Search recommendation method and device, equipment and storage medium

Abstract

The invention relates to the technical field of Internet, and discloses a search recommendation method and device, equipment and a storage medium. The method comprises the following steps: searching a work information base according to a target search word to obtain a corresponding search result; obtaining a corresponding first set from a target recall library according to the search result; the first set comprises a plurality of first candidate words obtained based on the search result; obtaining a corresponding second set from a corpus according to the target search word; the second set comprises a plurality of second candidate words related to the target search word; based on the first set and the second set, obtaining a plurality of recommendation words of the target search word; and outputting the plurality of recommendation words. By adopting the method, the recommendation words which are possibly interested in can be recommended to the user in the search process of the user, the click rate of the user is increased, and the search experience of the user is improved.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a search recommendation method, apparatus, device, and storage medium. Background Art

[0002] With the rapid development of the internet, online works (such as videos, audiobooks, and novels) are now available on online platforms. Users enter their desired search information into the website's search area to obtain search results. During the user's search, the website often provides clickable recommendations based on relevant algorithms to meet the user's search needs.

[0003] A common search recommendation method involves using a search engine to retrieve relevant terms based on user input and displaying them as recommendations in the user interface. However, these recommendations often lack relevance to the user's interests and fail to meet their personalized needs. Consequently, the resulting recommendations have low click-through rates and poor user engagement. Therefore, finding ways to accurately recommend terms of potential interest to users and meet their personalized needs is a pressing issue. Summary of the Invention

[0004] In view of this, the present application aims to propose a search recommendation method, apparatus, device and storage medium to accurately recommend words that may be of interest to users and meet their personalized needs.

[0005] To achieve the above objectives, the technical solutions of this application are as follows:

[0006] A first aspect of an embodiment of the present application provides a search recommendation method, the method comprising:

[0007] Searching a work information database according to a target search term to obtain corresponding search results; the work information database is used to store at least one of the following: film and television works, graphic works, and audiovisual works;

[0008] According to the search results, a corresponding first set is obtained from a target recall library; the first set includes: a plurality of first candidate words obtained based on the search results;

[0009] According to the target search word, a corresponding second set is obtained from the corpus; the second set includes: a plurality of second candidate words related to the target search word;

[0010] Based on the first set and the second set, obtaining a plurality of recommended words for the target search word;

[0011] The plurality of recommended words are output.

[0012] Optionally, before obtaining the corresponding first set from the target recall library according to the search result, the method further includes:

[0013] Constructing prompt words based on all target objects in the work information library; the prompt words include: a task description for generating first candidate words, metadata of the target objects, and examples of the first candidate words; the task description includes: output weights of first candidate words of different dimensions; the metadata includes descriptive information of the target objects in multiple dimensions; generating a corresponding first candidate word set based on the prompt words for each target object using a large language model; the first candidate word set includes first candidate words of multiple dimensions;

[0014] All first candidate word sets are added to the target recall library and associated with corresponding target objects.

[0015] Optionally, after outputting the plurality of recommendation words, the method further includes:

[0016] Obtaining a first click-through rate of a first candidate word in each dimension of the plurality of recommended words; the first click-through rate is determined based on historical search behaviors of a plurality of users;

[0017] Adjust the output weight of the first candidate word in each dimension according to the first click rate of the first candidate word in each dimension; wherein the type with a higher first click rate corresponds to a higher output weight;

[0018] Based on the output weights of the first candidate words in each dimension, the prompt words of the large language model are updated so that the large language model adjusts the proportions of the first candidate words in each dimension in the first candidate word set based on the output weights.

[0019] Optionally, according to the search result, obtaining a corresponding first set from a target recall library includes:

[0020] Obtaining relevance between all target objects in the search results and the target search term, and determining a target object whose relevance is higher than a first threshold;

[0021] Obtaining second click-through rates of all target objects in the search results, and determining target objects whose second click-through rates are higher than a second threshold;

[0022] Based on the target object whose relevance is higher than a first threshold and the target object whose second click rate is higher than a second threshold, respectively obtaining corresponding first candidate word sets from the target recall library;

[0023] The first set is generated based on all acquired first candidate word sets.

[0024] Optionally, based on the first set and the second set, a plurality of recommended words for the target search word are obtained, including:

[0025] Merging the first set and the second set to obtain a third set;

[0026] Obtaining first click-through rates of all candidate words in the third set, and sorting the candidate words from high to low according to the first click-through rates to obtain a sorting result; the first click-through rates are determined based on historical search behaviors of multiple users;

[0027] A plurality of candidate words with a higher ranking order are determined from the sorting results as a plurality of recommended words for the target search word.

[0028] Optionally, the first set and the second set are combined to obtain a third set, comprising:

[0029] calculating a target number of first candidate words for ranking based on the fusion ratio and the number of second candidate words in the second set;

[0030] According to the target number, selecting a corresponding number of first candidate words from the first set;

[0031] The third set is generated based on the filtered first candidate words and the second candidate words in the second set.

[0032] Optionally, after sorting the candidate words from high to low according to the first click rate to obtain the sorting result, the method further includes:

[0033] Adjusting the order of at least one first candidate word within the first interval so that the adjusted order of the first candidate word is within the second interval; wherein the order corresponding to the first position in the first interval is after the order corresponding to the last position in the second interval;

[0034] The sorting result is updated based on the adjusted arrangement order of each first candidate word.

[0035] According to a second aspect of an embodiment of the present application, a search recommendation device is provided for implementing the steps of the method provided in the first aspect of the embodiment of the present application, the device comprising:

[0036] A search module is configured to search a work information database according to a target search term to obtain corresponding search results; the work information database is used to store at least one of the following: film and television works, graphic works, and audiovisual works;

[0037] A first recall module is configured to obtain a corresponding first set from a target recall library according to the search results; the first set includes: a plurality of first candidate words obtained based on the search results;

[0038] A second recall module is configured to obtain a corresponding second set from the corpus according to the target search term; the second set includes: a plurality of second candidate words related to the target search term;

[0039] a screening module configured to obtain a plurality of recommended words for the target search word based on the first set and the second set;

[0040] The display module is configured to output the multiple recommended words.

[0041] According to a third aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the method provided in the first aspect of the embodiment of the present application are implemented.

[0042] According to the fourth aspect of the embodiments of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the computer program is executed by the processor, the steps in the method provided in the first aspect of the embodiments of the present application are implemented.

[0043] The search recommendation method provided in this application obtains a first set of candidate words related to the search results from a target recall library based on the search results obtained by the search engine based on a search term input by a user. Furthermore, based on the search term input by the user, a second set of words related to the search term is obtained from a corpus. Recommended words ultimately displayed to the user are obtained based on the first and second sets.

[0044] Since search results are based on the search terms entered by the user, the first candidate terms recalled based on the search results are more consistent with the content that the user may be interested in. Compared with traditional search recommendation solutions, this application can generate personalized recommendations based on the user's current search content and accurately display personalized recommendations that may be of interest in the user interface, meeting the personalized needs of different users, reducing user input costs, increasing the platform's click-through rate, and improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0046] Figure 1 This is a flowchart of a search recommendation method proposed in one embodiment of the present application;

[0047] Figure 2 is a schematic diagram of a process for obtaining a first set in one embodiment of the present application;

[0048] Figure 3 This is a flowchart of generating recommended words in a film and television drama search scenario in one embodiment of the present application;

[0049] Figure 4 This is a flowchart of obtaining recommendation words in one embodiment of the present application;

[0050] Figure 5 is a schematic diagram of a search recommendation device proposed in an embodiment of the present application;

[0051] Figure 6 FIG. 1 is a schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0053] It should be understood that references throughout this specification to "one embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearances of "in one embodiment" or "in an embodiment" throughout this specification do not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0054] In the various embodiments of the present application, it should be understood that the size of the serial numbers of the following processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0055] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different figures represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with certain aspects as detailed herein.

[0056] It should be noted that, unless there is any conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0057] When searching based on user input, common website search engines typically use relevance algorithms to retrieve related phrases from a general corpus based on the current search term, and then display them as recommendations in the user interface. For example, when a user searches for "costume drama" on a video streaming platform, the website's search engine may retrieve recommended terms such as "costume romance drama" and "costume detective drama" from a general corpus. However, these recommendations are poorly aligned with the user's interests, resulting in low click-through rates and a poor user experience.

[0058] Furthermore, after obtaining search results, users may be interested in certain information within them and continue searching based on their interests. For example, on a video streaming platform, a user may be interested in the starring actors and plot of TV series A in the search results for the target search term "costume suspense drama" and may continue searching for information about the starring actors or plot details. However, traditional search recommendation solutions cannot accurately recommend content that users are interested in.

[0059] This embodiment recalls relevant recommendation words from the current search results, thereby accurately recommending personalized recommendation words that may be of interest to the user.

[0060] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0061] Figure 1 This is a flow chart of the search recommendation method proposed in one embodiment of the present application. Figure 1 As shown, the method includes:

[0062] S1: Searching a work information database according to a target search term to obtain corresponding search results; the work information database is used to store at least one of the following: film and television works, graphic works, and audiovisual works;

[0063] S2: according to the search results, obtain a corresponding first set from a target recall library; the first set includes: a plurality of first candidate words obtained based on the search results;

[0064] S3: according to the target search word, obtain a corresponding second set from the corpus; the second set includes: a plurality of second candidate words related to the target search word;

[0065] S4: Obtain multiple recommended words for the target search word based on the first set and the second set;

[0066] S5: Output the plurality of recommended words.

[0067] In this embodiment, the search recommendation method is applied to the search engine of the website. First, the search engine searches the work information library according to the target search term to obtain the corresponding search results. In actual applications, the work information library can be a film and television library of a video streaming platform, a multimedia database that provides audio-visual works, a library of pictures and texts provided by a literary website, or a retrieval website that provides comprehensive information queries, etc. When the search engine searches according to the target search term, there may be multiple search results obtained. Therefore, after obtaining the search results, it is necessary to obtain the corresponding first candidate word from the target recall library according to each search result. The first set includes multiple first candidate words related to the corresponding search results. For example, when a user enters the target search term "costume suspense drama" on the streaming platform, there are 3 search results obtained: TV series A, TV series B, and TV series C. Based on the 3 search results, the corresponding multiple first candidate words are obtained from the target recall library respectively.

[0068] The search engine also obtains a corresponding second set from a general corpus based on the target search term. In practical applications, the corpus can be a film and television classification tag library, a film and television term dictionary library, etc., and this application does not limit this. The specific steps for obtaining the corresponding second set from the general corpus based on the target search term are as follows:

[0069] First, the words in the corpus are traversed based on the target search word, the target search word is matched with the words in the corpus by similarity, and multiple second candidate words whose similarity with the target search word is greater than the similarity threshold are obtained by calculating the cosine similarity.

[0070] Then, data cleaning is performed on all the obtained second candidate words, including:

[0071] (1) Remove duplicates from the two second candidate words with high similarity and retain only one of them;

[0072] (2) Remove invalid words, such as "free resources", "download", etc.

[0073] Then, according to a preset number of candidate words, a plurality of second candidate words that are recalled more frequently in a recent period of time are screened out from the second candidate words after data cleaning.

[0074] Finally, a candidate word set is generated based on the screened multiple second candidate words as the second set corresponding to the target search word.

[0075] After obtaining the first set and the second set corresponding to the target search term, multiple recommended terms are obtained based on the first set and the second set. Then, when the search results based on the target search term are displayed to the user, the recommended terms are displayed together in the user interface for the user to click and view, so as to achieve personalized recommendations for the user.

[0076] In this embodiment, considering that the target search term can reflect the information that the user currently wants to know, the first candidate term recalled based on the search results of the target search term can meet the user's subsequent potential search needs. That is, the first candidate term recalled in this embodiment has a greater chance of matching the content that the user is interested in. Compared with the traditional method of directly using the literal meaning of the search term to recall recommended terms from the corpus, this embodiment determines the final recommended term for display based on the recalled first and second candidate terms, thereby achieving the accurate display of personalized recommended terms to the user and meeting the user's potential search needs for subsequent searches. Moreover, in the subsequent search process, the user can directly click on the recommended term to search without manual input, which reduces the user's input cost, increases the platform's click-through rate, and improves the user experience.

[0077] As an implementation of the present application, before the step S2 of "acquiring the corresponding first set from the target recall library according to the search results", the following steps are further included:

[0078] S001: Constructing prompt words based on all target objects in the work information library; the prompt words include: a task description for generating a first candidate word, metadata of the target object, and examples of the first candidate word; the task description includes: output weights of the first candidate word in different dimensions; the metadata includes descriptive information of multiple dimensions of the target object;

[0079] S002: Generate a corresponding first candidate word set according to the prompt word of each target object through a large language model; the first candidate word set includes first candidate words of multiple dimensions;

[0080] S003: Add all first candidate word sets to the target recall library and associate them with corresponding target objects.

[0081] In one embodiment, prompt words are pre-constructed based on the target objects in the work information library, and a large language model (LLM) is used to generate corresponding first candidate words based on the prompt words. On a video streaming platform, the work information library can be the platform's film and television library. Descriptive information on multiple dimensions of the target objects is extracted from the work information library as metadata. Then, prompt words are constructed based on the metadata, the task description, and examples of the first candidate words.

[0082] Taking the video streaming platform as an example, constructing prompt words based on the target object includes the following steps:

[0083] (1) Extracting descriptive information of multiple dimensions for each film and television work (target object) from the film and television library, including: introduction, genre, starring information, episode changes, etc. The extracted descriptive information of each film and television work is used as the metadata of the corresponding film and television work;

[0084] (2) Based on the semantic understanding ability of the large language model, a task description for generating first candidate words is constructed, and based on the task description and metadata, a small number of examples of first candidate words are constructed. The task description is used to guide the large language model to generate first candidate words of specified dimensions and quantity according to the metadata. In the embodiment of the present application, the task description includes the task objectives, the definition of the dimensions for generating the first candidate words, the distribution ratio of the output weights of each dimension, the output format requirements, etc. Initially, the output weight ratios of each dimension are equal and the sum is 1. That is, the large language model generates the first candidate words of each dimension in the same proportion. The examples of the first candidate words are generated based on the description information of each dimension in the metadata.

[0085] For example, based on the metadata of the TV series "ABC", the first candidate words that users may be interested in for each dimension are constructed as follows:

[0086] 1) Dimension 1 - Related candidate words for basic information: "Who is the main actor of ABC?", "What type of drama is ABC?", "How many episodes does ABC have?", etc.

[0087] 2) Dimension 2 - Related candidate words of similar types: "Dramas similar to ABC", "What dramas starred by ABC", etc.

[0088] 3) Dimension 3 - Plot-related candidate words: "What is the ending of ABC?", "Where did character xx go in the end?", etc.

[0089] (3) Based on the task description, metadata and examples of the first candidate word generated by the target object, a prompt word for inputting the large language model is constructed, that is, prompt word prompt = task description (system) + metadata (Meta) + a small number of examples (fewshot).

[0090] After obtaining the prompt word for the target object, the prompt word is input into the large language model. The large language model generates corresponding first candidate words for multiple dimensions. Based on all the first candidate words generated by the large language model for each target object, a first candidate word set corresponding to the target object is constructed. Then, the first candidate word set for each target object is added to the target recall library and associated with the corresponding target object.

[0091] In this embodiment, each first candidate word in the target recall library is updated at a specified time interval so that the target recall library is synchronized with the target object stored in the work information library, ensuring that valid first candidate words can be quickly recalled in real time during the user search process. When a target object is added to the work information library or deleted, the corresponding first candidate word set in the target recall library is updated synchronously in a timely manner to ensure the effectiveness and real-time nature of the user search recommendation. For example, when a target object is added to the work information library, the corresponding first candidate word set is added to the target recall library through synchronous updating, so that the relevant recommended words of the newly added target object can be displayed in the user interface in a timely manner, achieving the purpose of timely recommending new works and improving the user experience. When a target object is deleted from the work information library, the first candidate word set corresponding to the target object is deleted from the target recall library in a timely manner through synchronous updating, which can avoid invalid data occupying storage resources and improve recall efficiency.

[0092] After the user enters the target search term, the corresponding multiple first candidate words are recalled online from the target recall library through the search results to form a first set, which is used together with the second set formed by the second candidate words recalled from the corpus to determine the recommended words that are finally output.

[0093] Traditional recommendation word generation methods generate recommendations based on information in a corpus, fixed generation templates, user-entered target search terms, or structured parameters of target objects. These generated recommendations are insufficiently rich and flexible, and fail to align with user interests. For example, based on the target search term "costume drama" and the tag parameters "suspense drama" and "detective" for all costume dramas in a film and television library, templated recommendation terms such as "costume suspense drama" and "costume detective drama" are generated, failing to align with user interests. However, the method provided in this embodiment can better meet the user's personalized needs. For example, a user wants to find a currently popular costume drama but doesn't know the title of the drama (target object). Using this solution, the user enters the search term "costume drama," and the search engine searches the work information library for multiple costume dramas based on the target search term "costume drama." These search results include the popular costume drama "123" recently added to the work information library. Furthermore, a corresponding first candidate word set is retrieved based on each target object (costume drama) in the search results, and these are merged to generate a first set. That is, the first set includes the first candidate word for the currently popular costume drama "123," such as "123's highlights." Based on this, the final recommended word is determined based on the first and second sets, so that the user interface has a probability of displaying the first candidate word for the currently popular costume drama "123" (such as "123's highlights"), allowing the user to watch the currently popular costume drama "123" episode by clicking on the recommended word.

[0094] This embodiment uses a large language model to generate multi-dimensional, flexible, and rich first candidate words based on metadata and examples of different dimensions in advance, which can better reflect the user's potential interests, meet the user's personalized needs during the search process, and improve the user experience.

[0095] As an implementation manner of the present application, after the above step S5, the following is further included:

[0096] S61: Obtaining a first click-through rate of a first candidate word in each dimension among the multiple recommended words; the first click-through rate is determined based on historical search behaviors of multiple users;

[0097] S62: adjusting the output weight of the first candidate word in each dimension according to the first click rate of the first candidate word in each dimension; wherein the type with a higher first click rate corresponds to a higher output weight;

[0098] S63: Based on the output weights of the first candidate words in each dimension, updating the prompt words of the large language model so that the large language model adjusts the proportions of the first candidate words in each dimension in the first candidate word set based on the output weights.

[0099] In one embodiment, the proportion of first candidate words in each dimension output by the large language model is adjusted based on the user's first click-through rate on the recommended word. Since the large language model generates first candidate words for each dimension of the target object based on pre-built prompt words, not all first candidate words are of interest to the user.

[0100] Based on this, in order to improve the user experience, it is necessary to adjust the large language model to focus on generating the first candidate words of the dimensions that users are more interested in. In this embodiment, the first click-through rate of the first candidate words displayed as recommended words in the user interface is counted to determine the first click-through rate of the first candidate words of each dimension in the first candidate word set of the target object. The first click-through rate of the first candidate words of each dimension is determined based on the historical search behavior of multiple users. In actual applications, the multiple users counted may be active users or all users in the most recent first time period. For example, the multiple users counted may be active users of a video website in the past month, or may be all users of a video website. According to the first click-through rate of the first candidate words of each dimension, the output weights of the first candidate words of each dimension generated by the large language model are adjusted.

[0101] Specifically, we first count the number of times the first candidate word of each dimension is displayed on the user interface and the number of times users click on the first candidate word, and calculate the first click rate of the first candidate word of the dimension:

[0102] The first click-through rate of dimension A = the total number of clicks on the recommended words of dimension A ÷ the total number of impressions of the recommended words of dimension A.

[0103] The total number of impressions and clicks of the recommended words are obtained by counting the search behaviors of multiple users in the first time period.

[0104] For example, statistics are collected on the total number of times that the recommendation words of dimension A of the "three, four, five" of film and television dramas are displayed in the user interface of all users of the video streaming platform in the past month, as well as the total number of times users click on the recommendation words of dimension A, and then the first click-through rate of the recommendation words of dimension A of the "three, four, five" of film and television dramas is calculated.

[0105] Then, the first click-through rate of the first candidate word in each dimension is counted, and the corresponding output weight is adjusted according to the first click-through rate. The dimension with a higher first click-through rate means that the user is more interested in the information of this dimension, so the corresponding output weight needs to be increased so that the large language model focuses more on generating the first candidate word in this dimension. Specifically, according to the first click-through rate of the first candidate word in each dimension, the output weight of the first candidate word in each dimension is adjusted. The specific steps are as follows:

[0106] (1) Calculate the average click-through rate of all current dimensions:

[0107] Average click-through rate = the sum of the first click-through rates of all dimensions ÷ the number of dimensions;

[0108] (2) Based on the average click-through rate, calculate the adjustment factors for each dimension:

[0109] Adjustment factor for dimension A = CTR for dimension A ÷ average CTR;

[0110] (3) Based on the adjustment factors of each dimension and the current output weight, calculate the temporary weight:

[0111] Temporary weight of dimension A = current weight of dimension A × adjustment factor of dimension A;

[0112] (4) Normalize the temporary weights of all dimensions to obtain the adjusted output weights.

[0113] For example, initially, the weights of the first candidate words in the four dimensions are equal to 25%, and the first click-through rates of each dimension are 0.3, 0.2, 0.1, and 0.4, respectively. The average click-through rate is calculated to be 0.25. Based on the average click-through rate, the adjustment factors of the four dimensions are calculated as 1.2, 0.8, 0.4, and 1.6, respectively. Furthermore, based on the adjustment factors and the initial weight of 25%, the temporary weights are calculated as 30%, 20%, 10%, and 40%, respectively. Normalizing the temporary weights obtains the adjusted output weights of each dimension. In this example, since the sum of the temporary weights is 100%, the values ​​remain unchanged after normalization.

[0114] Finally, based on the output weights of the first candidate words in each dimension of each target object, the prompt words of the target object are updated, thereby adjusting the proportion of the first candidate words in each dimension of the target object generated by the large language model, further improving the user click-through rate and user experience.

[0115] As an implementation of the present application, in the above step S2, "obtaining the corresponding first set from the target recall library according to the search results" includes:

[0116] S21: Obtaining relevance between all target objects in the search results and the target search term, and determining a target object whose relevance is higher than a first threshold;

[0117] S22: Obtaining second click-through rates of all target objects in the search results, and determining a target object whose second click-through rate is higher than a second threshold;

[0118] S23: Based on the target object whose relevance is higher than a first threshold and the target object whose second click rate is higher than a second threshold, respectively obtain corresponding first candidate word sets from the target recall library;

[0119] S24: Generate the first set based on all acquired first candidate word sets.

[0120] In one embodiment, because the search results for the target search term contain a large number of target objects, retrieving first candidate terms for each target object in the search results would result in excessive data processing and high latency, making it impossible to display recommended terms in real time on the user interface. Therefore, by filtering some of the target objects in the search results to further obtain corresponding first candidate terms, we ensure that recommended terms can be displayed in a timely manner.

[0121] Specifically, the relevance between each target object in the search results and the target search term is calculated. Based on the relevance between each target object and the target search term, target objects with a relevance higher than a first threshold are selected from the search results as first candidate objects. Furthermore, the second click-through rate of each target object (e.g., TV series) in the adjacent first time period is obtained:

[0122] Second click rate of target object = historical clicks ÷ historical impressions;

[0123] The historical click counts and historical display counts are obtained by counting the search and click behaviors of all users in the first time period.

[0124] For example, the number of times users across the video streaming platform click on the TV series "123" within one month, as well as the number of times the TV series "123" is displayed on the user interface (including the browsing interface and the search interface), are counted, and the second click-through rate of the TV series "123" is then calculated.

[0125] The duration of the first time period can be set based on actual needs. Based on the second click-through rate of each target object within the first time period, target objects with a second click-through rate greater than a second threshold are filtered from the search results as second candidate objects. Based on the first and second candidate objects, a corresponding first set is retrieved from the target recall library.

[0126] Figure 2 This is a flow chart of obtaining the first set in one embodiment of the present application. Figure 2 As shown, the search engine first searches based on the target search term to obtain search results. Then, the relevance between each target object in the search results and the target search term, as well as the second click-through rate of each target object in the first time period, is calculated. First candidate target objects with a relevance higher than a first threshold and second candidate target objects with a second click-through rate higher than a second threshold are screened out. Based on the screened target objects (first candidate target objects and second candidate target objects), corresponding first candidate word sets are obtained from the target recall library respectively, and all the obtained first candidate word sets are merged to obtain the first set.

[0127] Optionally, if the number of first and second candidate target objects screened out exceeds a third threshold, third candidate target objects having a relevance higher than the first threshold and a second click-through rate higher than the second threshold are further screened out, and the corresponding first candidate word set is retrieved from the target recall library based on the third candidate target objects. In actual application scenarios, the third threshold can be set based on the search engine's data processing capabilities and the latency requirements for displaying recommended words in real time on the user interface.

[0128] In this embodiment, before recalling the first candidate word based on the search results, the target object in the search results is screened based on the relevance to the target search word and the second click-through rate, so as to increase the probability that the first candidate word finally displayed on the user interface is a recommended word of interest to the user, better provide personalized recommendations to the user, and increase the user's click probability.

[0129] As an implementation of the present application, the above step S4 includes:

[0130] S41: Merge the first set and the second set to obtain a third set;

[0131] S42: Obtaining first click rates of all candidate words in the third set, and sorting the candidate words from high to low according to the first click rates to obtain a sorting result; the first click rates are determined based on historical search behaviors of multiple users;

[0132] S43: Determine a plurality of candidate words with a higher ranking from the sorting results as a plurality of recommended words for the target search word.

[0133] In one embodiment, the first set obtained from the search results is merged with the second set to obtain a third set, and the candidate words in the third set are sorted from high to low according to their first click-through rates to obtain a sorted result. Then, based on the target number of recommended words that can be displayed on the user interface at one time, multiple candidate words that are ranked high in the sorted result are determined as recommended words and displayed on the user interface.

[0134] Optionally, a scoring and ranking method is used when ranking each candidate word. A refined ranking model is used to score and rank each candidate word based on its first click rate, and the ranking result is output. Then, based on the target number of terms that can be displayed at one time in the user interface, a corresponding number of candidate words with the highest ranking are selected for display. Optionally, the refined ranking model is constructed based on a deep neural network (DNN).

[0135] Figure 3 This is a flow chart of generating recommended words in a film and TV drama search scenario in one embodiment of the present application. Figure 3 As shown, a large language model is pre-generated based on the metadata of each film or television work (target object) in the film and television library using a large language model. This set of corresponding first candidate terms is then stored in a target recall library. When the search engine receives the target search term "XX" (film or television title) entered by the user, it searches the library based on the target search term, obtaining search results containing multiple target objects. Furthermore, target objects for recalling first candidate terms are selected based on the relevance of each target object to the target search term and the second click-through rate (CTR) of each target object within a first time period. Based on the selected target objects, the corresponding first candidate term set is retrieved from the target recall library, and all first candidate term sets are merged to obtain a first set. Simultaneously, multiple second candidate terms similar to the target search term are retrieved from the corpus using similarity matching to generate a second set. The first and second sets are merged to obtain a third set. The candidate terms in the third set are ranked from high to low based on their first CTR within the first time period. Based on the target number of recommended terms that can be displayed in the user interface at one time, the top-ranked candidate terms are selected as recommended terms and displayed in the user interface.

[0136] Optionally, when the user enters repeated target search terms, the recommended terms are displayed in a carousel manner. Specifically, based on the number of target terms that can be displayed at one time on the user interface, all candidate terms in the sorted third set are divided to generate multiple groups of recommended terms. When the user enters the target search term for the first time, the first group of recommended terms will be displayed on the user interface. When the user enters the target search term again, the next group of undisplayed recommended terms will be selected in sequence and displayed on the user interface. Furthermore, after the recommended terms are displayed on the user interface, it is detected whether the user has clicked on the currently displayed recommended terms. If the user has clicked on part of the currently displayed recommended terms, it means that the currently displayed recommended terms include content that the user is interested in. Based on this, when the user enters the same target search term again, the recommended terms that were displayed last time will continue to be displayed on the user interface, making it easier for the user to click and view, thereby improving the user experience.

[0137] As an implementation of the present application, the above step S41 includes:

[0138] S411: Calculating a target number of first candidate words for ranking based on the fusion ratio and the number of second candidate words in the second set;

[0139] S412: Filtering a corresponding number of first candidate words from the first set according to the target number;

[0140] S413: Generate the third set based on the screened first candidate words and the second candidate words in the second set.

[0141] In one embodiment, due to the excessive amount of data in the recalled first candidate words, the search engine's processing efficiency is low when sorting and filtering the recommendations, resulting in a significant delay in displaying the recommended words on the user interface, affecting the user experience. Therefore, in this embodiment, before sorting and filtering the candidate words, the first candidate words for sorting are pre-screened using the fusion ratio. In this embodiment, the fusion ratio = the number of first candidate words ÷ the number of second candidate words.

[0142] Figure 4 This is a flow chart of obtaining recommended words in one embodiment of the present application. Figure 4 As shown, according to the preset fusion ratio and the number of second candidate words in the second set, the target number of first candidate words for merging is calculated:

[0143] Target number = number of second candidate words in the second set × fusion ratio.

[0144] For example, when the fusion ratio is 0.5, the target number of first candidate words to be screened is half the number of second candidate words in the second set. Based on this target number, the target number of first candidate words is screened from all currently recalled first sets. Then, the target number of first candidate words that have been screened are merged with the second set to obtain a third set. The third set is then sorted and screened to obtain the final recommended words for display.

[0145] Optionally, the default value of the fusion ratio is set to 1, so that the number of first candidate words used for merging is equal to the number of second candidate words in the second set, ensuring that the number of two types of candidate words sorted in the initial stage is uniform. Furthermore, in the subsequent process of users using the search engine, the fusion ratio is fine-tuned based on the click-through rates of the first candidate words and the second candidate words displayed in the user interface, so that the search engine displays more recommended words in the dimensions that users are more interested in in the user interface. For example, in the case where the click-through rate of the first candidate word is higher, the fusion ratio is increased so that the first candidate word is used more for sorting and display in the user interface to increase the user's click-through rate.

[0146] Optionally, while increasing the fusion ratio, the number of second candidate words in the second set is simultaneously reduced, thereby controlling the total number of candidate words for ranking within an appropriate range to avoid increasing the load on the search engine and affecting the user experience.

[0147] As an implementation of the present application, after the step S42 of "ranking the candidate words from high to low according to the first click rate to obtain a ranking result", the following steps are further included:

[0148] S42-1: Adjusting the order of at least one first candidate word within the first interval so that the adjusted order of the first candidate word is within the second interval; wherein the order corresponding to the first position in the first interval is after the order corresponding to the last position in the second interval;

[0149] S42-2: Based on the adjusted arrangement order of each first candidate word, update the sorting result.

[0150] In traditional solutions, the recommended words displayed on the user interface are multiple second candidate words recalled from the corpus. Before display, a refined ranking model (deep neural network) is used to learn the characteristics of each second candidate word, and score each second candidate word based on its characteristics. The order in which each second candidate word is displayed on the user interface is determined from high to low according to the score.

[0151] In one embodiment, considering that the recommended words in the present application include the first candidate word and the second candidate word, the refined ranking model needs to score the first candidate word and the second candidate word together. However, since the refined ranking model has not learned the characteristics of the first candidate word before, the score of the first candidate word is relatively low, and the display order of the first candidate word is relatively late. In the case that the number of recommended words displayed on the user interface is limited, the first candidate word that is ranked at the end may not be displayed, which is not conducive to personalized recommendations for users. Therefore, in order to ensure that the recommended words containing the first candidate word can be displayed on the user interface every time the search results are displayed, after the ranking results are obtained by scoring through the refined ranking model, some of the first candidate words that are ranked relatively late are subjected to a power-up operation, and the arrangement order of these first candidate words is adjusted to a front position.

[0152] Specifically, based on the ranking results, a first interval with a lower order and a second interval with a higher order are determined. For example, the interval ranked 10-30 is used as the first interval, and the interval ranked 1-10 is used as the second interval. Multiple first candidate words are selected from the first interval and inserted into the second interval, and the adjusted ranking results are updated.

[0153] By increasing the weight of the first candidate words that are ranked lower, we can ensure that the first candidate words that may be of interest to users can be displayed when they search for the first time, thereby increasing the exposure rate of the first candidate words when the click rate of users is not high in the initial stage, and improving the user experience.

[0154] Optionally, multiple first candidate words in the first interval that are ranked lower are inserted into the second interval that is ranked higher, and a weighting method of randomly assigning weights is used. Specifically, according to the number of first candidate words that need to be weighted, a corresponding number of first candidate words are randomly screened from the first interval, and the adjusted relative ranking of each first candidate word is calculated by randomly assigning weights. Based on the adjusted relative ranking, each first candidate word is sequentially inserted into a plurality of randomly designated positions in the second interval, and the overall ranking result of the third set is updated.

[0155] In one embodiment, the updated sorting results are used as historical sorting data, and the parameters of the refined ranking model are adjusted based on the historical sorting data, so that the refined ranking model learns the sorting strategy of the first candidate word, increases the exposure weight of the first candidate word in the subsequent sorting process, and thus advances the sorting order of the first candidate word.

[0156] Based on the same inventive concept, an embodiment of the present application provides a search recommendation device. Figure 5 . Figure 5 FIG is a schematic diagram of a search recommendation device 100 proposed in an embodiment of the present application. Figure 5 As shown, the device includes:

[0157] Search module 101 is configured to search a work information database according to a target search term to obtain corresponding search results; the work information database is used to store at least one of the following: film and television works, graphic works, and audiovisual works;

[0158] The first recall module 102 is configured to obtain a corresponding first set from a target recall library according to the search results; the first set includes: a plurality of first candidate words obtained based on the search results;

[0159] The second recall module 103 is configured to obtain a corresponding second set from the corpus according to the target search term; the second set includes: a plurality of second candidate words related to the target search term;

[0160] A screening module 104 is configured to obtain a plurality of recommended words for the target search word based on the first set and the second set;

[0161] The display module 105 is configured to output the multiple recommended words.

[0162] As an embodiment of the present application, the device further includes a maintenance module configured to perform the following steps:

[0163] Constructing a prompt word based on all target objects in the work information library; the prompt word includes: a task description for generating a first candidate word, metadata of the target object, and an example of the first candidate word; the task description includes: output weights of the first candidate word in different dimensions; the metadata includes descriptive information of the target object in multiple dimensions;

[0164] Generate a corresponding first candidate word set according to the prompt word of each target object through a large language model; the first candidate word set includes first candidate words of multiple dimensions;

[0165] All first candidate word sets are added to the target recall library and associated with corresponding target objects.

[0166] As an implementation manner of the present application, after outputting the plurality of recommendation words, the maintenance module is further configured to perform the following steps:

[0167] Obtaining a first click-through rate of a first candidate word in each dimension of the plurality of recommended words; the first click-through rate is determined based on historical search behaviors of a plurality of users;

[0168] Adjust the output weight of the first candidate word in each dimension according to the first click rate of the first candidate word in each dimension; wherein the type with a higher first click rate corresponds to a higher output weight;

[0169] Based on the output weights of the first candidate words in each dimension, the prompt words of the large language model are updated so that the large language model adjusts the proportions of the first candidate words in each dimension in the first candidate word set based on the output weights.

[0170] As an embodiment of the present application, the first recall module 102 is configured to obtain a corresponding first set from a target recall library according to the search results, including:

[0171] Obtaining relevance between all target objects in the search results and the target search term, and determining a target object whose relevance is higher than a first threshold;

[0172] Obtaining second click-through rates of all target objects in the search results, and determining target objects whose second click-through rates are higher than a second threshold;

[0173] Based on the target object whose relevance is higher than a first threshold and the target object whose second click rate is higher than a second threshold, respectively obtaining corresponding first candidate word sets from the target recall library;

[0174] The first set is generated based on all acquired first candidate word sets.

[0175] As an embodiment of the present application, the screening module 104 is configured to obtain multiple recommended words for the target search word based on the first set and the second set, including:

[0176] Merging the first set and the second set to obtain a third set;

[0177] Obtaining first click-through rates of all candidate words in the third set, and sorting the candidate words from high to low according to the first click-through rates to obtain a sorting result; the first click-through rates are determined based on historical search behaviors of multiple users;

[0178] A plurality of candidate words with a higher ranking order are determined from the sorting results as a plurality of recommended words for the target search word.

[0179] As an embodiment of the present application, the screening module 104 is configured to merge the first set and the second set to obtain a third set, including:

[0180] calculating a target number of first candidate words for ranking based on the fusion ratio and the number of second candidate words in the second set;

[0181] According to the target number, selecting a corresponding number of first candidate words from the first set;

[0182] The third set is generated based on the filtered first candidate words and the second candidate words in the second set.

[0183] As an embodiment of the present application, after sorting the candidate words from high to low according to the first click rate to obtain the sorting result, the screening module 104 is further configured to perform the following steps:

[0184] Adjusting the order of at least one first candidate word within the first interval so that the adjusted order of the first candidate word is within the second interval; wherein the order corresponding to the first position in the first interval is after the order corresponding to the last position in the second interval;

[0185] The sorting result is updated based on the adjusted arrangement order of each first candidate word.

[0186] Based on the same inventive concept, an embodiment of the present application provides a readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the search recommendation method described in any of the above embodiments of the present application are implemented.

[0187] Based on the same inventive concept, an embodiment of the present application provides an electronic device, referring to Figure 6 , Figure 6 Schematic diagram of an electronic device according to one embodiment of the present application. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of the search recommendation method described in any of the above embodiments of the present application.

[0188] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0189] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0190] For the sake of simplicity, the method embodiments are described as a series of action combinations. However, those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and components involved are not necessarily required by this application.

[0191] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the embodiments of the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0193] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0195] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the underlying inventive concepts. Therefore, this application is intended to include the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0196] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

[0197] The above is a detailed introduction to the search recommendation method, device, equipment and storage medium provided by this application. Specific examples are used in this article to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method of this application and its core ideas. At the same time, for general technical personnel in this field, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on this application.

Claims

1. A search recommendation method, characterized in that: include: Search the work information database according to the target search term to obtain the corresponding search results; The work information database is used to store at least one of the following: film and television works, graphic works, and audio-visual works; According to the search results, obtaining a corresponding first set from a target recall library; The first set includes: a plurality of first candidate words obtained based on the search results; According to the target search word, a corresponding second set is obtained from the corpus; the second set includes: a plurality of second candidate words related to the target search word; Based on the first set and the second set, obtaining a plurality of recommended words for the target search word; The plurality of recommended words are output.

2. The search recommendation method according to claim 1, characterized in that: Before obtaining the corresponding first set from the target recall library according to the search result, the method further includes: Constructing a prompt word based on all target objects in the work information library; the prompt word includes: a task description for generating a first candidate word, metadata of the target object, and an example of the first candidate word; the task description includes: output weights of the first candidate word in different dimensions; the metadata includes descriptive information of the target object in multiple dimensions; Generate a corresponding first candidate word set according to the prompt word of each target object through a large language model; the first candidate word set includes first candidate words of multiple dimensions; All first candidate word sets are added to the target recall library and associated with corresponding target objects.

3. The search recommendation method according to claim 2, characterized in that After outputting the plurality of recommendation words, the method further includes: Obtaining a first click-through rate of a first candidate word in each dimension of the plurality of recommended words; the first click-through rate is determined based on historical search behaviors of a plurality of users; Adjust the output weight of the first candidate word in each dimension according to the first click rate of the first candidate word in each dimension; wherein the type with a higher first click rate corresponds to a higher output weight; Based on the output weights of the first candidate words in each dimension, the prompt words of the large language model are updated so that the large language model adjusts the proportions of the first candidate words in each dimension in the first candidate word set based on the output weights.

4. The search recommendation method according to claim 1, wherein: According to the search results, a corresponding first set is obtained from the target recall library, including: Obtaining relevance between all target objects in the search results and the target search term, and determining a target object whose relevance is higher than a first threshold; Obtaining second click-through rates of all target objects in the search results, and determining target objects whose second click-through rates are higher than a second threshold; Based on the target object whose relevance is higher than a first threshold and the target object whose second click rate is higher than a second threshold, respectively obtaining corresponding first candidate word sets from the target recall library; The first set is generated based on all acquired first candidate word sets.

5. The search recommendation method according to claim 1, wherein: Based on the first set and the second set, a plurality of recommended words for the target search word are obtained, including: Merging the first set and the second set to obtain a third set; Obtaining first click-through rates of all candidate words in the third set, and sorting the candidate words from high to low according to the first click-through rates to obtain a sorting result; the first click-through rates are determined based on historical search behaviors of multiple users; A plurality of candidate words with a higher ranking order are determined from the sorting results as a plurality of recommended words for the target search word.

6. The search recommendation method according to claim 5, characterized in that: The first set and the second set are combined to obtain a third set, including: Calculating a target number of first candidate words for ranking based on the fusion ratio and the number of second candidate words in the second set; According to the target number, selecting a corresponding number of first candidate words from the first set; The third set is generated based on the filtered first candidate words and the second candidate words in the second set.

7. The search recommendation method according to claim 5, characterized in that: After sorting the candidate words from high to low according to the first click rate and obtaining the sorting results, the following steps are also included: Adjusting the order of at least one first candidate word within the first interval so that the adjusted order of the first candidate word is within the second interval; wherein the order corresponding to the first position in the first interval is after the order corresponding to the last position in the second interval; The sorting result is updated based on the adjusted arrangement order of each first candidate word.

8. A search recommendation device, characterized in that: Used to perform the method according to any one of claims 1 to 7, comprising: A search module is configured to search a work information database according to a target search term to obtain corresponding search results; the work information database is used to store at least one of the following: film and television works, graphic works, and audiovisual works; A first recall module is configured to obtain a corresponding first set from a target recall library according to the search results; the first set includes: a plurality of first candidate words obtained based on the search results; A second recall module is configured to obtain a corresponding second set from the corpus according to the target search term; the second set includes: a plurality of second candidate words related to the target search term; a screening module configured to obtain a plurality of recommended words for the target search word based on the first set and the second set; The display module is configured to output the multiple recommended words.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images