Multi-word search implementation method, apparatus, medium, and device

Abstract

The application provides a multi-word search implementation method and device, a medium and equipment. The method comprises the following steps: determining a plurality of query words received by a search box, wherein the plurality of query words at least comprise a first query word and a second query word corresponding to different search items with different semantics; in response to a confirmation instruction of a multi-word intelligent collocation mode, sending the plurality of query words to a server to request intelligent collocation search based on the plurality of query words; receiving at least one intelligent collocation search result returned by the server, wherein a target first search item and a target second search item are determined according to an intelligent collocation strategy representing a dependency relationship between query words, and the intelligent collocation search result is obtained; and displaying the at least one intelligent collocation search result. The application can improve search efficiency.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, medium and device for implementing multi-word search. Background Technology

[0002] When using apps or mini-programs, users frequently use the search function to find target objects. For example, in e-commerce scenarios such as retail, catering, and pharmaceuticals, users enter query terms in the search box to search for target objects (target products, target dishes, target medicines, etc.). The efficiency with which users browse and select valid search items directly affects user conversion rates. Therefore, improving search efficiency is a technical problem that needs to be solved. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, medium and device for multi-word search, the main purpose of which is to improve search efficiency.

[0004] According to one aspect of this application, a multi-word search implementation method is provided for use on a terminal side. The method includes: determining multiple query terms received by a search box, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; in response to a confirmation instruction of a multi-word intelligent collocation mode, sending the multiple query terms to a server and requesting intelligent collocation search based on the multiple query terms; receiving at least one intelligent collocation search result returned by the server, wherein a target first search term and a target second search term are determined according to an intelligent collocation strategy representing the dependency relationship between query terms, thereby obtaining the intelligent collocation search result; and displaying the at least one intelligent collocation search result.

[0005] According to one aspect of this application, a multi-word search implementation method is provided for a server side. The method includes: receiving multiple query terms sent by a terminal and confirming that the terminal requests a multi-word intelligent collocation search, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; determining at least one intelligent collocation search result and returning the at least one intelligent collocation search result to the terminal, wherein a target first search term and a target second search term are determined according to an intelligent collocation strategy representing the dependency relationship between query terms, and the intelligent collocation search result is obtained. According to one aspect of this application, a storage medium is provided, wherein a computer program is stored in the storage medium, wherein the computer program is configured to execute the above method at runtime.

[0006] According to one aspect of this application, a training method for a multi-word intelligent collocation prediction model is provided, comprising: acquiring sample data and performing feature engineering processing on the sample data to obtain input features of the prediction model: at least one of user-end features, product-end features, query term-end features, logistics-end features, and spatiotemporal features; learning a neural network structure with dependencies between multiple query terms based on at least one of the user-end features, product features, query term-end features, logistics features, and spatiotemporal features to obtain a click-through rate (CTR) prediction model.

[0007] According to one aspect of this application, a multi-word search implementation apparatus is provided for use on a terminal side. The apparatus includes: a query term determination unit, configured to determine multiple query terms received by a search box, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; a smart collocation request unit, configured to send the multiple query terms to a server in response to a confirmation instruction of a multi-word smart collocation mode, requesting a smart collocation search based on the multiple query terms; a smart collocation confirmation unit, configured to receive at least one smart collocation search result returned by the server, wherein a target first search term and a target second search term are determined according to a smart collocation strategy representing the dependency relationship between query terms, thereby obtaining the smart collocation search result; and a smart collocation display unit, configured to display the at least one smart collocation search result.

[0008] According to one aspect of this application, a multi-word search implementation apparatus is provided for use on a server side. The apparatus includes: a query term receiving unit for receiving multiple query terms sent by a terminal, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; and an intelligent collocation determination unit for confirming that the terminal requests a multi-word intelligent collocation search, and determining at least one intelligent collocation search result and returning it to the terminal, wherein a target first search term and a target second search term are determined according to an intelligent collocation strategy that characterizes the dependency relationship between query terms, thereby obtaining the intelligent collocation search result.

[0009] According to one aspect of this application, a training apparatus for a multi-word intelligent collocation prediction model is provided, comprising: a sample acquisition unit for acquiring sample data; a feature engineering processing unit for performing feature engineering processing on the sample data to obtain input features of the prediction model: at least one of user-side features, product-side features, query term-side features, logistics-side features, and spatiotemporal features; and a learning unit for learning a neural network structure with dependencies between multiple query terms based on at least one of the user-side features, product features, query term-side features, logistics features, and spatiotemporal features, to obtain a click-through rate (CTR) prediction model.

[0010] According to one aspect of this application, a storage medium is provided that stores a computer program, wherein the computer program is configured to execute the above-described method at runtime.

[0011] According to one aspect of this application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the methods described above.

[0012] By utilizing the above technical solutions, this application provides a multi-word search implementation method, apparatus, medium, and device that can enable users to place orders with a single click. For example, in near-field e-commerce and food delivery scenarios, users may wish to enhance their ability to shop across multiple stores in one stop by searching for multiple products and finding merchants that offer the same products to place an order. Alternatively, it can help users find a combination of stores that can simultaneously meet their order requirements (e.g., those that offer shorter delivery times, lower delivery fees, or the best value for money), reducing the time users spend searching and filtering, minimizing search steps, and improving search efficiency.

[0013] The multi-word search implementation method provided in this application, in the default intelligent matching search mode, automatically provides intelligent matching search results based on the dependencies between multiple query terms, enabling one-click searching of multiple items, shortening the search process, and improving search efficiency. In addition to providing the intelligent matching search mode, it also offers a user-defined matching search mode. In this mode, users can obtain self-matched search results by selecting items corresponding to each query term. Therefore, by combining the intelligent matching search mode and the self-matching search mode, the needs of different users in multi-word search scenarios can be met. The intelligent matching search mode addresses users' concerns regarding cost and delivery time, while the self-matching search mode addresses users' needs regarding the quality of the items themselves, satisfying personalized matching requirements.

[0014] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 The illustration shows a schematic diagram of an implementation scenario for a multi-word search method provided in this application.

[0017] Figure 2 This paper illustrates a flowchart of a multi-word search implementation method for a terminal provided by an embodiment of this application;

[0018] Figure 3 The present application provides a schematic diagram of the intelligent collocation mode in a multi-word search implementation method.

[0019] Figure 4 This illustration shows a schematic diagram of the self-matching mode in a multi-word search implementation method provided in an embodiment of this application;

[0020] Figure 5 This paper illustrates a flowchart of a multi-word search implementation method for a server provided in an embodiment of this application.

[0021] Figure 6 The flowchart illustrates a training method for a multi-word intelligent collocation prediction model provided in an embodiment of this application.

[0022] Figure 7 This illustration shows a training method for a multi-word intelligent collocation prediction model provided in an embodiment of this application;

[0023] Figure 8 This illustration shows a schematic diagram of a multi-word search implementation device located on a terminal according to an embodiment of this application;

[0024] Figure 9 This illustration shows a schematic diagram of a multi-word search implementation device located on a server side, according to an embodiment of this application.

[0025] Figure 10 This diagram illustrates the structure of a training device for a multi-word intelligent collocation prediction model provided in an embodiment of this application. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0027] The terminology is explained below.

[0028] query: Represents the user's query term (or search term).

[0029] item: Indicates the search terms to be displayed, such as goods, dishes, medicines, or services.

[0030] Relevance prediction model: An algorithmic module that calculates the relevance between user query terms and retrieved search terms such as goods / services. For example, a relevance prediction model can be trained based on the click-through rate (CTR) algorithm.

[0031] During their research, the inventors of this application discovered that when users search for target objects using query terms, they sometimes need to search for multiple target objects. For example, in a food ordering scenario, a user might want to order both "pork rib soup" and "braised fish." Using traditional search methods, the user would need to perform two separate searches and place two separate orders, which is not only cumbersome and increases browsing and selection time, but could also potentially extend delivery times and increase payment amounts. Therefore, based on this, the inventors of this application propose a multi-word search implementation method that allows users to enter multiple query terms into the search box at once and perform a one-click search based on these multiple query terms, thus improving search efficiency.

[0032] See Figure 1 This is a schematic diagram illustrating an implementation scenario of a multi-word search method provided in this application. The scenario illustrates a terminal and a server. The terminal can be a user terminal, such as a mobile phone, computer, or smartwatch. The server refers to a network-side device, such as a server, that responds to a terminal's search request, performs a search, and returns search results.

[0033] In this example, we'll use the search page of a specific app as an illustration. The app typically has a search box at the top (or elsewhere), where users can input search terms using Chinese characters, images, or voice. Specifically, in this embodiment, the search box supports multi-word (multiple search terms) input. These search terms are distinguished by special characters, such as spaces, backslashes (\), or commas. For example, if a user inputs "boiled fish, Kung Pao chicken," boiled fish is the first search term, and Kung Pao chicken is the second, separated by a comma ("、"). After the app receives the multiple search terms, it transmits them to a network server, which then returns search results based on its recommendation system.

[0034] In this scenario, a recommendation system is illustrated on the server side. This system includes a recall module, a ranking (CTR prediction) module, and a mechanism strategy module. For each query term, the recall module targets all items, while the ranking module targets only the recalled items. The ranking module sorts the items based on a query-result relevance model. In this embodiment, for multiple query terms, after obtaining the recalled items for each query term, a target item is selected from the recalled items based on an intelligent matching strategy. The target items corresponding to each query term are then combined to obtain the intelligent matching search results. For example, in the above example, for the user-inputted queries "boiled fish" and "Kung Pao chicken," the server first recalls multiple items corresponding to "boiled fish" and multiple items corresponding to "Kung Pao chicken." Then, based on the intelligent matching strategy, one item is selected from the multiple items corresponding to "boiled fish," and another item is selected from the multiple items corresponding to "Kung Pao chicken." These two items constitute the intelligent matching search results. It should be noted that the intelligent matching search results generally include multiple results. According to the priority of the intelligent matching strategy, the first intelligent matching search result, the second intelligent matching search result, the third intelligent matching search result, and so on are determined according to the priority. These intelligent search results are returned to the terminal for display, allowing the user to select a search result and place an order.

[0035] See Figure 2 The diagram shows a flowchart of a multi-word search implementation method for a terminal provided by an embodiment of this application. The multi-word search implementation method includes the following steps S201-S204.

[0036] S201: Determine the multiple query terms received by the search box, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms.

[0037] In information retrieval search scenarios, users may enter multiple query terms in the search box. For example, a user might enter "Beijing Children's Hospital bus," which includes the three query terms "Beijing," "Children's Hospital," and "bus," aiming to obtain information about "buses to Beijing Children's Hospital." These three query terms do not correspond to individual items (search terms) but rather represent a collective search intent. Unlike information retrieval search scenarios, this application's embodiments can be applied to product search and order placement scenarios. For example, in near-field e-commerce and food delivery scenarios, users can search for and place orders for products by entering query terms. Each query term corresponds to a search for a product or service item, and different query terms correspond to different items.

[0038] Query terms refer to the meaning information provided by the user in the search box through Chinese characters or pinyin input, image insertion, voice input, etc. In this embodiment, multiple query terms can be input; for example, the user can input three query terms A, B, and C at once. Each query term corresponds to a different semantic meaning and a different search item. For example, query term A represents dish 1, query term B represents dish 2, and query term C represents beverage 1. Generally, multiple query terms correspond to products under the same product category. For example, the above three query terms all correspond to products in the catering category, thus facilitating the retrieval of all products from the same store as much as possible.

[0039] To distinguish between different query terms, special character recognition can be used to differentiate them. In one implementation, after determining the multiple query terms received by the search box, the following steps are also included:

[0040] 1. Identify special characters that are not text in the search box;

[0041] 2. Determine whether the special character is a preset multi-word delimiter. If so, use the multi-word delimiter to distinguish each query term and obtain the first query term and the second query term.

[0042] For example, the special character can be a space character, "、", " / ", etc., and this application embodiment does not limit it.

[0043] S202: In response to the confirmation command of the multi-word intelligent collocation mode, send multiple query terms to the server and request intelligent collocation search based on multiple query terms.

[0044] To minimize user interaction and provide convenient and quick multi-word search results, thereby improving search efficiency, a multi-word intelligent collocation mode can be provided by default on the search page of the app or mini-program. Specifically, when a multi-word search is detected in the search box, the multi-word intelligent collocation mode is activated by default, automatically providing one-click search results that satisfy multiple query terms. Corresponding to the multi-word intelligent collocation mode, a multi-word self-collagement mode can also be provided; the specific implementation of the multi-word self-collagement mode is explained in detail below.

[0045] See Figure 3 The illustration shows a schematic diagram of the intelligent collocation mode in a multi-word search implementation method provided in this application embodiment. The search box contains three query terms: "fish," "chicken," and "beef," separated by spaces. The intelligent collocation mode is selected by default, and the intelligent collocation results are displayed.

[0046] S203: Receive at least one intelligent collocation search result returned by the server, wherein, based on the intelligent collocation strategy representing the dependency relationship between query terms, the first target search term and the second target search term are determined, and the intelligent collocation search result is obtained.

[0047] During their research, the inventors of this application discovered that, in order to improve search efficiency in multi-word search scenarios, compared to traditional single-word search scenarios, multi-word search not only needs to consider factors related to the "product" itself (such as item visit rate), but also the dependencies between query terms. The dependencies between query terms can be understood as the degree of closeness between query terms influenced by the spatiotemporal interaction information of each item. This spatiotemporal interaction information can include whether they belong to the same store, visit rate, delivery time, price, and spatiotemporal differences. For example, if the corresponding items for query terms A and B both exist in the same store, then these two query terms have a strong dependency. Furthermore, when combining items from different stores, the focus is on visit rate and delivery time; item combinations with high visit rates and short delivery times indicate a strong dependency between the corresponding query terms.

[0048] In one implementation, the intelligent matching strategy is confirmed as: a first strategy of prioritizing matching with items from the same store; based on the first strategy of prioritizing matching with items from the same store, target first search terms and target second search terms belonging to the same store are determined from at least one first search term corresponding to the first query term and from at least one second search term corresponding to the second query term, respectively.

[0049] In another implementation, the intelligent matching strategy is confirmed as follows: A second strategy for inter-store matching is determined based on combined influencing factors; based on this second strategy, target first and second search terms corresponding to the combined influencing factors are determined from at least one first search term corresponding to the first query term and at least one second search term corresponding to the second query term. For example, combined influencing factors refer to at least one of the following: visit rate factor, delivery factor, price factor, and spatiotemporal factor. In one implementation, the priority of each combined influencing factor can be set, and the current combination of one or more combined influencing factors can be determined based on their priority, thereby determining the target search term based on the current combined factors. In another implementation, a combined influencing factor filtering interface can be provided to the user, allowing the user to select influencing factors, such as displaying visit rate factor (most popular), delivery factor (shortest delivery time), and price factor (lowest total price). Based on one or more combined influencing factors selected by the user, items are matched using the selected combined influencing factors to obtain the matching results.

[0050] Therefore, based on the intelligent matching strategy that represents the dependencies between query terms, the target first search term and the target second search term are determined, and the intelligent matching search results are obtained. For example, the query terms include a first query term and a second query term. Assume that there are multiple first search terms recalled for the first query term: first search term 1, first search term 2, ..., first search term n, and multiple second search terms recalled for the second query term: second search term 1, second search term 2, ..., second search term n. In the search term set (first search term 1-n, second search term 1-n), based on the dependencies between query terms, that is, based on the closeness between the first and second query terms due to the influence of the spatiotemporal interaction information of each item, assuming that the first search term i and the second search term i are selected, the first search term i is taken as the target first search term, and the second search term i is taken as the target second search term. The target first search term and the second search term are combined to obtain the intelligent matching search results.

[0051] There can be multiple smart combination search results. For example, the first set of smart combination search results could be {first search item i, second search item i}, the second set of smart combination search results could be {first search item j, second search item j}, and so on. The priority ranking of smart combination search results can be determined based on the smart combination strategy. For example, items from the same store could be prioritized and displayed first as the first set of smart combination search results, followed by items from different stores but with the shortest delivery time. Other smart combination search results could also be displayed based on other factors influencing the combination. There are no restrictions on this.

[0052] S204: Display at least one of the above-mentioned intelligent matching search results.

[0053] In the multi-word intelligent matching mode, users only need to enter multiple query terms in the search box and click a button such as "search" to obtain at least one intelligent matching search result. The intelligent matching search result includes the combination of items corresponding to each query term, and then users can choose one of the multiple intelligent matching search results to place an order.

[0054] Again Figure 3For example, using the three search terms "fish," "chicken," and "beef," a smart combination search is performed, providing two sets of smart combination results. The first set includes steamed sea bass, braised chicken pieces, and tomato stewed beef brisket, all from the same store (Store A). The second set includes braised ribbonfish, braised chicken with pork belly, and braised beef, from two different stores (Store B and Store C). More smart combination search results can be displayed, but these are not detailed here. Furthermore, during the display of each smart combination search result, the total price, total delivery time, and customer reviews can be displayed for user reference. For instance, if a user confirms the selection of the first smart combination result, clicking the "Proceed to Checkout" button next to that result allows for a one-click purchase of all three items corresponding to the three search terms. Compared to traditional word searches, this shortens the search process and improves search efficiency.

[0055] The intelligent collocation mode of the multi-word search implementation method provided in this application embodiment has the following characteristics:

[0056] (1) The search box can perform multi-word searches, for example, by using spaces to distinguish them. After clicking the "Search" button, it will be anchored to the "Smart Matching" mode by default.

[0057] (2) Users who engage in intelligent matching are relatively sensitive to cost and time, and the characteristics of items can be reduced;

[0058] (3) Determine the intelligent matching logic based on the intelligent matching strategy:

[0059] (3.1) Prioritize launching products from the same store where multiple products are available, as this results in lower delivery costs and less impact on delivery capacity;

[0060] (3.2) Combining different stores, for example, focusing on shopping visit rate and delivery time; supply focusing on the same business district to reduce delivery time and cost;

[0061] (3.3) Other single indicator recommendations: such as lowest price, shortest time, etc.; you can also filter by filter options, such as displaying combination influencing factors on the interface, such as displaying purchase rate factor (most popular), delivery factor (shortest time), price factor (lowest total price), etc. Based on one or more combination influencing factors selected by the user, the items are matched with the selected combination influencing factors to obtain the matching results.

[0062] As can be seen, the multi-word search implementation method proposed in this application can satisfy users' one-click ordering needs. For example, in near-field e-commerce and food delivery scenarios, users hope to increase their ability to shop across stores in one stop. By searching for multiple products, they can find merchants that have supply at the same time and place an order. Alternatively, it can help users find a combination of stores that can satisfy their ordering needs at the same time (e.g., satisfying different dimensions such as shorter delivery time, lower delivery fee, and best cost performance), reducing the time users spend searching and filtering, reducing search links, and improving search efficiency.

[0063] As described above, the multi-word search implementation method provided in this application embodiment, in addition to providing a multi-word intelligent matching mode, also provides a multi-word self-matching mode. In the multi-word self-matching mode, users can independently select the items corresponding to each query term, thereby combining and ordering each selected item.

[0064] Therefore, based on the above-mentioned multi-word search implementation method, the following steps can also be included:

[0065] S1: In response to the command to switch to the multi-word auto-combination mode, send multiple query terms to the server and request an auto-combination search based on the multiple query terms;

[0066] S2: Receive at least one first search term returned by the server for the first query term and at least one second search term returned for the second query term, and in response to the selection instructions for the target first search term and the target second search term, determine the search results to be automatically combined.

[0067] S3: Display search results for custom combinations.

[0068] In one implementation, item pages can be displayed for each query term, and page switching can be performed by selecting a query term. Therefore, before the above-mentioned S2 "responding to the selection instruction of the first target search term and the second target search term", the following steps may also be included:

[0069] S2-0: Display a first selection page for at least one first search term and a second selection page for at least one second search term for the second query term, and control the switching between the first selection page and the second selection page by clicking on the first query term and the second query term;

[0070] S2-1 receives a selection instruction for selecting a first search item on the first selection page, and a selection instruction for selecting a second search item on the second selection page.

[0071] Specifically, S2-0, "controlling the switching between the first selection page and the second selection page by clicking the first and second query terms," ​​can include: after confirming the receipt of a selection instruction for the target first search term, when the second query term is clicked, switching to the second selection page to display at least one second search term. This involves determining the combined influence factors of the target first search term and determining the order of at least one second search term on the second selection page based on these combined influence factors. The combined influence factors include at least one of the following: same-store factor, purchase rate factor, delivery factor, and price factor. Furthermore, when at least one second search term is displayed on the second selection page, a prompt regarding the combined influence factors can be provided. In other words, based on the previously selected item, the order of items corresponding to subsequent query terms is dynamically adjusted, so that items with a strong dependency on the previously selected item are ranked first, and a prompt regarding the combined influence factors can be provided. For example, given the search terms "chicken" and "beef," when a user selects a "fried chicken nugget" item on the "chicken" item page and then switches to the "beef" item page, the multiple "beef" items are sorted according to the selected "fried chicken nugget" item. For instance, "beef" items from the same store as the "fried chicken nugget" item might be prioritized, with a message indicating they belong to the same store. Alternatively, "beef" items from the same business district or with the lowest delivery cost might be displayed first, along with relevant prompts. The advantage of this approach is that it strengthens the dependencies between different search terms, guiding users to choose the optimal combination.

[0072] See Figure 4 This illustration shows a schematic diagram of the self-matching mode in a multi-word search implementation method provided in this application embodiment. Users can switch from the default intelligent matching search mode to the self-matching search mode. Different item selection pages can be displayed for different query terms. For example, when performing a self-matching search based on the query terms "fish," "chicken," and "beef," the page displays multiple items related to "fish" for the user to select. For instance, clicking "+" adds the item to the cart. After selecting the "fish" item, the user can click "chicken," and the page switches to the "chicken" item selection page for similar selection; the same applies to "beef." In this case, such users value unique products, are relatively insensitive to cost, and have greater freedom in product selection; they do not know the delivery time and price before making a selection. When users add items to their cart, the cart calculates the overall price and delivery time. It can be understood that the item page display for a specific query term can simultaneously display prompts for each item's price, delivery time, and other information.

[0073] The multi-word search implementation method provided in this application offers users a solution for searching multiple items corresponding to multiple query terms at once. In the default intelligent matching search mode, it automatically provides intelligent matching search results based on the dependencies between multiple query terms, enabling one-click searching of multiple items, shortening the search process, and improving search efficiency. In addition to providing the intelligent matching search mode, it also offers a self-matching search mode. In this mode, users can obtain self-matching search results by selecting the items corresponding to each query term. Therefore, by combining the intelligent matching search mode and the self-matching search mode, the needs of different users in multi-word search scenarios can be met. The intelligent matching search mode addresses users' concerns regarding cost, delivery time, etc., while the self-matching search mode addresses users' needs regarding the items themselves, satisfying personalized matching requirements.

[0074] See Figure 5 The diagram illustrates a flowchart of a server-side multi-word search implementation method provided in an embodiment of this application, including the following steps:

[0075] S501: Receive multiple query terms sent by the terminal and confirm that the terminal requests to perform multi-word intelligent collocation search, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search items;

[0076] S502: Determine at least one intelligent collocation search result and return the at least one intelligent collocation search result to the terminal, wherein, according to the intelligent collocation strategy representing the dependency relationship between query terms, a first target search term and a second target search term are determined, and the intelligent collocation search result is obtained.

[0077] In one implementation, determining the target first search term and the target second search term according to the intelligent matching strategy includes: setting the intelligent matching strategy to a first strategy of prioritizing matching with items from the same store; and determining the target first search term and the target second search term belonging to the same store from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term according to the first strategy of prioritizing matching with items from the same store.

[0078] In one implementation, determining the target first search term and the target second search term according to the intelligent matching strategy includes: setting the intelligent matching strategy as: determining a second strategy for inter-store combination matching based on the combination influence factor; and determining the target first search term and the target second search term corresponding to the combination influence factor from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term according to the second strategy for inter-store combination matching.

[0079] In one implementation, determining the target first search term and target second search term corresponding to the combined influencing factors includes: setting the combined influencing factors to include at least one of a purchase rate factor, a delivery factor, a price factor, and a spatiotemporal factor; and determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term based on the purchase rate factor, the delivery factor, the price factor, and the spatiotemporal factor.

[0080] In one implementation, determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term based on at least one of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor includes: determining a current at least one combined influencing factor based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, and determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term based on the current at least one combined influencing factor; or, receiving the filtering results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor sent by the terminal, determining the current at least one combined influencing factor, and determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term based on the current at least one combined influencing factor.

[0081] In one implementation, the method further includes: confirming the terminal request to perform a multi-word auto-combination search; recalling at least one first search term for the first query term and at least one second search term for the second query term, and returning the at least one first search term and the at least one second search term to the terminal; controlling the terminal to display the at least one first search term and the at least one second search term in a page-by-page format, and providing selection buttons for the target first search term and the target second search term to determine the auto-combination search results.

[0082] In one implementation, the control terminal displays the at least one first search term and the at least one second search term in separate pages, including: controlling the terminal to display a first selection page for the first search term and a second selection page for the second search term, and controlling the first selection page and the second selection page to switch between them by clicking on the first search term and the second search term; controlling the terminal to receive a selection instruction for selecting the target first search term on the first selection page, and controlling the terminal to receive a selection instruction for selecting the target second search term on the second selection page.

[0083] In one implementation, controlling the switching between the first selection page and the second selection page by clicking on the first query term and the second query term includes: after determining that the terminal has received a selection instruction for the target first search term, when the second query term is clicked, controlling the terminal to switch to the second selection page to display at least one second search term, wherein a combined influence factor of the target first search term is determined, and the order of the at least one second search term on the second selection page is determined based on the combined influence factor.

[0084] In one implementation, the combined influencing factors include at least one of the following: same-store factor, purchase rate factor, delivery factor, and price factor; the method further includes: controlling the terminal to display the combined influencing factors when sorting and displaying the at least one second search item on the second selection page.

[0085] In one implementation, after receiving multiple query terms, the method further includes: identifying non-text special characters among the multiple query terms; determining whether the special character is a preset multi-word distinguisher; if so, distinguishing each query term with the multi-word distinguisher to obtain the first query term and the second query term.

[0086] For details on the specific implementation principles and methods of the intelligent matching mode and the autonomous matching mode, please refer to the aforementioned relevant information. Figure 1-4 The description is omitted here.

[0087] See Figure 6 The diagram illustrates a flowchart of a training method for a multi-word intelligent collocation prediction model provided in an embodiment of this application. The training method for this multi-word intelligent collocation prediction model includes the following steps:

[0088] S601: Obtain sample data and perform feature engineering on the sample data to obtain the input features of the prediction model: at least one of the following: user-side features, product-side features, query term-side features, logistics-side features, and spatiotemporal features;

[0089] S602: Based on at least one of user-end features, product features, query term features, logistics features, and spatiotemporal features, learn a neural network structure that shows the dependencies between multiple query terms, and obtain a click-through rate (CTR) prediction model.

[0090] Referring to the above Figure 1 This multi-word intelligent collocation prediction model can be implemented in the fine ranking module.

[0091] Specifically, step S602, "learning a neural network structure to obtain a click-through rate (CTR) prediction model based on at least one of user-end features, product features, query term features, logistics features, and spatiotemporal features," may further include:

[0092] S602-1: Based on user-side features and product features, train the first attention network to obtain the first fusion feature;

[0093] S602-2: Cross-link the first fusion feature with the query term features, product features, and logistics features to obtain cross features;

[0094] S602-3: Train a second attention network based on query term features and spatiotemporal features to obtain the second fusion feature;

[0095] S602-4: Based on the cross features and the second fusion features, train the third attention network to obtain the third fusion features;

[0096] S602-5: Based on the third fusion feature, a click-through rate (CTR) prediction model is trained.

[0097] Among them, user-side features include user profile features and user behavior sequences, while query term features include query term features and query term sequences. The query term sequence is a feature obtained by the intersection of query term features and spatiotemporal features.

[0098] The following is a specific example to illustrate the multi-word intelligent collocation prediction model.

[0099] See Figure 7 The diagram illustrates a training method for a multi-word intelligent collocation prediction model provided in an embodiment of this application.

[0100] The challenge of multi-word search lies in the fact that user intent is composed of multiple queries, and product combinations are influenced by the intersection of these queries. For example, a user searching for "claypot rice" and "cola" at their office on a weekday and "milk" and "cola" in their residential area on a weekend actually intends to buy a 330ml bottle of cola, while the latter wants a 6-pack of 330ml cola. Therefore, we model user intent based on query sequences and spatiotemporal information, introduce logistics signals (such as the number of stores included in the order, delivery fees before and after combination, and delivery time), and use an auxiliary loss function to model multi-word scenarios based on single-word scenarios.

[0101] The model structure typically includes the following parts:

[0102] 1. Embedding Layer

[0103] The bottom layer of the model is the feature embedding layer. Each input feature field in the embedding layer corresponds to an embedding matrix, which can transform high-dimensional sparse data into low-dimensional dense data. The embedding representation of the feature is obtained from the embedding matrix based on each feature value. The user behavior sequence and query sequence enter the first attention network (target attention) and the second attention network (multi-head attention). All other features are concatenated and then enter the fully connected layer (MLP layer). The input features mainly include the following categories:

[0104] (1) User characteristics: including user profile characteristics (such as user gender, whether the user is a member), and user behavior sequences, user behavior sequences ( 如是否多品下单 ， 是否短时间多次下单) Features such as...

[0105] (2) Query features: including query features (such as search term category, monthly exposure of search term, monthly purchase rate of search term, etc.) and query sequence. The query sequence is a feature obtained by the intersection of query features and spatiotemporal features.

[0106] (3) Product characteristics: such as monthly sales volume, monthly purchase rate, price, and image.

[0107] (4) Logistics characteristics: such as delivery method, delivery cost, delivery time, etc.

[0108] (5) Spatiotemporal characteristics: such as geohash, city ID, number of purchases made by users on geohash, time period, etc.

[0109] 2. Attention Network

[0110] This model architecture utilizes a multi-attention structure, including a first attention network for learning user historical behavior, a second attention network for learning query dependencies, and a third attention network for characterizing user preference for items under multiple intents. Both the first and second attention networks are located between the Embedding Layer and the MLP Layer. User behavior sequences and the items (targets) to be combined are embedded and fed into the first attention network to learn the correlation between the target and the behavior sequence. For query dependencies, the multi-attention mechanism learns the dependencies between different queries, and the captured user interest information is output to the MLP layer. Furthermore, in the second attention network, spatiotemporal features and some prior features can be added, and after shallow interaction, these can be directly used as feature representations in the output, while also acting as feature judges to filter out user interest representations that are more important to the target task. The third attention network, based on the cross-features between a single query and item and user intent, adds an attention layer to characterize the preference for individual items under multiple intents.

[0111] 3. Fully Linked Layer (MLP Layer)

[0112] After obtaining the user's embedding representation, the general task requires high-dimensional feature cross processing, which usually adopts a multi-layer MLP structure. At the same time, BatchNorm (an algorithm to accelerate convergence speed and stability) and the LRelu activation function can be added.

[0113] 4. Auxiliary Loss Module

[0114] Considering the problem of insufficient samples or cold scenarios for directly learning multiple queries to model the behavior between combined products, we can leverage the relatively abundant single query and product behavioral interaction data to assist model learning and ultimately improve the model's expressive power.

[0115] From a model training perspective, firstly, it's necessary to collect samples. In practice, there are no usable samples in the initial launch phase, so real samples can be accumulated over a certain period based on the strategy. Secondly, feature engineering is performed, mainly constructing user-side features, query-side features, product features, logistics features, and spatiotemporal features. User-side features, in addition to user profile data, also incorporate user behavior sequences to capture user behavior habits. Known knowledge guides the model to learn users' historical purchasing behavior, primarily focusing on order sequence features, such as whether multiple items are ordered or whether multiple orders are placed in a short period. Query-side features mainly consist of real-time multi-query sequences, using an attention network structure to learn the dependencies between queries. Finally, the model is built, introducing multiple attention mechanisms to learn personalized user behavior preferences and multi-query dependencies, and to capture spatiotemporal differences in user behavior. In practice, a neural network structure can be learned to obtain a CTR prediction model.

[0116] After training the CTR prediction model, when a user performs a multi-word search, the model can provide intelligent matching suggestions. First, it receives multiple query requests from the user's front end and determines the recall product set based on the recall model. Then, it constructs the inputs required by the model on the recall product set, such as user class features, query class features, product features, logistics features, spatiotemporal features, etc. The model will provide a fine-grained ranking prediction score for each product for filtering. Finally, it generates intelligent matching results according to the optimal combination strategy and transmits them to the user.

[0117] See Figure 8 This diagram illustrates a multi-word search implementation device located on a terminal according to an embodiment of this application. The multi-word search implementation device is used on the terminal side, and the method includes:

[0118] The query term determination unit 801 is used to determine multiple query terms received by the search box, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms;

[0119] The intelligent matching request unit 802 is used to send the multiple query words to the server in response to the confirmation instruction of the multi-word intelligent matching mode, and request intelligent matching search based on the multiple query words;

[0120] The intelligent matching confirmation unit 803 is used to receive at least one intelligent matching search result returned by the server, wherein the first target search term and the second target search term are determined according to the intelligent matching strategy that represents the dependency relationship between query terms, and the intelligent matching search result is obtained.

[0121] The intelligent matching display unit 804 is used to display the at least one intelligent matching search result.

[0122] In one implementation, the intelligent matching confirmation unit 803 is specifically used to confirm that the intelligent matching strategy is: a first strategy of prioritizing matching with items from the same store; and, based on the first strategy of prioritizing matching with items from the same store, to determine the target first search item and the target second search item belonging to the same store from at least one first search item corresponding to the first query term and at least one second search item corresponding to the second query term, respectively.

[0123] In one implementation, the intelligent matching confirmation unit 803 is specifically used to confirm that the intelligent matching strategy is: determining a second strategy for inter-store matching based on the combination influence factor; and determining, based on the second strategy for inter-store matching, the target first search term and the target second search term corresponding to the combination influence factor from at least one first search term corresponding to the first query term and at least one second search term corresponding to the second query term.

[0124] In one implementation, the intelligent matching confirmation unit 803 is specifically used to confirm that the combination influencing factors include at least one of the following: purchase rate factor, delivery factor, price factor, and spatiotemporal factor; and to determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively, based on the purchase rate factor, delivery factor, price factor, and spatiotemporal factor.

[0125] In one implementation, the intelligent matching confirmation unit 803 is specifically used to: determine at least one current combined influencing factor based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor; and determine corresponding target first search item and target second search item from the at least one first search item and the at least one second search item, respectively, based on the at least one current combined influencing factor; or, in response to the screening results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine the screened at least one current combined influencing factor; and determine corresponding target first search item and target second search item from the at least one first search item and the at least one second search item, respectively, based on the at least one current combined influencing factor.

[0126] One implementation also includes:

[0127] The self-matching request unit 805 is used to respond to the switching instruction of the multi-word self-matching mode, send the multiple query words to the server, and request self-matching search based on the multiple query words;

[0128] The autonomous matching confirmation unit 806 is used to receive at least one first search term returned by the server for the first query term and at least one second search term returned by the server for the second query term, and in response to the selection instructions of the target first search term and the target second search term, determine the autonomous matching search results.

[0129] The self-matching display unit 807 is used to display the self-matching search results.

[0130] One implementation also includes:

[0131] The switching control unit 808 is configured to display a first selection page for at least one first search term and a second selection page for at least one second search term for a second query term, and to control the switching between the first selection page and the second selection page by clicking on the first query term and the second query term; to receive a selection instruction for selecting a target first search term on the first selection page, and to receive a selection instruction for selecting a target second search term on the second selection page.

[0132] In one implementation, the switching control unit 808 is specifically configured to, after determining that a selection instruction for a target first search item has been received, switch to the second selection page to display at least one second search item when the second query term is clicked, wherein a combined influence factor of the target first search item is determined, and the order of the at least one second search item on the second selection page is determined based on the combined influence factor.

[0133] In one implementation, the combined influencing factors include at least one of the following: same-store factor, purchase rate factor, delivery factor, and price factor;

[0134] The device further includes an impact factor prompting unit 809, used to provide prompts for the combined impact factors when the at least one second search item is displayed in sorted order on the second selection page.

[0135] One implementation also includes:

[0136] The query term differentiation unit 810 is used to identify special characters that are not text in the search box; determine whether the special character is a preset multi-word separator; if so, differentiate each query term with the multi-word separator to obtain the first query term and the second query term.

[0137] See Figure 9 This diagram illustrates a multi-word search implementation device located on a server side, according to an embodiment of this application. The multi-word search implementation device is used on the server side and includes:

[0138] The query term receiving unit 901 is used to receive multiple query terms sent by the terminal, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms;

[0139] The intelligent collocation determination unit 902 is used to confirm the terminal request to perform a multi-word intelligent collocation search, and to determine at least one intelligent collocation search result and return it to the terminal. The intelligent collocation search result is obtained by determining the first target search term and the second target search term according to the intelligent collocation strategy that represents the dependency relationship between query terms.

[0140] In one implementation, the intelligent matching determination unit 902 is specifically used to set the intelligent matching strategy as a first strategy of prioritizing matching with items from the same store; and, based on the first strategy of prioritizing matching with items from the same store, determine a target first search term and a target second search term belonging to the same store from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term.

[0141] In one implementation, the intelligent matching determination unit 902 is specifically used to set the intelligent matching strategy as follows: determining a second strategy for inter-store combination matching based on the combination influence factor; and, based on the second strategy for inter-store combination matching, determining the target first search term and the target second search term corresponding to the combination influence factor from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term.

[0142] In one implementation, the intelligent matching determination unit 902 is specifically used to: set the combination influencing factors to include at least one of the following: purchase rate factor, delivery factor, price factor, and spatiotemporal factor; and determine corresponding target first search item and target second search item from the at least one first search item and the at least one second search item, respectively, based on the purchase rate factor, delivery factor, price factor, and spatiotemporal factor.

[0143] In one implementation, the intelligent matching determination unit 902 is specifically used to: determine at least one current combined influencing factor based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor; and determine corresponding target first search item and target second search item from the at least one first search item and the at least one second search item based on the current at least one combined influencing factor; or, receive the filtering results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor sent by the terminal, determine at least one current combined influencing factor, and determine corresponding target first search item and target second search item from the at least one first search item and the at least one second search item based on the current at least one combined influencing factor.

[0144] One implementation also includes:

[0145] The autonomous collocation determination unit 903 is used to confirm the terminal's request to perform a multi-word autonomous collocation search; and to recall at least one first search term for the first query term and at least one second search term for the second query term, and return the at least one first search term and the at least one second search term to the terminal, control the terminal to display the at least one first search term and the at least one second search term in a page, and provide selection buttons for the target first search term and the target second search term, and determine the autonomous collocation search results.

[0146] In one implementation, the autonomous matching determination unit 903 is specifically used to control the terminal to display a first selection page for a first search term and a second selection page for a second search term, and to control the switching between the first selection page and the second selection page by clicking on the first search term and the second search term; to control receiving a selection instruction for selecting the first search term on the first selection page, and to receive a selection instruction for selecting the second search term on the second selection page.

[0147] In one implementation, the autonomous combination determination unit 903 is specifically used to control the terminal to switch to the second selection page to display at least one second search item after determining that it has received a selection instruction for the target first search item, when the second query term is clicked. The terminal determines the combination influence factor of the target first search item and determines the order of the at least one second search item on the second selection page based on the combination influence factor.

[0148] In one implementation, the combined influencing factors include at least one of the following: same-store factor, purchase rate factor, delivery factor, and price factor;

[0149] The autonomous combination determination unit 903 is specifically used to control the terminal to provide prompts for the combination influence factors when the terminal displays the at least one second search item in the sorted order on the second selection page.

[0150] One implementation also includes:

[0151] The query term differentiation unit 904 is used to identify non-text special characters in the plurality of query terms; and to determine whether the special character is a preset multi-word separator. If so, the multi-word separator is used to differentiate each query term to obtain the first query term and the second query term.

[0152] See Figure 10This illustration shows a schematic diagram of the training device structure for a multi-word intelligent collocation prediction model provided in an embodiment of this application. The training device for the multi-word intelligent collocation prediction model includes:

[0153] The sample acquisition unit 1001 is used to acquire sample data;

[0154] The feature engineering processing unit 1002 is used to perform feature engineering processing on the sample data to obtain the input features of the prediction model: at least one of the following: user-side features, product-side features, query term-side features, logistics-side features, and spatiotemporal features;

[0155] The learning unit 1003 is used to learn a neural network structure with dependencies between multiple query terms based on at least one of the user-end features, product features, query term features, logistics features and spatiotemporal features, and obtain a click-through rate (CTR) prediction model.

[0156] In one implementation, the learning unit 1003 is specifically used to: train a first attention network based on user-end features and product features to obtain a first fusion feature; cross-link the first fusion feature with query term features, product features, and logistics features to obtain a cross feature; train a second attention network based on query term features and spatiotemporal features to obtain a second fusion feature; train a third attention network based on the cross feature and the second fusion feature to obtain a third fusion feature; and train a click-through rate (CTR) prediction model based on the third fusion feature.

[0157] In one implementation, the user-side features include user profile features and user behavior sequences, and the query term-side features include query term features and query term sequences, wherein the query term sequences are features obtained by the intersection of query term features and spatiotemporal features.

[0158] Embodiments of this application also provide a storage medium storing a computer program, wherein the computer program is configured to execute the steps in the above-described target method embodiments at runtime.

[0159] Optionally, in this embodiment, the storage medium can be configured to store data for performing the following... Figure 2 , Figure 5 or Figure 6 A computer program that performs the steps.

[0160] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0161] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in the above-described object method embodiments.

[0162] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0163] Optionally, in this embodiment, the processor can be configured to execute via a computer program. Figure 2 , Figure 5 or Figure 6 A computer program that performs the steps.

[0164] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0165] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0166] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0167] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, at least one unit or component may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0168] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across at least one network unit. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0170] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0171] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for implementing multi-word search, characterized in that, For use on the terminal side, the method includes: The search box receives multiple query terms, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; In response to the confirmation command of the multi-word intelligent collocation mode, the system sends the multiple query words to the server and requests intelligent collocation search based on the multiple query words; The system receives at least one intelligent matching search result returned by the server. Based on an intelligent matching strategy representing the dependencies between query terms, it determines a first target search term and a second target search term to obtain the intelligent matching search result. The intelligent matching strategy is either a first strategy prioritizing matching within the same store or a second strategy determining inter-store matching based on combination influence factors. These combination influence factors include at least one of visitor rate factors, delivery factors, price factors, and spatiotemporal factors. If a CTR prediction model is trained, when performing multi-word searches, the system determines an intelligent matching scheme based on the CTR prediction model and generates intelligent matching results that are then transmitted to the user. The CTR prediction model is trained based on a third fusion feature constructed using cross-features and second fusion features. The cross-features are obtained by cross-linking a first fusion feature obtained from user-side features and product features with query term-side features, product features, and logistics features. The second fusion feature is obtained based on the query term-side features and spatiotemporal features. The query term-side features include query term features and query term sequences, where the query term sequence is a feature obtained by crossing the query term features with the spatiotemporal features. Display at least one of the intelligent matching search results.

2. The method according to claim 1, characterized in that, The intelligent matching strategy based on the dependencies between query terms determines the first target search term and the second target search term, including: The intelligent pairing strategy is confirmed to be: prioritizing pairings from the same store as the primary strategy; According to the first strategy of prioritizing matching with the same store, target first search terms and target second search terms belonging to the same store are determined from at least one first search term corresponding to the first query term and at least one second search term corresponding to the second query term, respectively.

3. The method according to claim 1, characterized in that, The intelligent matching strategy based on the dependencies between query terms determines the first target search term and the second target search term, including: The intelligent matching strategy is confirmed to be: a second strategy for determining the combination and matching between stores based on combination influencing factors; Based on the second strategy for combining stores, the target first search term and the target second search term corresponding to the combined influence factor are determined from at least one first search term corresponding to the first query term and at least one second search term corresponding to the second query term.

4. The method according to claim 3, characterized in that, The determination of the target first search term and target second search term corresponding to the combined influence factors includes: Confirm the combined influencing factors; Based on at least one of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively.

5. The method according to claim 4, characterized in that, The step of determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively, based on at least one of the following factors: purchase rate factor, delivery factor, price factor, and time-space factor, includes: Based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine at least one combined influencing factor, and based on the at least one combined influencing factor, determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively; or, In response to the screening results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, at least one current combined influencing factor is determined, and based on the at least one current combined influencing factor, the corresponding target first search term and target second search term are determined from the at least one first search term and the at least one second search term, respectively.

6. The method according to claim 1, characterized in that, Also includes: In response to the instruction to switch to the multi-word auto-combination mode, the system sends the multiple query words to the server and requests an auto-combination search based on the multiple query words. Receive at least one first search term returned by the server for the first query term and at least one second search term returned for the second query term, and in response to the selection instructions of the target first search term and the target second search term, determine the search results to be automatically combined; Display the search results for the self-matching feature.

7. The method according to claim 6, characterized in that, Prior to the selection instruction in response to the first and second target search terms, the method further includes: A first selection page is displayed for a first query term, and a second selection page is displayed for a second query term, with at least one second search term. The first selection page and the second selection page are switched by clicking on the first query term and the second query term. The first selection page receives a selection instruction to select a first search item, and the second selection page receives a selection instruction to select a second search item.

8. The method according to claim 7, characterized in that, The control of switching between the first selection page and the second selection page by clicking the first query term and the second query term includes: After confirming that a selection instruction for the target first search term has been received, when the second query term is clicked, the system switches to the second selection page to display at least one second search term. In this process, a combined influence factor of the target first search term is determined, and the ranking of the at least one second search term on the second selection page is determined based on the combined influence factor.

9. The method according to claim 8, characterized in that, The combined influencing factors include at least one of the following: same-store factor, visit rate factor, delivery factor, and price factor; The method further includes: providing a prompt for the combined influence factors when the at least one second search item is displayed in sorted order on the second selection page.

10. The method according to any one of claims 1-9, characterized in that, After determining the multiple query terms received by the search box, the method further includes: Identify special non-text characters in the search box; Determine whether the special character is a preset multi-word distinguisher. If so, use the multi-word distinguisher to distinguish each query word and obtain the first query word and the second query word.

11. A method for implementing multi-word search, characterized in that, For use on the server side, the method includes: The receiver receives multiple query terms sent by the terminal and confirms the terminal's request to perform multi-word intelligent collocation search, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search items; At least one intelligent matching search result is determined and returned to the terminal. Specifically, a first target search term and a second target search term are determined based on an intelligent matching strategy that characterizes the dependencies between query terms, resulting in the intelligent matching search result. The intelligent matching strategy is either a first strategy prioritizing matching within the same store or a second strategy determining matching between stores based on combination influence factors. These combination influence factors include at least one of visitor rate factors, delivery factors, price factors, and spatiotemporal factors. If a CTR prediction model is trained, when performing multi-word searches, an intelligent matching scheme is determined based on the CTR prediction model, and intelligent matching results are generated and transmitted to the user. The CTR prediction model is trained based on a third fusion feature constructed using cross-features and second fusion features. The cross-features are obtained by cross-linking the first fusion feature obtained from user-side features and product features with query-side features, product features, and logistics features. The second fusion feature is obtained based on the query-side features and spatiotemporal features. The query-side features include query-term features and query-term sequences, where the query-term sequence is a feature obtained by crossing the query-term features and the spatiotemporal features.

12. The method according to claim 11, characterized in that, The intelligent matching strategy based on the dependencies between query terms determines the first target search term and the second target search term, including: The intelligent pairing strategy is set as follows: the first strategy is to prioritize pairing items from the same store. Based on the first strategy of prioritizing matching with the same store, target first search terms and target second search terms belonging to the same store are determined from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term.

13. The method according to claim 11, characterized in that, The intelligent matching strategy based on the dependencies between query terms determines the first target search term and the second target search term, including: The intelligent matching strategy is set as follows: a second strategy for determining the combination and matching between stores based on the combination influencing factors; According to the second strategy for combining stores, the target first search term and the target second search term corresponding to the combined influence factor are determined from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term.

14. The method according to claim 13, characterized in that, The determination of the target first search term and target second search term corresponding to the combined influence factors includes: Based on at least one of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively.

15. The method according to claim 14, characterized in that, The step of determining the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively, based on at least one of the following factors: purchase rate factor, delivery factor, price factor, and time-space factor, includes: Based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine at least one combined influencing factor, and based on the at least one combined influencing factor, determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively; or, The system receives the screening results of purchase rate factor, delivery factor, price factor, and spatiotemporal factor sent by the receiving terminal, determines at least one combined influencing factor, and determines the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively, based on the at least one combined influencing factor.

16. The method according to claim 11, characterized in that, Also includes: Confirm the terminal request to perform a multi-word auto-combination search; For the first query term, at least one first search term is recalled, and for the second query term, at least one second search term is recalled. The at least one first search term and the at least one second search term are returned to the terminal, which controls the terminal to display the at least one first search term and the at least one second search term in separate pages, and provides selection buttons for the target first search term and the target second search term, allowing users to determine the search results they want to combine.

17. The method according to claim 16, characterized in that, The control terminal displays the at least one first search item and the at least one second search item in a page-by-page format, including: The terminal is controlled to display a first selection page for a first search term and a second selection page for a second search term, and the first selection page and the second selection page are controlled to switch between the first selection page and the second selection page by clicking on the first search term and the second search term; The system controls the receiving of a selection instruction for a first search term on the first selection page, and the receiving of a selection instruction for a second search term on the second selection page.

18. The method according to claim 17, characterized in that, The control of switching between the first selection page and the second selection page by clicking the first query term and the second query term includes: After the terminal determines that it has received a selection instruction for the target first search term, when the second query term is clicked, it switches to the second selection page to display at least one second search term. The terminal determines the combined influence factor of the target first search term and determines the order of the at least one second search term on the second selection page based on the combined influence factor.

19. The method according to claim 18, characterized in that, The combined influencing factors include at least one of the following: same-store factor, visit rate factor, delivery factor, and price factor; The method further includes: when the terminal displays the at least one second search item in sorted order on the second selection page, prompting the combined influence factors.

20. The method according to any one of claims 11-19, characterized in that, After receiving multiple query terms, it also includes: Identify special non-text characters in the multiple query terms; Determine whether the special character is a preset multi-word distinguisher. If so, use the multi-word distinguisher to distinguish each query word and obtain the first query word and the second query word.

21. A training method for a multi-word intelligent collocation prediction model, characterized in that, include: Obtain sample data and perform feature engineering on the sample data to obtain the input features of the prediction model: at least one of the following: user-side features, product-side features, query term-side features, logistics-side features, and spatiotemporal features; Based on at least one of the user-end features, product features, query term features, logistics features, and spatiotemporal features, a neural network structure with dependencies between multiple query terms is learned to obtain a click-through rate (CTR) prediction model. When performing multi-word searches, the CTR prediction model determines a smart matching scheme and generates a smart matching result that is then transmitted to the user. The CTR prediction model is trained using a third fusion feature constructed using cross-features and a second fusion feature. The cross-features are obtained by cross-linking the first fusion feature obtained from user-end features and product features with the query term features, product features, and logistics features. The second fusion feature is obtained based on the query term features and spatiotemporal features. The query term features include query term features and query term sequences, where the query term sequence is a feature obtained by crossing the query term features with the spatiotemporal features.

22. The method according to claim 21, characterized in that, The step of learning a neural network structure based on at least one of the user-end features, product features, query term features, logistics features, and spatiotemporal features to obtain a click-through rate (CTR) prediction model includes: Based on user-side features and product features, train the first attention network to obtain the first fusion feature; The first fusion feature is cross-linked with the query term features, product features, and logistics features to obtain cross features; Based on query terminology features and spatiotemporal features, a second attention network is trained to obtain the second fusion feature; Based on the cross features and the second fusion features, a third attention network is trained to obtain the third fusion features; Based on the third fusion feature, a click-through rate (CTR) prediction model is trained.

23. The method according to claim 21 or 22, characterized in that, The user-side features include user profile features and user behavior sequences, and the query term features include query term features and query term sequences. The query term sequences are features obtained by crossing query term features with spatiotemporal features.

24. A multi-word search implementation device, characterized in that, For use on the terminal side, the device includes: The query term determination unit is used to determine multiple query terms received by the search box, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; The intelligent collocation request unit is used to respond to the confirmation instruction of the multi-word intelligent collocation mode, send the multiple query words to the server, and request intelligent collocation search based on the multiple query words; A smart matching confirmation unit is used to receive at least one smart matching search result returned by the server. Based on a smart matching strategy representing the dependencies between query terms, a first target search term and a second target search term are determined to obtain the smart matching search result. The smart matching strategy is either a first strategy prioritizing matching within the same store or a second strategy determining inter-store matching based on combination influence factors. These combination influence factors include at least one of a purchase rate factor, a delivery factor, a price factor, and a spatiotemporal factor. If a CTR prediction model is trained, when performing multi-word searches, a smart matching scheme is determined based on the CTR prediction model, and a smart matching result is generated and transmitted to the user. The CTR prediction model is trained based on a third fusion feature constructed using cross-features and second fusion features. The cross-features are obtained by cross-linking a first fusion feature obtained from user-side features and product features with query term-side features, product features, and logistics features. The second fusion feature is obtained based on the query term-side features and spatiotemporal features. The query term-side features include query term features and a query term sequence, where the query term sequence is a feature obtained by crossing the query term features with the spatiotemporal features. The intelligent matching display unit is used to display at least one intelligent matching search result.

25. The apparatus according to claim 24, characterized in that, The intelligent matching confirmation unit is specifically used to confirm that the intelligent matching strategy is: a first strategy of prioritizing matching with items from the same store; and, based on the first strategy of prioritizing matching with items from the same store, to determine the target first search item and the target second search item belonging to the same store from at least one first search item corresponding to the first query term and at least one second search item corresponding to the second query term, respectively.

26. The apparatus according to claim 24, characterized in that, The intelligent matching confirmation unit is specifically used to confirm that the intelligent matching strategy is: to determine a second strategy for inter-store matching based on the combination influence factor; and to determine the target first search term and target second search term corresponding to the combination influence factor from at least one first search term corresponding to the first query term and at least one second search term corresponding to the second query term based on the second strategy for inter-store matching.

27. The apparatus according to claim 26, characterized in that, The intelligent combination confirmation unit is specifically used to confirm the combination influencing factors; and to determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, based on at least one of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor.

28. The apparatus according to claim 27, characterized in that, The intelligent combination confirmation unit is specifically used to determine at least one combined influencing factor based on the priority of the purchase rate factor, delivery factor, price factor, and time and space factor, and to determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term based on the at least one combined influencing factor. Alternatively, in response to the screening results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor, determine the current at least one combined influencing factor that has been screened, and determine the corresponding target first search term and target second search term from the at least one first search term and the at least one second search term, respectively, based on the current at least one combined influencing factor.

29. The apparatus according to claim 24, characterized in that, Also includes: The self-matching request unit is used to respond to the switching instruction of the multi-word self-matching mode, send the multiple query words to the server, and request self-matching search based on the multiple query words; The self-matching confirmation unit is used to receive at least one first search term returned by the server for the first query term and at least one second search term returned by the server for the second query term, and in response to the selection instructions of the target first search term and the target second search term, determine the self-matching search results; The self-matching display unit is used to display the self-matching search results.

30. The apparatus according to claim 29, characterized in that, Also includes: A switching control unit is configured to display a first selection page for at least one first search term and a second selection page for at least one second search term for a second query term, and to control the switching between the first selection page and the second selection page by clicking on the first query term and the second query term; to receive a selection instruction for selecting a target first search term on the first selection page, and to receive a selection instruction for selecting a target second search term on the second selection page.

31. The apparatus according to claim 30, characterized in that, The switching control unit is specifically used to, after determining that a selection instruction for the target first search item has been received, switch to the second selection page to display at least one second search item when the second query term is clicked, wherein the combined influence factor of the target first search item is determined, and the order of the at least one second search item on the second selection page is determined based on the combined influence factor.

32. The apparatus according to claim 31, characterized in that, The combined influencing factors include at least one of the following: same-store factor, visit rate factor, delivery factor, and price factor; The device further includes an impact factor prompting unit, used to provide prompts for the combined impact factors when the at least one second search item is displayed in sorted order on the second selection page.

33. The apparatus according to any one of claims 25-32, characterized in that, Also includes: The query term differentiation unit is used to identify special characters that are not text in the search box; determine whether the special character is a preset multi-word separator; if so, differentiate each query term with the multi-word separator to obtain the first query term and the second query term.

34. A multi-word search implementation device, characterized in that, For use on the server side, the device includes: The query term receiving unit is used to receive multiple query terms sent by the terminal, wherein the multiple query terms include at least a first query term and a second query term with different semantics and corresponding to different search terms; The intelligent matching determination unit is used to confirm a terminal request for multi-word intelligent matching search, and to determine at least one intelligent matching search result and return it to the terminal. Specifically, based on an intelligent matching strategy representing the dependency relationship between query terms, a first target search term and a second target search term are determined to obtain the intelligent matching search result. The intelligent matching strategy is either a first strategy prioritizing matching within the same store or a second strategy determining inter-store matching based on combination influence factors. These combination influence factors include at least one of the following: purchase rate factor, delivery factor, price factor, and spatiotemporal factor. If a CTR prediction model is trained, then when performing… During multi-word search, the intelligent matching scheme is determined according to the CTR prediction model and the intelligent matching result is generated and passed to the user. The CTR prediction model is trained based on the third fusion feature constructed using cross features and second fusion features. The cross features are obtained by cross-linking the first fusion feature obtained based on user-end features and product features with query term features, product features, and logistics features. The second fusion feature is obtained based on the query term features and spatiotemporal features. The query term features include query term features and query term sequences. The query term sequences are features obtained by crossing the query term features and the spatiotemporal features.

35. The apparatus according to claim 34, characterized in that, The intelligent matching determination unit is specifically used to set the intelligent matching strategy as a first strategy of prioritizing matching with items from the same store; and, based on the first strategy of prioritizing matching with items from the same store, to determine the target first search item and the target second search item belonging to the same store from at least one first search item recalled for the first query term and at least one second search item recalled for the second query term.

36. The apparatus according to claim 34, characterized in that, The intelligent matching determination unit is specifically used to set the intelligent matching strategy as follows: determining a second strategy for inter-store matching based on the combination influence factor; and, based on the second strategy for inter-store matching, determining the target first search term and the target second search term corresponding to the combination influence factor from at least one first search term recalled for the first query term and at least one second search term recalled for the second query term.

37. The apparatus according to claim 36, characterized in that, The intelligent matching determination unit is specifically used to determine the corresponding target first search item and target second search item from the at least one first search item and the at least one second search item, respectively, based on at least one of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor.

38. The apparatus according to claim 37, characterized in that, The intelligent combination determination unit is specifically used to: determine at least one combined influencing factor based on the priority of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor; and determine the corresponding target first search item and target second search item from the at least one first search item and the at least one second search item based on the at least one combined influencing factor; or, receive the filtering results of the purchase rate factor, delivery factor, price factor, and spatiotemporal factor sent by the terminal, determine at least one combined influencing factor, and determine the corresponding target first search item and target second search item from the at least one first search item and the at least one second search item based on the at least one combined influencing factor.

39. The apparatus according to claim 34, characterized in that, Also includes: The self-matching determination unit is used to confirm the terminal's request to perform a multi-word self-matching search; Furthermore, for the first query term, at least one first search term is recalled, and for the second query term, at least one second search term is recalled. The at least one first search term and the at least one second search term are returned to the terminal, which controls the terminal to display the at least one first search term and the at least one second search term in separate pages, and provides selection buttons for the target first search term and the target second search term to determine the search results.

40. The apparatus according to claim 39, characterized in that, The autonomous matching determination unit is specifically used to control the terminal to display a first selection page for the first search term and a second selection page for the second search term, and to control the switching between the first selection page and the second selection page by clicking on the first search term and the second search term; to control the receiving of a selection instruction for the first search term on the first selection page, and to receive a selection instruction for the second search term on the second selection page.

41. The apparatus according to claim 40, characterized in that, The autonomous combination determination unit is specifically used to control the terminal to switch to the second selection page to display at least one second search item after receiving a selection instruction for the target first search item and when the second query term is clicked. The unit determines the combination influence factor of the target first search item and determines the order of the at least one second search item on the second selection page based on the combination influence factor.

42. The apparatus according to claim 41, characterized in that, The combined influencing factors include at least one of the following: same-store factor, visit rate factor, delivery factor, and price factor; The autonomous combination determination unit is specifically used to control the terminal to provide prompts for the combined influence factors when the terminal displays the at least one second search item in the sorted order on the second selection page.

43. The apparatus according to any one of claims 34-42, characterized in that, Also includes: The query term differentiation unit is used to identify special non-text characters among the multiple query terms; In addition, it is determined whether the special character is a preset multi-word distinguisher. If so, the multi-word distinguisher is used to distinguish each query word to obtain the first query word and the second query word.

44. A training device for a multi-word intelligent collocation prediction model, characterized in that, include: The sample acquisition unit is used to acquire sample data; The feature engineering processing unit is used to perform feature engineering processing on the sample data to obtain the input features of the prediction model: at least one of the following: user-side features, product-side features, query term-side features, logistics-side features, and spatiotemporal features; The learning unit is used to learn a neural network structure with dependencies between multiple query terms based on at least one of the user-end features, product features, query term features, logistics features, and spatiotemporal features, to obtain a click-through rate (CTR) prediction model. When performing a multi-word search, the CTR prediction model determines a smart matching scheme and generates a smart matching result that is then transmitted to the user. The CTR prediction model is trained using a third fusion feature constructed using cross-features and a second fusion feature. The cross-features are obtained by cross-linking a first fusion feature obtained from user-end features and product features with query term features, product features, and logistics features. The second fusion feature is obtained based on the query term features and spatiotemporal features. The query term features include query term features and query term sequences, where the query term sequence is a feature obtained by crossing the query term features with the spatiotemporal features.

45. The apparatus according to claim 44, characterized in that, The learning unit is specifically used to train a first attention network based on user-end features and product features to obtain a first fusion feature; and to cross-link the first fusion feature with query term features, product features, and logistics features to obtain cross features. Based on query term features and spatiotemporal features, a second attention network is trained to obtain a second fusion feature; based on the cross features and the second fusion feature, a third attention network is trained to obtain a third fusion feature; and based on the third fusion feature, a click-through rate (CTR) prediction model is trained.

46. ​​The apparatus according to claim 44 or 45, characterized in that, The user-side features include user profile features and user behavior sequences, and the query term features include query term features and query term sequences. The query term sequences are features obtained by crossing query term features with spatiotemporal features.

47. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1 to 23 when it is run.

48. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1 to 23.

Images