attention processing mechanism module, and the intent prediction module. After multiple training iterations, the trained self-attention processing mechanism module, the first target attention processing mechanism module, the second target attention processing mechanism module, and the intent prediction module can be obtained.

[0096] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein.

[0097] Figure 4 This diagram illustrates a structural block diagram of a search intent determination device according to an embodiment of the present disclosure. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 4 As shown, the search intent determination device includes:

[0098] The individual feature acquisition module 401 is configured to acquire current scene features and pre-recorded individual historical search sequences generated by the navigated object during historical navigation. This solution is applied to the navigation process of a navigated object traveling along a target navigation route. It determines the search intent through a trained search intent prediction model. In specific implementation, current scene features refer to the features generated when the navigated object travels along the target navigation route. These features may include the origin and destination features of the target navigation route (e.g., origin and destination coordinates, type of point of interest at the destination), travel time (i.e., the time when the target navigation route was requested), weather conditions during travel along the target navigation route, whether the travel occurred on a weekday or a holiday, etc. The historical navigation process refers to the navigation process before the navigated object travels along the target navigation route. An individual historical search sequence includes a series of sequence elements of the navigated object, ordered by travel time. Each sequence element corresponds to a historical navigation process, and each sequence element includes historical scene features and the historical search tags corresponding to those features.

[0099] The group feature acquisition module 402 is configured to acquire the group historical search sequence from the historical search records corresponding to the historical navigation records of the historical navigated object based on the starting point and ending point of the target navigation route; In this embodiment, the historical navigation objects may include the navigation object and other navigation objects, or only other navigation objects. Each sequence element in the group historical search sequence corresponds to a historical navigation process, and the sequence elements are also arranged according to the travel time. Each sequence element includes: historical scene features and historical search tags.

[0100] Unless otherwise specified, the historical scene features and historical search tags recorded in this disclosure refer to the historical route search scene features and historical route search tags recorded when all the navigated objects use the route search service.

[0101] The sequence elements included in the individual historical search sequence and the group historical search sequence are preferably sorted in order of travel time from earliest to latest.

[0102] In a preferred embodiment of this disclosure, the historical search tag can be a sequence of length N, where N equals the total number of interest point types. That is, each bit in the sequence corresponds to an interest point of a certain type, and the value of each bit in the historical search tag is used to characterize whether the navigated object / other navigated objects used the search term of this type of interest point in the historical navigation process.

[0103] For example, if there are five types of points of interest (POIs): food, gas station, convenience store, charging station, and restroom, then the historical search tags would be a sequence of length 5, where each element corresponds to one type of POI. Therefore, the sequence of historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 indicates that this type of POI was searched during the corresponding historical navigation process via a "follow-the-path" search, while a value of 0 indicates that this type of POI was not searched. Assuming the historical search tags are specifically [0,1,0,0,1], it means that gas stations and restrooms were searched during the corresponding historical navigation process via a "follow-the-path" search. This example is only for clearer illustration of historical search tags and should not be considered a limitation on the implementation of historical search tags.

[0104] The search intent prediction model in this embodiment includes a trained self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module. The following details the processing procedure after inputting individual historical search sequences and group historical search sequences into the model. The specific processing procedure of the model includes: The model includes a self-attention mechanism processing module 403, which obtains the preference features of the navigated object based on the input individual historical search sequence. The self-attention mechanism module employs a self-attention model structure. When processing the input individual historical search sequence, it considers the relationship between each element in the sequence and all other elements. This allows for a more accurate understanding of the contextual information within the input sequence, ensuring that the obtained preference features of the navigated object accurately represent its personal preferences when using the "follow-the-path" search. For example, the navigated object might prefer searching for food in one travel scenario, or gas stations in another. In some embodiments, the self-attention mechanism module calculates the correlation (or weight) between each element in the input individual historical search sequence and all other elements.

[0105] The model includes a first target attention mechanism processing module 404, which obtains a first recommendation feature based on the input preference features, the current scene features, and the individual's historical search sequence; The first target attention mechanism processing module adopts the target-attention model structure. Target-attention is generally used in recommendation tasks, mainly for attention operations on candidate data and used data. Therefore, the first target attention mechanism processing module is used to perform attention operations on preference features, current scene features and individual historical search sequences.

[0106] Furthermore, the input of the target-attention model structure includes two parts: key input (also known as K input) and query input (also known as Q input). In one embodiment of this disclosure, the concatenated feature obtained by concatenating preference features and current scene features can be used as the query input to the first target attention mechanism processing module, and the individual's historical search sequence can be used as the key input to the first target attention mechanism processing module. The first target attention mechanism processing module performs attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene features corresponding to each element of the individual's historical search sequence and the query, and using the calculated similarity value as a weight to weight the historical search label of that element in the individual's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the first recommendation feature.

[0107] Therefore, the group training data acquisition module can be implemented in the following manner: The preference features and the current scene features are concatenated to obtain the query object; The first target attention mechanism processing module, trained with the individual's historical search sequence as a key and the query object as input, obtains a first recommendation feature. The first recommendation feature is a weighted sum of the similarity value of the historical scene features in the individual's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

[0108] The aforementioned similarity value is the similarity value between the historical scene features, preference features, and current scene features corresponding to each element in the individual's historical search sequence. The weighted sum representation is the weighted summation of the historical search tags in the individual's historical search sequence, with the similarity value corresponding to each element in the historical scene features as the weight.

[0109] The model includes a second target attention mechanism processing module 405, which obtains a second recommendation feature based on the input preference features, the current scene features, and the group's historical search sequence. Similarly, the second target attention mechanism processing module also adopts the target-attention model structure, which is used to perform attention operations on preference features, current scene features and group historical search sequences, thereby obtaining the second recommendation features based on the similarity between the current situation represented by preference features and current scene features and the historical situation related to the historical navigation object represented by the group historical search sequence.

[0110] Similar to the first target attention mechanism processing module, in this disclosure, the concatenated feature obtained by combining preference features and current scene features is used as the query input, and the group's historical search sequence is used as the key input to the second target attention mechanism processing module. This module performs an attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene features corresponding to each element in the group's historical search sequence and the query, and using the calculated similarity value as a weight to weight the historical search tags corresponding to that element in the group's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the second recommendation feature. Therefore, the second target attention mechanism processing module can be implemented as follows:

[0111] The group's historical search sequence is used as a key and the query object is input into the second target attention mechanism processing module to obtain the second recommendation feature. The second recommendation feature is a weighted sum of the similarity value of the historical scene features in the group's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

[0112] The aforementioned similarity value is the similarity value between the concatenated features of historical scene features, preference features, and current scene features corresponding to each element in the group's historical search sequence. The weighted sum representation is the sum of the weighted similarity values ​​corresponding to each element in the group's historical search sequence and the corresponding historical search tags in the group's historical search sequence.

[0113] The following example illustrates the calculation process of the first and second recommendation features: Assuming an individual's or group's historical search sequence comprises n elements, represented as: {historical scene feature 1, historical search tag 1}, {historical scene feature 2, historical search tag 2}, ..., {historical scene feature n, historical search tag n}, the similarity between historical scene feature 1 and the query object is represented by similarity 1, the similarity between historical scene feature 2 and the query object is represented by similarity 2, and the similarity between historical scene feature n and the query object is represented by similarity n. Then, the first recommended feature or the second recommended feature = similarity 1 × historical search tag 1 + similarity 2 × historical search tag 2 + ... + similarity n × historical search tag n.

[0114] The model includes an intent prediction module 406, which, based on the input first recommendation features and second recommendation features, obtains the search intent of the navigated object to search along the target navigation route.

[0115] In some embodiments, the intent prediction module may employ a neural network model, such as an MLP model, or other models; this disclosure does not impose specific limitations on this. The first recommendation feature and the second recommendation feature are concatenated and input into the intent prediction module to obtain the search intent.

[0116] When the historical search tags are a sequence of N, the search intent output by the intent prediction module is a scoring sequence of length N+1. This scoring sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether the navigated object will search for that type of interest point on the target navigation route. If the score is greater than or equal to the recommendation threshold, it means that the navigated object is predicted to search for that type of interest point on the target navigation route; if the score is less than the recommendation threshold, it means that the navigated object is predicted not to search for that type of interest point on the target navigation route.

[0117] Using the example above, if the point of interest type is [food, gas station, convenience store, charging station, toilet], the search intent output by the intent prediction module is [5, 7, 3, 6, 4, 0]. The first score is the recommendation threshold, and the second to sixth scores correspond to the five point of interest types mentioned above. The point of interest types with scores higher than the threshold are food and convenience stores. Therefore, food and convenience stores are the point of interest types to be recommended to the navigation target. Finally, the search box on the "follow-the-path" page will actively provide search keywords corresponding to food and convenience stores.

[0118] In the search intent determination device proposed in the above embodiments, the group feature acquisition module 402 can be implemented in the following manner: Filter historical navigation records to find those with origin and destination that match the target navigation route; wherein, the historical navigation records can be the historical navigation records of the navigated object and other navigation objects besides the navigated object, or the historical navigation records of other navigation objects besides the navigated object; the travel time of these historical navigation records is earlier than the travel time corresponding to the target navigation route of the navigated object; Based on the historical search tags and historical scene features in the target's historical navigation records, the group's historical search sequence is constructed.

[0119] The above describes the search intent determination apparatus proposed in this disclosure. The search intent prediction model used in this apparatus is an offline-trained model. During offline training of the search intent prediction model, the loss function adjusts the parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module included in the search intent prediction model simultaneously based on the scoring sequence and sample labels output by the search intent prediction model. For an understanding of the sample labels and the offline training process of the model, please refer to the model training apparatus provided in this disclosure.

[0120] Figure 5A structural block diagram of a model training apparatus according to an embodiment of the present disclosure is shown. This apparatus can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 5 As shown, the model training device includes:

[0121] The individual training data acquisition module 501 is configured to acquire the target historical scene features and target historical search tag features corresponding to the target historical navigation route of the navigated object, as well as the individual historical search sequence generated by the navigated object during the navigation process before the target historical navigation route. In practical implementation, target historical scene features refer to the features generated when the navigated object travels along the target historical navigation route. These features may include the origin and destination features of the target historical navigation route (such as origin and destination coordinates, point of interest type at the destination, etc.), travel time (i.e., the time when the target navigation route was requested), weather conditions during the journey, whether the journey occurred on a weekday or a holiday, etc. The historical navigation process prior to the target historical navigation route refers to the navigation process before the navigated object travels along the target historical navigation route. An individual historical search sequence includes a series of sequence elements of the navigated object, ordered by travel time. Each sequence element corresponds to a historical navigation process, and each sequence element includes historical scene features and the historical search tags corresponding to those features.

[0122] In a preferred embodiment of this disclosure, the target historical search tag can be a sequence of length N, where N equals the total number of types of points of interest. That is, each bit in the sequence corresponds to a type of point of interest, and the value of each bit in the target historical search tag is used to characterize whether the navigated object has used the search term of this type of point of interest to search along the target historical navigation route.

[0123] For example, if there are five types of points of interest (POIs): food, gas station, convenience store, charging station, and restroom, then the target historical search tags would be a sequence of length 5. Each element in the sequence corresponds to one type of POI, so the sequence of target historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 in the sequence indicates that POIs of that type were searched along the target historical navigation route using the "follow-the-path" search, while a value of 0 indicates that POIs of that type were not searched. Assuming the historical search tags are specifically [0,1,0,0,1], this means that gas stations and restrooms were searched along the target historical navigation route using the "follow-the-path" search. This example is only for more clearly illustrating target historical search tags and should not be considered a limitation on the implementation of target historical search tags.

[0124] The group training data acquisition module 502 is configured to acquire the group historical search sequence from the historical search records corresponding to the historical navigation records of the historical navigated object based on the starting point and ending point of the target historical navigation route; In this embodiment, the historical navigation objects may include the navigation object and other navigation objects, or only other navigation objects. Each sequence element in the group historical search sequence corresponds to a historical navigation process, and the sequence elements are also arranged according to the travel time. Each sequence element includes: historical scene features and historical search tags.

[0125] Unless otherwise specified, the historical scene features and historical search tags recorded in this disclosure refer to the historical route search scene features and historical route search tags recorded when all the navigated objects use the route search service.

[0126] The sequence elements included in the individual historical search sequence and the group historical search sequence are preferably sorted in order of travel time from earliest to latest.

[0127] In a preferred embodiment of this disclosure, the historical search tag can be a sequence of length N, where N equals the total number of interest point types. That is, each bit in the sequence corresponds to an interest point of a certain type, and the value of each bit in the historical search tag is used to characterize whether the navigated object / other navigated objects used the search term of this type of interest point in the historical navigation process.

[0128] For example, if there are five types of points of interest (POIs): food, gas station, convenience store, charging station, and restroom, then the historical search tags would be a sequence of length 5, where each element corresponds to one type of POI. Therefore, the sequence of historical search tags would be [food, gas station, convenience store, charging station, restroom]. A value of 1 indicates that this type of POI was searched during the corresponding historical navigation process via a "follow-the-path" search, while a value of 0 indicates that this type of POI was not searched. Assuming the historical search tags are specifically [0,1,0,0,1], it means that gas stations and restrooms were searched during the corresponding historical navigation process via a "follow-the-path" search. This example is only for clearer illustration of historical search tags and should not be considered a limitation on the implementation of historical search tags.

[0129] The search intent prediction model in this embodiment includes a self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module. The following will detail the process of training the model using individual historical search sequences and group historical search sequences as training data. The training process of the model specifically includes: The training data input module 503 is configured to input the individual historical search sequence and the group historical search sequence of the navigated object as training data into the search intent prediction model; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical search label and historical scene feature; In this embodiment, the model includes a self-attention mechanism processing module, a first target attention mechanism processing module, a second target attention mechanism processing module, and an intent prediction module, all of which are trained in the same training process. The first target attention mechanism processing module and the second target attention mechanism processing module are two separate modules that do not share model parameters.

[0130] The training data for the above model includes: target historical scene features, individual historical search sequences, group historical search sequences, and target historical search labels. When collecting samples, a navigation route from one of the historical navigation processes of the navigated object can be selected as the target historical navigation route. For example, the most recent historical navigation route completed by the navigated object can be selected as the target historical navigation route, and individual historical search sequences can be collected from the historical navigation processes prior to this latest historical navigation route. It can be understood that one target historical navigation route corresponds to one set of training data. This training data includes the target historical scene features, individual historical search features, and group historical search features corresponding to the target historical navigation route. The sample labels of this training data are the target historical search labels within that historical navigation route. To improve the prediction accuracy of the search intent prediction model, multiple sets of training data need to be collected. Each set of training data is input into the search intent prediction model to complete one training session. Each training process can be understood as inputting training samples into the search intent prediction model. After processing the training data, the search intent prediction model outputs a prediction result, which is the search intent corresponding to the target historical scene features output by the intent prediction module of the search intent prediction model. Then, by constructing a loss function, the search intent and the target historical search label used as the sample label are brought into the loss function. By optimizing the loss function, the model parameters of the search intent model are adjusted. Adjusting the model parameters of the search intent model includes adjusting the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module.

[0131] The model includes a self-attention mechanism processing module 504, which obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The self-attention mechanism module employs a self-attention model structure. When processing the input individual historical search sequence, it considers the relationship between each element in the sequence and all other elements. This allows for a more accurate understanding of the contextual information within the input sequence, ensuring that the obtained preference features of the navigated object accurately represent its personal preferences when using the "follow-the-path" search. For example, the navigated object might prefer searching for food in one travel scenario, or gas stations in another. In some embodiments, the self-attention mechanism module calculates the correlation (or weight) between each element in the input individual historical search sequence and all other elements.

[0132] The model includes a self-attention mechanism processing module 505, which obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The first target attention mechanism processing module adopts the target-attention model structure. Target-attention is generally used in recommendation tasks, mainly for attention operations on candidate data and used data. Therefore, the first target attention mechanism processing module is used to perform attention operations on preference features, target historical scene features and individual historical search sequences.

[0133] Furthermore, the input to the target-attention model structure includes two parts: The input consists of a key value (also known as K-input) and a query object (also known as Q-input). In one embodiment of this disclosure, the concatenated feature obtained by combining preference features and current scene features can be used as the query object input to the first target attention mechanism processing module. The individual's historical search sequence can be used as the key value input to the first target attention mechanism processing module. The first target attention mechanism processing module performs attention operations on the key value and the query object. This attention operation includes calculating the similarity between the historical scene features corresponding to each element of the individual's historical search sequence and the query object, and using the calculated similarity value as a weight to weight the historical search tags of that element in the individual's historical search sequence. Finally, the weighted values ​​corresponding to all elements are summed to obtain the first recommendation feature.

[0134] Therefore, the first target attention mechanism processing module can be implemented as follows: The preference features and the target historical scene features are concatenated to obtain the query object; The individual's historical search sequence is used as a key and the query object is input into the first target attention mechanism processing module to obtain the first recommendation feature. The first recommendation feature is a weighted sum of the similarity value of the historical scene features in the individual's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the target historical scene features.

[0135] The aforementioned similarity value is the similarity value between the historical scene features, preference features, and current scene features corresponding to each element in the individual's historical search sequence. The weighted sum representation is the weighted summation of the historical search tags in the individual's historical search sequence, with the similarity value corresponding to each element in the historical scene features as the weight.

[0136] The model includes a second target attention mechanism processing module 506, which obtains a second recommendation feature based on the preference features, the target historical scene features, and the group historical search sequence; Similarly, the second target attention mechanism processing module also adopts the target-attention model structure, which is used to perform attention operations on preference features, target historical scene features and group historical search sequences, thereby obtaining the second recommendation feature, which is similar to the current situation represented by preference features and target historical scene features and the historical situation related to the historical navigation object represented by the group historical search sequence.

[0137] Similar to the first target attention mechanism processing module, in this disclosure, the concatenated feature obtained by concatenating the preference feature and the target historical scene feature is used as the query input, and the group historical search sequence is used as the key input to the second target attention mechanism processing module. The module performs attention operation on the key and the query. This attention operation includes calculating the similarity between the historical scene feature corresponding to each element in the group historical search sequence and the query, and using the calculated similarity value as a weight value to weight the historical search tag corresponding to that element in the group historical search sequence. Finally, the weight values ​​corresponding to all elements are summed to obtain the second recommendation feature.

[0138] Therefore, the second target attention mechanism processing module can be implemented as follows: The group's historical search sequence is used as a key and the query object is input into the second target attention mechanism processing module to obtain the second recommendation feature. The second recommendation feature is a weighted sum of the similarity value of the historical scene features in the group's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the target historical scene features.

[0139] The aforementioned similarity value is the similarity value between the concatenated features of historical scene features, preference features, and current scene features corresponding to each element in the group's historical search sequence. The weighted sum representation is the sum of the weighted similarity values ​​corresponding to each element in the group's historical search sequence and the corresponding historical search tags in the group's historical search sequence.

[0140] The following example illustrates the calculation process of the first and second recommendation features: Assuming an individual's or group's historical search sequence comprises n elements, represented as: {historical scene feature 1, historical search tag 1}, {historical scene feature 2, historical search tag 2}, ..., {historical scene feature n, historical search tag n}, the similarity between historical scene feature 1 and the query object is represented by similarity 1, the similarity between historical scene feature 2 and the query object is represented by similarity 2, and the similarity between historical scene feature n and the query object is represented by similarity n. Then, the first recommended feature or the second recommended feature = similarity 1 × historical search tag 1 + similarity 2 × historical search tag 2 + ... + similarity n × historical search tag n.

[0141] The model includes an intent prediction module 507, which obtains the search intent corresponding to the target historical scene features based on the first recommendation feature and the second recommendation feature. The output of the intent prediction module is a scoring sequence of length N+1. This sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether the navigated object will search for that type of interest point on the target historical navigation route. If the score is greater than or equal to the recommendation threshold, it indicates that the interest point of that type is predicted to be searched by the navigated object on the target historical navigation route; if the score is less than the recommendation threshold, it indicates that the interest point of that type is predicted not to be searched by the navigated object on the target historical navigation route.

[0142] In some embodiments, the intent prediction module may employ a neural network model, such as an MLP model, or other models; this disclosure does not impose specific limitations on this. The first recommendation feature and the second recommendation feature are concatenated and input into the intent prediction module to obtain the search intent.

[0143] The search intent output by the intent prediction module is a scoring sequence of length N+1. This sequence includes a recommendation threshold and N scores, each corresponding to a type of interest point. Each score represents a predicted score indicating whether that type of interest point will be searched by the navigated object on the target navigation route. If the score is greater than or equal to the recommendation threshold, it indicates that the interest point of that type is predicted to be searched by the navigated object on the target navigation route; if the score is less than the recommendation threshold, it indicates that the interest point of that type is predicted not to be searched by the navigated object on the target navigation route.

[0144] Using the example above, if the point of interest type is [food, gas station, convenience store, charging station, toilet], the search intent output by the intent prediction module is [5, 7, 3, 6, 4, 0]. The first score is the scoring threshold, and the second to sixth scores correspond to the five point of interest types mentioned above. The point of interest types with scores higher than the scoring threshold are food and convenience stores. Therefore, food and convenience stores are the point of interest types to be recommended to the navigation target. Finally, the search box on the "Search Along the Way" page will actively provide search keywords corresponding to food and convenience stores.

[0145] The loss function 508 adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module simultaneously, based on the search intent output by the intent prediction module and the target historical search tag features used as sample tags.

[0146] In the training apparatus for the search intent prediction model proposed in the above embodiments, the group training data acquisition module can be implemented in the following manner: Filter historical navigation records from the historical navigation records to find those with origin and destination that match the target historical navigation route; wherein, the historical navigation records can be historical navigation records of the navigated object and other navigation objects besides the navigated object, or historical navigation records of other navigation objects besides the navigated object; the travel time of these historical navigation records is earlier than the travel time corresponding to the target navigation route of the navigated object; Based on the historical search tags and historical scene features in the target's historical navigation records, the group's historical search sequence is constructed.

[0147] The following section, using the search intent output by the intent prediction module as an example, describes the construction process of the loss function: in, This represents the value of the loss function. Indicates the recommendation threshold. This indicates the score corresponding to the negative label. This indicates the score corresponding to the positive label. A negative label refers to the type of interest point in the sample labels that has not been searched by the navigated object, while a positive label refers to the type of interest point in the sample labels that has been searched by the navigated object.

[0148] Each training session is also a process of optimizing the loss function, with the goal of optimizing the loss function value. It gets smaller and smaller. To make the loss function value... To make si-sj, si-s0, and s0-sj increasingly smaller, if si-sj, si-s0, and s0-sj become negative, then si is less than sj, si is less than s0, and s0 is less than sj. In other words, the score si corresponding to the negative label is less than the score corresponding to the positive label, and the corresponding score si is less than the recommendation threshold s0. The score sj corresponding to the positive label is greater than the recommendation threshold s0.

[0149] Since the scoring and recommendation thresholds mentioned above are all calculated from the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module, the model parameters that minimize the loss function can be obtained when optimizing the loss function. These model parameters are the model parameters adjusted during this training process.

[0150] This disclosure also discloses an electronic device. Figure 6 This diagram illustrates a structural block diagram of an electronic device according to an embodiment of the present disclosure, such as... Figure 6 As shown, the electronic device 600 includes a memory 601 and a processor 602; wherein, The memory 601 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 602 to implement the above method steps.

[0151] Figure 7 This is a schematic diagram of the structure of a computer system suitable for implementing a search intent determination method and / or model training method according to an embodiment of the present disclosure.

[0152] like Figure 7As shown, the computer system 700 includes a processing unit 701, which can be implemented as a CPU, GPU, FPGA, NPU, or other processing unit. The processing unit 701 can execute various processes according to any of the methods described above in this disclosure, based on a program stored in the read-only memory (ROM) 702 or a program loaded from the storage portion 708 into the random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the computer system 700. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0153] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 710 as needed so that computer programs read from it can be installed into the storage section 708 as needed.

[0154] In particular, according to embodiments of this disclosure, any of the methods described above in the embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program containing program code for performing any of the methods in the embodiments of this disclosure. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711.

[0155] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0156] The units or modules described in the embodiments of this disclosure can be implemented in software or hardware. The described units or modules can also be located in a processor, and the names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0157] In another aspect, this disclosure also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores one or more programs that are used by one or more processors to perform the methods described in this disclosure.

[0158] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A method for determining search intent, wherein, Applied to the process of a navigating object traveling along a target navigation route, the search intent is determined through a trained search intent prediction model, including: Obtain the current scene features and the pre-recorded individual historical search sequences generated by the navigated object during the historical navigation process; Based on the starting point and ending point of the target navigation route, a group historical search sequence is obtained from the historical search records corresponding to the historical navigation records of the historically navigated object; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical scene features and historical search tags; The individual's historical search sequence is input into the self-attention mechanism processing module included in the model to obtain the preference features of the navigated object; The preference features, the current scene features, and the individual's historical search sequence are input into the first target attention mechanism processing module included in the model to obtain the first recommendation features; The preference features, the current scene features, and the group's historical search sequence are input into the second target attention mechanism processing module included in the model to obtain the second recommendation features; The first recommendation feature and the second recommendation feature are input into the intent prediction module included in the model to obtain the search intent of the navigated object to search along the target navigation route.

2. The method according to claim 1, wherein, The preference features, the current scene features, and the individual's historical search sequence are input into the first target attention mechanism processing module included in the model to obtain the first recommendation features, including: The preference features and the current scene features are concatenated to obtain the query object; The first target attention mechanism processing module, trained with the individual's historical search sequence as a key and the query object as input, obtains a first recommendation feature. The first recommendation feature is a weighted sum of the similarity value of the historical scene features in the individual's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

3. The method according to claim 1, wherein, The preference features, the current scene features, and the group's historical search sequence are input into the second target attention mechanism processing module included in the model to obtain the second recommendation features, including: The group's historical search sequence is used as a key and the query object is input into the second target attention mechanism processing module to obtain the second recommendation feature. The second recommendation feature is a weighted sum of the similarity value of the historical scene features in the group's historical search sequence and the historical search tags corresponding to the historical scene features. The similarity value of the historical scene features represents the similarity between the historical scene features and the preference features and the current scene features.

4. The method according to any one of claims 1-3, wherein, The historical search tags are a sequence of length N, where N equals the total number of interest point types. Each bit in the sequence corresponds to an interest point of a certain type, and the value of each bit is used to indicate whether the search term for that type of interest point was used during the historical navigation process.

5. The method according to claim 4, wherein, The search intent is a scoring sequence with a length of N+1. The scoring sequence includes a recommendation threshold and N scores, where each of the N scores corresponds to a type of interest point.

6. The method according to claim 4, wherein, When training the search intent prediction model, the loss function is based on the scoring sequence and sample labels output by the search intent prediction model, and the parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module and the intent prediction module included in the search intent prediction model are adjusted simultaneously.

7. A model training method, wherein, include: Obtain the target historical scene features and target historical search tag features corresponding to the target historical navigation route of the navigated object, as well as the individual historical search sequence generated by the navigated object during the historical navigation process before the target historical navigation route; Based on the starting point and ending point of the target's historical navigation route, obtain the group's historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object; The individual historical search sequence and the group historical search sequence of the navigated object are used as training data and input into the search intent prediction model; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical search label and historical scene feature; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The model includes a first target attention mechanism processing module that obtains a first recommendation feature based on the preference features, the target historical scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module that obtains a second recommendation feature based on the preference features, the target historical scene features, and the group's historical search sequence; The model includes an intent prediction module that, based on the first recommendation feature and the second recommendation feature, obtains the search intent corresponding to the target historical scene feature; The loss function adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module simultaneously based on the search intent output by the intent prediction module and the target historical search tag features used as sample tags.

8. A search intent determination device, wherein, Used to determine search intent through a trained search intent prediction model during the travel of the navigated object along the target navigation route, including: The individual feature acquisition module is configured to acquire current scene features and pre-recorded individual historical search sequences generated by the navigated object during historical navigation. The group feature acquisition module is configured to acquire a group historical search sequence from the historical search records corresponding to the historical navigation records of the historically navigated object, based on the start and end points of the target navigation route; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical scene features and historical search tags; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the input individual's historical search sequence. The model includes a first target attention mechanism processing module, which obtains a first recommendation feature based on the input preference features, the current scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module, which obtains a second recommendation feature based on the input preference features, the current scene features, and the group's historical search sequence; The model includes an intent prediction module that, based on the input first recommendation features and second recommendation features, obtains the search intent of the navigated object to search along the target navigation route.

9. A model training device, wherein, include: The individual training data acquisition module is configured to acquire the target historical scene features and target historical search tag features corresponding to the target historical navigation route of the navigated object, as well as the individual historical search sequence generated by the navigated object during the navigation process before the target historical navigation route. The group training data acquisition module is configured to acquire the group historical search sequence from the historical search records corresponding to the historical navigation records of the historical navigated object, based on the starting point and ending point of the target's historical navigation route; The training data input module is configured to input the individual historical search sequence and the group historical search sequence of the navigated object as training data into the search intent prediction model; wherein, each sequence element in the individual historical search sequence and the group historical search sequence includes: historical search tags and historical scene features; The model includes a self-attention mechanism processing module that obtains the preference features of the navigated object based on the individual historical search sequence of the navigated object. The model includes a first target attention mechanism processing module, which obtains a first recommendation feature based on the preference features, the target historical scene features, and the individual's historical search sequence; The model includes a second target attention mechanism processing module, which obtains a second recommendation feature based on the preference features, the target historical scene features, and the group's historical search sequence; The model includes an intent prediction module that, based on the first recommendation feature and the second recommendation feature, obtains the search intent corresponding to the target historical scene feature; The loss function, based on the search intent output by the intent prediction module and the target historical search tag features used as sample tags, simultaneously adjusts the model parameters of the self-attention mechanism processing module, the first target attention mechanism processing module, the second target attention mechanism processing module, and the intent prediction module included in the model.

10. A computer-readable storage medium having stored thereon computer instructions, wherein, When executed by a processor, the computer instructions implement the method described in any one of claims 1-7.

Images