Training method of search term generation model and search term generation method and device

Abstract

The invention provides a training method of a search term generation model and a search term generation method and device, and relates to the technical field of computers, in particular to the application fields of machine learning, large language models, multi-modal models, generative search, search engines, information retrieval, visual content search and the like. The specific implementation scheme is as follows: acquiring target training data; wherein the target training data comprises a target visual content description text and a target search word tag; utilizing a search term generation model to describe the text based on the target visual content to obtain a search term reasoning result; and based on the target search term tag and the search term reasoning result, training a search term generation model to obtain a trained search term generation model.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to application fields such as machine learning, large language models, multimodal models, generative search, search engines, information retrieval, and visual content search, and specifically to a training method for a search term generation model, a search term generation method, and a device. Background Art

[0002] Search engine advertising is a marketing method in which advertisers use the advertising platform provided by search engines to display product advertisements to users. Its main purpose is that when users enter a specific search term through a search engine, the advertising system can intelligently display product advertisements that are associated with the specific search term and are relevant to them through a built-in content recommendation strategy, thereby increasing the market exposure of the product and promoting the growth of product sales. In traditional advertising systems, advertisers usually need to select search terms themselves in order to establish an association between these search terms and their own product advertisements. Summary of the invention

[0003] The present disclosure provides a method for training a search term generation model, a method for generating a search term, and a device.

[0004] According to a first aspect of the present disclosure, a method for training a search term generation model is provided, comprising:

[0005] Obtain target training data; wherein the target training data includes target visual content description text and target search term labels;

[0006] Using the search term generation model, the text is described based on the target visual content to obtain the search term inference result;

[0007] Based on the target search term label and the search term inference result, the search term generation model is trained to obtain a trained search term generation model.

[0008] According to a second aspect of the present disclosure, a search term generation method is provided, comprising:

[0009] Get the visual content description text;

[0010] The trained search term generation model is used to obtain search term generation results based on the visual content description text.

[0011] According to a third aspect of the present disclosure, a training device for a search term generation model is provided, comprising:

[0012] A training data acquisition unit, used to acquire target training data; wherein the target training data includes a target visual content description text and a target search term label;

[0013] An inference result acquisition unit, used to obtain a search word inference result based on a target visual content description text by using a search word generation model;

[0014] The model training unit is used to train the search term generation model based on the target search term label and the search term inference result to obtain a trained search term generation model.

[0015] According to a fourth aspect of the present disclosure, a search term generating device is provided, comprising:

[0016] A description text acquisition unit, used to acquire a description text of the visual content;

[0017] The search word generation unit is used to obtain a search word generation result based on the visual content description text by using a trained search word generation model.

[0018] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0019] at least one processor;

[0020] a memory communicatively coupled to the at least one processor;

[0021] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method provided in the first aspect of the present disclosure.

[0022] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method provided according to the first aspect of the present disclosure.

[0023] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program, which implements the method provided according to the first aspect of the present disclosure when executed by a processor.

[0024] The present disclosure can realize automatic generation of search terms and ensure the matching degree between the search term generation results and specific visual content (for example, online advertisements).

[0025] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0027] Figure 1A flowchart of a method for training a search term generation model provided in an embodiment of the present disclosure;

[0028] Figure 2 A diagram illustrating a selection process of candidate training data provided by an embodiment of the present disclosure;

[0029] Figure 3 A diagram illustrating another process of selecting candidate training data provided by an embodiment of the present disclosure;

[0030] Figure 4 A diagram illustrating another process of selecting candidate training data provided by an embodiment of the present disclosure;

[0031] Figure 5 A flowchart of a search term generation method provided by an embodiment of the present disclosure;

[0032] Figure 6 A diagram for explaining a method for training a search term generation model and an application process of the trained search term generation model provided in an embodiment of the present disclosure;

[0033] Figure 7 A schematic diagram of an application scenario of a method for training a search term generation model provided in an embodiment of the present disclosure;

[0034] Figure 8 A schematic diagram of an application scenario of a search term generation method provided by an embodiment of the present disclosure;

[0035] Fig. 9 A schematic structural block diagram of a training device for a search term generation model provided in an embodiment of the present disclosure;

[0036] Fig.10 A schematic structural block diagram of a search term generating device provided in an embodiment of the present disclosure;

[0037] Fig.11 A schematic structural block diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0038] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0039] As mentioned in the background technology, search engine advertising is a marketing method in which advertisers use the advertising platform provided by search engines to display product advertisements to users. Its main purpose is that when users enter a specific search term through a search engine, the advertising delivery system can intelligently display product advertisements that are associated with the specific search term and have high relevance to them through a built-in content recommendation strategy, thereby increasing the market exposure of the product and promoting the growth of product sales.

[0040] In traditional advertising systems, advertisers usually need to select search terms by themselves in order to establish associations between these search terms and their own product advertisements. This not only consumes a lot of human resources, but also, in order to increase advertising coverage, some advertisers have the phenomenon of "randomly selecting terms" and / or "selecting multiple terms", which will greatly reduce the matching degree between search terms and product advertisements. When the matching degree between search terms and product advertisements is low, the user's search satisfaction will also decrease, and ultimately, there will be a risk of losing search engine users.

[0041] In view of the above problems, the embodiments of the present disclosure provide a method for training a search term generation model, which can be applied to electronic devices. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing unit or other similar computing devices. Figure 1 The flowchart diagram shown illustrates a method for training a search term generation model provided by an embodiment of the present disclosure. It should be noted that although a logical order is shown in the flowchart diagram, in some cases, the steps shown or described in the flowchart can also be performed in other orders.

[0042] Step S101, obtaining target training data.

[0043] The target training data may include target visual content description text and target search term labels. Here, the target visual content description text may be text content used to describe the target visual content sample; the target search term label may be a predetermined search term with a high degree of matching with the target visual content sample.

[0044] In addition, in the embodiments of the present disclosure, the target visual content samples may be online advertisements, video resources, graphic resources, etc. Among them, online advertisements may be static advertisements, animated image advertisements, video advertisements, etc.; graphic resources may be static graphics, dynamic graphics, etc., which are not limited in the embodiments of the present disclosure.

[0045] Step S102, using the search word generation model, based on the target visual content description text, obtains the search word inference result.

[0046] The search term generation model may be a large language model. Here, the large language model may be a pre-trained neural network model that has general language knowledge, world knowledge, expertise in various fields, etc., and is stored in the model in the form of parameters. In one example, the large language model may be an autoregressive generation model of a Transformer architecture.

[0047] Step S103: training the search term generation model based on the target search term label and the search term inference result to obtain a trained search term generation model.

[0048] In one example, after obtaining the target search term label and the search term inference result, a target loss value between the target search term label and the search term inference result can be calculated, and based on the target loss value, the search term generation model is trained to obtain a trained search term generation model. For example, based on the target loss value, the model parameters of the search term generation model are adjusted to obtain a trained search term generation model.

[0049] By adopting the training method of the search term generation model provided by the embodiment of the present disclosure, it is possible to obtain target training data including target visual content description text and target search term labels, and use the search term generation model to obtain search term inference results based on the target visual content description text, and then train the search term generation model based on the target search term labels and the search term inference results to obtain a trained search term generation model. Thereafter, the trained search term generation model can be used to obtain search term generation results. For example, the visual content description text can be obtained, and the trained search term generation model can be used to obtain search term generation results based on the visual content description text. Compared with the prior art, which manually selects keywords, this can not only save a lot of human resources, but also avoid human bias, thereby ensuring the matching degree of the search term generation results with specific visual content (for example, online advertising).

[0050] It should be noted that, in the embodiment of the present disclosure, step S101, step S102 and step S103 can be executed repeatedly until the search term generation model meets the convergence condition to obtain the final trained search term generation model. The convergence condition may be that the target loss value is less than or equal to the preset loss value. Here, the preset loss value can be set according to actual business needs or application scenarios, and the embodiment of the present disclosure does not limit this.

[0051] Further, in some optional implementations, step S101, i.e., “obtaining target training data” may include:

[0052] Acquire multiple original training data; wherein each of the multiple original training data includes original visual content description text and original search term label;

[0053] According to a relevance selection strategy and / or a diversity selection strategy, multiple candidate training data are selected from multiple original training data; wherein each candidate training data in the multiple candidate training data includes a candidate visual content description text and a candidate search term label;

[0054] Based on the plurality of candidate training data, target training data is determined.

[0055] Among them, the original visual content description text and the candidate visual content description text have the same data structure as the target visual content description text, which is not elaborated in the present embodiment; correspondingly, the original search term label and the candidate search term label have the same data structure as the target search term label, which is not elaborated in the present embodiment.

[0056] In addition, it should be noted that in the disclosed embodiment, historical visual content samples and original search word labels with corresponding relationships can be obtained from the advertising delivery system, and the historical visual content samples can be summarized to obtain the original visual content description text, and then the original visual content description text and the original search word labels can be used together as original training data. Among them, "summarizing the historical visual content samples to obtain the original visual content description text" can be: using a large language model, based on the inductive prompt words, summarizing the historical visual content samples to obtain the original visual content description text.

[0057] Repeat the above steps to obtain multiple original training data.

[0058] It should also be noted that in the embodiment of the present disclosure, the relevance selection strategy is intended to ensure that each of the multiple candidate training data selected has a greater content relevance. The content relevance of the candidate training data is used to characterize the relevance between the candidate visual content description text and the candidate search term label in the candidate training data. The diversity selection strategy is intended to ensure that the multiple candidate training data selected have data diversity in a specified dimension. The specified dimension can be a search term performance evaluation result and / or a search term field.

[0059] After selecting multiple candidate training data from multiple original training data according to the relevance selection strategy and / or the diversity selection strategy, target training data can be determined based on the multiple candidate training data. For example, each candidate training data in the multiple candidate training data can be used as target training data.

[0060] Through the above method, in the embodiment of the present disclosure, multiple original training data can be obtained, and multiple candidate training data can be selected from the multiple original training data according to the relevance selection strategy and / or the diversity selection strategy, and then the target training data is determined based on the multiple candidate training data. In this way, at least it can be ensured that each of the multiple candidate training data selected has a greater content relevance, or that the multiple candidate training data selected have data diversity of a specified dimension, that is, at least the quality of the multiple candidate training data can be ensured, or the diversity of the multiple candidate training data can be ensured. In this case, after the target training data is determined based on the multiple candidate training data and used to train the search term generation model to obtain the trained search term generation model, the performance and / or generalization of the trained search term generation model can be ensured. In this way, when the search term generation result is obtained by using the trained search term generation model, the matching degree of the search term generation result with the specific visual content (for example, online advertising) can be further ensured.

[0061] In the above implementation, “selecting multiple candidate training data from multiple original training data according to a relevance selection strategy and / or a diversity selection strategy” may include:

[0062] According to the relevance selection strategy, a plurality of first preliminary selected training data are selected from the plurality of original training data; wherein each of the plurality of first preliminary selected training data comprises a first preliminary selected visual content description text and a first preliminary selected search term label;

[0063] According to the diversity selection strategy, a plurality of candidate training data are selected from a plurality of first preliminary selected training data.

[0064] Among them, the first preliminary visual content description text has the same data structure as the target visual content description text, which is not elaborated in the embodiment of the present disclosure; correspondingly, the first preliminary search term label has the same data structure as the target search term label, which is not elaborated in the embodiment of the present disclosure.

[0065] Through the above method, in the embodiment of the present disclosure, a plurality of first preliminary selected training data can be first selected from a plurality of original training data according to the relevance selection strategy, and then a plurality of candidate training data can be selected from the plurality of first preliminary selected training data according to the diversity selection strategy. In other words, the relevance selection strategy and the diversity selection strategy can be implemented serially. In this way, it can be ensured that each of the plurality of selected candidate training data has a greater content relevance, and it can be ensured that the plurality of selected candidate training data has a diversity of specified dimensions. Therefore, after determining the target training data based on the plurality of candidate training data and using it to train the search term generation model to obtain the trained search term generation model, the performance and generalization of the trained search term generation model can be further ensured.

[0066] In one example, “selecting a plurality of first preliminary selected training data from a plurality of original training data according to a relevance selection strategy” may include:

[0067] Obtaining content relevance of each original training data among the plurality of original training data;

[0068] A plurality of first preliminary selected training data with the greatest content relevance are selected from the plurality of original training data; or, a plurality of first preliminary selected training data with content relevance meeting a preset relevance requirement are selected from the plurality of original training data.

[0069] The content relevance of the original training data is used to characterize the relevance between the original visual content description text and the original search term label in the original training data.

[0070] In a specific example, the content relevance of each of the multiple original training data can be obtained by a similarity algorithm, wherein the similarity algorithm can be a cosine similarity algorithm.

[0071] In another specific example, a similarity discrimination model can be used to obtain the content relevance of each of the multiple original training data. The similarity discrimination model can be a large language model or a trained dedicated similarity discrimination model. Here, the dedicated similarity discrimination model can be an autoregressive generative model of a Transformer architecture or a twin neural network, for example, a Siamese model.

[0072] After obtaining the content relevance of each of the multiple original training data, N original training data with the largest content relevance can be selected from the multiple original training data as the first preliminary training data. Wherein, N ≥ 2 and is a positive integer. Here, the specific value of N can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure are not limited to this. After obtaining the content relevance of each of the multiple original training data, multiple original training data whose content relevance meets the preset relevance requirements can also be selected from the multiple original training data as the first preliminary training data. Wherein, the preset relevance requirement can be that the content relevance is greater than or equal to a preset relevance threshold. Here, the preset relevance threshold can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure are not limited to this.

[0073] In the above example, after obtaining the content relevance of each of the multiple original training data, multiple first preliminary selected training data with the greatest content relevance can be selected from the multiple original training data, or multiple first preliminary selected training data with content relevance that meets the preset relevance requirements can be selected from the multiple original training data. That is, in the above example, a specific relevance selection strategy can be selected according to actual business needs or application scenarios to ensure the applicability and effectiveness of the multiple candidate training data finally obtained.

[0074] In yet another example, “selecting a plurality of candidate training data from a plurality of first preliminary selected training data according to a diversity selection strategy” may include:

[0075] According to the first diversity selection strategy and / or the second diversity selection strategy, a plurality of candidate training data are selected from a plurality of first preliminary selected training data.

[0076] Among them, the first diversity selection strategy is a diversity selection strategy set for the search term performance evaluation result; the second diversity selection strategy is a diversity selection strategy set for the search term field. Here, the search term performance evaluation result can be the search term performance score corresponding to the first preliminary training data; the search term field can be used to represent the business field related to the first preliminary training data, such as the home decoration field, the home appliance field, the computer and office supplies field, the mobile device field, the women's clothing field, the men's clothing field, the beauty and skin care field, etc.

[0077] In the above examples, multiple candidate training data can be selected from multiple first preliminary training data according to the first diversity selection strategy and / or the second diversity selection strategy. That is to say, in the above examples, a specific diversity selection strategy can be selected according to actual business needs or application scenarios to ensure the applicability and effectiveness of the multiple candidate training data finally obtained.

[0078] In a specific example, “selecting a plurality of candidate training data from a plurality of first preliminary selected training data according to the first diversity selection strategy and / or the second diversity selection strategy” may include:

[0079] According to the first diversity selection strategy, a plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data; wherein each of the plurality of second preliminary selected training data comprises a second preliminary selected visual content description text and a second preliminary selected search term label;

[0080] According to the second diversity selection strategy, a plurality of candidate training data are selected from the plurality of second preliminary selected training data.

[0081] Among them, the second preliminary visual content description text has the same data structure as the target visual content description text, which is not elaborated in the embodiment of the present disclosure; correspondingly, the second preliminary search term label has the same data structure as the target search term label, which is not elaborated in the embodiment of the present disclosure.

[0082] In the above example, a plurality of second preliminary training data may be first selected from a plurality of first preliminary training data according to a first diversity selection strategy, and then a plurality of candidate training data may be selected from a plurality of second preliminary training data according to a second diversity selection strategy. In other words, the first diversity selection strategy and the second diversity selection strategy may be implemented serially. In this way, it is possible to ensure that the selected plurality of candidate training data have data diversity in a specified dimension of search term performance evaluation results, and it is possible to ensure that the selected plurality of candidate training data have data diversity in a specified dimension of search term fields. Therefore, after determining target training data based on a plurality of candidate training data and using them to train a search term generation model to obtain a trained search term generation model, the generalization of the trained search term generation model can be further ensured.

[0083] Further, in a more specific example, “selecting a plurality of second preliminary selected training data from a plurality of first preliminary selected training data according to a first diversity selection strategy” may include:

[0084] Obtaining a search term performance evaluation result of each of the plurality of first preliminary selected training data;

[0085] A plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data by using a probability sampling algorithm based on a search term performance evaluation result of each of the plurality of first preliminary selected training data.

[0086] As mentioned above, the search term performance evaluation result may be a search term performance score corresponding to the first preliminary training data. In actual implementation, the search term performance score corresponding to the first preliminary training data may be obtained in the following manner:

[0087] Obtaining the total number of times the user inputs the first preliminary search term label in the first preliminary training data through the search engine as the total number of searches;

[0088] Obtaining a total number of visual content displays and a total number of visual content clicks related to a first preliminary selected visual content description text in the first preliminary selected training data after a user inputs a first preliminary selected search term tag in the first preliminary selected training data through a search engine;

[0089] Based on the total number of searches, the total number of visual content displays, and the total number of visual content clicks, a search term performance score corresponding to the first preliminary training data is obtained.

[0090] For example, the ratio of the total number of visual content displays and the total number of searches can be calculated as the first scoring parameter, and the ratio of the total number of visual content clicks and the total number of searches can be calculated as the second scoring parameter. The first scoring parameter and the second scoring parameter are fused to obtain a search term performance score corresponding to the first preliminary training data.

[0091] Among them, "fusing the first scoring parameter and the second scoring parameter to obtain the search term performance score corresponding to the first preliminary training data" may include: calculating the product of the first scoring parameter and the first weight value as the first parameter to be fused, and calculating the product of the second scoring parameter and the second weight value as the second parameter to be fused, and calculating the sum of the first parameter to be fused and the second parameter to be fused as the search term performance score corresponding to the first preliminary training data. Among them, the first weight value and the second weight value can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure are not limited to this.

[0092] Exemplarily, there is first preliminary selected training data A, which includes first preliminary selected search term labels A1 and first preliminary selected visual content description text A2. After user B inputs the first preliminary search term label A1 through the search engine, the advertising delivery system does not display the visual content sample corresponding to the first preliminary visual content description text A2 to user B, then the number of searches recorded is 1, the number of visual content displays is 0, and the number of visual content clicks is 0; after user C inputs the first preliminary search term label A1 through the search engine, the advertising delivery system displays the visual content sample corresponding to the first preliminary visual content description text A2 to user C, but user C does not click on the visual content sample corresponding to the first preliminary visual content description text A2, then the number of searches recorded is 1, the number of visual content displays is 1, and the number of visual content clicks is 0; after user D inputs the first preliminary search term label A1 through the search engine, the advertising delivery system displays the visual content sample corresponding to the first preliminary visual content description text A2 to user D, and user D clicks on the visual content sample corresponding to the first preliminary visual content description text A2, then the number of searches recorded is 1, the number of visual content displays is 1, and the number of visual content clicks is 1.

[0093] After that, the ratio of the total number of visual content presentations to the total number of searches can be calculated as the first scoring parameter, that is, 2 / 3 can be used as the first scoring parameter, and the ratio of the total number of visual content clicks to the total number of searches can be calculated as the second scoring parameter, that is, 1 / 3 can be used as the second scoring parameter, and the first scoring parameter and the second scoring parameter can be merged to obtain the search term performance score corresponding to the first preliminary training data. Assuming that the first weight value is 1 / 3 and the second weight value is 2 / 3, the search term performance score corresponding to the first preliminary training data can be obtained as: 2 / 3*1 / 3+1 / 3*2 / 3=4 / 9=0.44.

[0094] After obtaining the search term performance evaluation result of each of the plurality of first preliminary training data, a probability sampling algorithm can be used to select a plurality of second preliminary training data from the plurality of first preliminary training data based on the search term performance evaluation result of each of the plurality of first preliminary training data. The probability sampling algorithm can be a probability sampling algorithm implemented by a probability sampling function obtained from the Numpy library of Python. Here, the probability sampling function can be Np.random.choice().

[0095] In the above example, the search term performance evaluation results of each of the multiple first preliminary training data can be obtained, and a probability sampling algorithm can be used to select multiple second preliminary training data from the multiple first preliminary training data based on the search term performance evaluation results of each of the multiple first preliminary training data. In this way, the selected multiple second preliminary training data will include most of the first preliminary training data with relatively good search term performance evaluation results, and a small part of the first preliminary training data with relatively poor search term performance evaluation results. In this way, while ensuring the quality of the multiple candidate training data finally obtained, it is possible to ensure that the multiple candidate training data selected have data diversity in the specified dimension of the search term performance evaluation results, so that after the target training data is determined based on the multiple candidate training data and used to train the search term generation model to obtain the trained search term generation model, the performance and generalization of the trained search term generation model can be further ensured.

[0096] In another more specific example, “selecting a plurality of candidate training data from a plurality of second preliminary selected training data according to a second diversity selection strategy” may include:

[0097] determining a search term field for each of the plurality of second preliminary selected training data;

[0098] Grouping the plurality of second preliminary selected training data according to the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data to obtain a plurality of preliminary selected training data groups;

[0099] Selecting a plurality of high-quality training data from a plurality of preliminary training data groups respectively;

[0100] Based on the multiple high-quality training data, multiple candidate training data are obtained.

[0101] In actual implementation, for each of the multiple second preliminary training data, the second preliminary training data (for example, the second preliminary search term label in the second preliminary training data) can be processed by word segmentation and named entity recognition to obtain subject information related to the second preliminary training data (for example, product name), and based on the subject information related to the second preliminary training data, the search term field of the second preliminary training data is obtained.

[0102] In addition, in the above example, the high-quality training data can be the second preliminary selected training data with relatively good search term performance evaluation results. Based on this, here, a first implementation method can be provided for "selecting multiple high-quality training data from multiple preliminary selected training data groups" in the above example:

[0103] Taking the plurality of preliminary selected training data groups as target data groups, obtaining the search term performance evaluation results of each second preliminary selected training data in the target data groups;

[0104] Select multiple high-quality training data with the best search term performance evaluation results from the target data group; or select multiple high-quality training data with the search term performance evaluation results that meet preset evaluation requirements from the target data group.

[0105] Exemplarily, after determining the search term field of each second preliminary training data in a plurality of second preliminary training data, and grouping the plurality of second preliminary training data according to the search term field of each second preliminary training data in a plurality of second preliminary training data, and obtaining a plurality of preliminary training data groups, M second preliminary training data with the best search term performance evaluation results can be selected from the target data group as high-quality training data. Wherein, M≥2 and is a positive integer. Here, the specific value of M can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure are not limited to this. After determining the search term field of each second preliminary training data in a plurality of second preliminary training data, and grouping the plurality of second preliminary training data according to the search term field of each second preliminary training data in a plurality of second preliminary training data, and obtaining a plurality of preliminary training data groups, the second preliminary training data whose search term performance evaluation results meet the preset evaluation requirements can also be selected from the target data group as high-quality training data. Wherein, the preset evaluation requirement can be that the search term performance evaluation result is greater than or equal to the preset performance scoring threshold. Here, the preset performance scoring threshold can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure are not limited to this.

[0106] The first implementation method of “selecting a plurality of high-quality training data from a plurality of preliminary training data groups” will be further described below through specific examples.

[0107] Assume that there are second preliminary selected training data E1, second preliminary selected training data E2, second preliminary selected training data E3, second preliminary selected training data E4, second preliminary selected training data E5, second preliminary selected training data E6, second preliminary selected training data E7, second preliminary selected training data E8, second preliminary selected training data E9 and second preliminary selected training data E10.

[0108] Assume again that the search term field of the second preliminary training data E1 is the field of home appliances, and its search term performance evaluation result is 0.90, the search term field of the second preliminary training data E2 is the field of computers and office supplies, and its search term performance evaluation result is 0.56, the search term field of the second preliminary training data E3 is the field of computers and office supplies and the field of mobile devices (for example, the subject information related to the second preliminary training data E3 is a laptop computer, which can be attributed to both the field of computers and office supplies and the field of mobile devices), and its search term performance evaluation result is 0.66, the search term field of the second preliminary training data E4 is the field of home appliances, and its search term performance evaluation result is 0.87, and the search term field of the second preliminary training data E 5, the search term field is the mobile device field, and its search term performance evaluation result is 0.91, the search term field of the second preliminary training data E6 is the home appliance field, and its search term performance evaluation result is 0.48, the search term field of the second preliminary training data E7 is the computer office supplies field, and its search term performance evaluation result is 0.76, the search term field of the second preliminary training data E8 is the computer office supplies field, and its search term performance evaluation result is 0.85, the search term field of the second preliminary training data E9 is the mobile device field, and its search term performance evaluation result is 0.69, and the search term field of the second preliminary training data E10 is the home appliance field, and its search term performance evaluation result is 0.80.

[0109] Then, by grouping the above 10 second primary training data, we can get 3 primary training data groups:

[0110] The first preliminary training data group (home appliance field): the second preliminary training data E1, the second preliminary training data E4, the second preliminary training data E6 and the second preliminary training data E10, and the ranking of the search term performance evaluation results is: the second preliminary training data E1> the second preliminary training data E4>

[0111] The second preliminary selected training data E10> the second preliminary selected training data E6;

[0112] The second preliminary selected training data group (computer office supplies field): second preliminary selected training data E2, second preliminary selected training data E3, second preliminary selected training data E7, second preliminary selected training data E8, and the ranking of the search term performance evaluation results is: second preliminary selected training data E8> second preliminary selected training data E7> second preliminary selected training data E3> second preliminary selected training data E2;

[0113] The third preliminary training data group (mobile device field): second preliminary training data E3, second preliminary training data E5 and second preliminary training data E9, and the search term performance evaluation results are ranked as follows: second preliminary training data E5>second preliminary training data E9>second preliminary training data E3.

[0114] Please combine Figure 2 After determining the search term field of each second preliminary training data in the multiple second preliminary training data, and grouping the multiple second preliminary training data according to the search term field of each second preliminary training data in the multiple second preliminary training data to obtain multiple preliminary training data groups, if the data selection strategy of "taking the multiple preliminary training data groups as the target data groups, obtaining the search term performance evaluation results of each second preliminary training data in the target data groups, and selecting multiple high-quality training data with the best search term performance evaluation results from the target data groups" is adopted, then, when the multiple high-quality training data with the best search term performance evaluation results selected from the target data group are specifically 2, finally, the second preliminary training data E1, the second preliminary training data E4, the second preliminary training data E5, the second preliminary training data E6, the second preliminary training data E8 and the second preliminary training data E9 can be selected as high-quality training data and used to form multiple candidate training data.

[0115] Please combine Figure 3 After determining the search term field of each second preliminary training data in the multiple second preliminary training data, and grouping the multiple second preliminary training data according to the search term field of each second preliminary training data in the multiple second preliminary training data to obtain multiple preliminary training data groups, if the data selection strategy of "taking the multiple preliminary training data groups as the target data groups, obtaining the search term performance evaluation results of each second preliminary training data in the target data groups, and selecting multiple high-quality training data whose search term performance evaluation results meet the preset evaluation requirements from the target data groups" is adopted, then, when the preset evaluation requirement is that the search term performance evaluation result is greater than or equal to the preset performance score threshold, and the preset performance score threshold is 0.75, finally, the second preliminary training data E1, the second preliminary training data E4, the second preliminary training data E5, the second preliminary training data E7, the second preliminary training data E8 and the second preliminary training data E10 can be selected as high-quality training data, and used to form multiple candidate training data.

[0116] Here, a second implementation method can also be provided for "selecting multiple high-quality training data from multiple preliminary training data groups" in the above example:

[0117] Obtaining the inter-group selection order of multiple preliminary training data groups;

[0118] According to the order of group selection, multiple preliminary training data groups are traversed;

[0119] When traversing to one of the multiple preliminary training data groups, a target number of high-quality training data with the best search term performance evaluation results are selected from the preliminary training data group until the selected high-quality training data meets the preset number requirement.

[0120] The target quantity can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure do not limit this; the preset quantity requirement can be that the selected high-quality training data is greater than or equal to the preset quantity threshold. Here, the preset quantity threshold can be set according to actual business needs or application scenarios, and the embodiments of the present disclosure do not limit this.

[0121] In addition, in actual implementation, the inter-group selection order of multiple preliminary training data groups can be obtained by the following method:

[0122] Taking multiple preliminary training data groups as the data groups to be processed, obtaining the mean values ​​of the search term performance evaluations of the data groups to be processed, so as to obtain multiple search term performance evaluation means corresponding to the multiple preliminary training data groups one by one;

[0123] Based on a plurality of search term performance evaluation means corresponding one-to-one to a plurality of preliminary training data groups, an inter-group selection order of the plurality of preliminary training data groups is obtained.

[0124] Among them, the search term performance evaluation mean of the data group to be processed can be the average of the search term performance evaluation results of multiple second preliminary training data in the data group to be processed; "based on multiple search term performance evaluation means corresponding one by one to multiple preliminary training data groups, obtaining the inter-group selection order of multiple preliminary training data groups" can be: obtaining the inter-group selection order of multiple preliminary training data groups in a descending order of multiple search term performance evaluation means.

[0125] The second implementation method of “selecting a plurality of high-quality training data from a plurality of preliminary training data groups” will be further described below through specific examples.

[0126] First, suppose that Figure 2 and Figure 3 Multiple preliminary training data sets are shown.

[0127] Please combine Figure 4, after determining the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data, and grouping the plurality of second preliminary selected training data according to the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data, to obtain a plurality of preliminary selected training data groups, if the method of "obtaining the inter-group selection order of the plurality of preliminary selected training data groups, and traversing the plurality of preliminary selected training data groups according to the inter-group selection order, so as to select a target number of high-quality training data with the best search term performance evaluation result from the preliminary selected training data group each time one of the plurality of preliminary selected training data groups is traversed to is adopted, According to the data selection strategy of "until the selected high-quality training data meets the preset quantity requirement", then, when the target quantity is 1, the preset quantity requirement is that the selected high-quality training data is greater than or equal to the preset quantity threshold, and the preset quantity threshold is 8, finally, the second preliminary selected training data E1, the second preliminary selected training data E3, the second preliminary selected training data E4, the second preliminary selected training data E5, the second preliminary selected training data E7, the second preliminary selected training data E8, the second preliminary selected training data E9 and the second preliminary selected training data E10 can be selected as high-quality training data and used to form multiple candidate training data.

[0128] In the above example, the multiple second preliminary training data can be grouped according to the search term field of each of the multiple second preliminary training data, and multiple candidate training data can be selected from the multiple second preliminary training data by group selection. In this way, it can be further ensured that the selected multiple candidate training data have data diversity in the specified dimension of the search term field, so that after the target training data is determined based on the multiple candidate training data and used to train the search term generation model to obtain the trained search term generation model, the performance and generalization of the trained search term generation model can be further ensured.

[0129] In some optional implementations, step S102, i.e., "using the search word generation model to obtain the search word inference result based on the target visual content description text" may include:

[0130] Obtain entity organization logo samples related to the target visual content description text;

[0131] Get a sample of the expected number of search terms;

[0132] Splicing the entity organization logo sample, the expected search word quantity sample and the target visual content description text to obtain a data splicing result sample;

[0133] The search term generation model is used to obtain the search term inference results based on the data splicing result samples.

[0134] The entity organization identification sample may be a company name; the specific value of the expected search term quantity sample may be set according to actual business needs or application scenarios, and the embodiments of the present disclosure do not limit this. Here, the expected search term quantity sample is used to characterize the number of search terms included in the expected search term inference result.

[0135] In one example, the entity organization identification sample, the expected search word quantity sample and the target visual content description text are spliced, and the obtained data splicing result sample can be: entity organization identification sample-target visual content description text-expected search word quantity sample.

[0136] Through the above method, in the embodiment of the present disclosure, it is possible to obtain entity organization identification samples related to the target visual content description text, as well as to obtain samples of the expected number of search terms, and to splice the entity organization identification samples, the expected number of search terms samples and the target visual content description text to obtain data splicing result samples, and then use the search term generation model to obtain search term inference results based on the data splicing result samples. In this way, richer and more valuable learnable data can be provided for the search term generation model, which can further ensure the performance of the trained search term generation model.

[0137] The disclosed embodiment provides a visual search term generation method, which can be applied to an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing unit or other similar computing device. Figure 5 The flowchart diagram shown in the figure illustrates a search term generation method provided by an embodiment of the present disclosure. It should be noted that although the logical order is shown in the flowchart diagram, in some cases, the steps shown or described in the flowchart can also be performed in other orders.

[0138] Step S501, obtaining visual content description text.

[0139] The visual content description text may be a text content used to describe the visual content sample to be processed. Here, the visual content sample to be processed may be an online advertisement, a video resource, a graphic resource, etc. The online advertisement may be a static advertisement, a dynamic image advertisement, a video advertisement, etc.; the graphic resource may be a static graphic, a dynamic graphic, etc., which is not limited in the embodiments of the present disclosure.

[0140] Step S502, using the trained search word generation model, based on the visual content description text, to obtain a search word generation result.

[0141] Among them, the trained search term generation model can be a search term generation model trained by the aforementioned search term generation model training method.

[0142] By using the visual search word generation method provided in the embodiment of the present disclosure, it is possible to obtain a visual content description text, and use a trained search word generation model to obtain a search word generation result based on the visual content description text. Compared with the solution of manually selecting keywords in the prior art, this can not only save a lot of human resources, but also avoid human bias, thereby ensuring the matching degree of the search word generation result with the specific visual content (for example, online advertisement).

[0143] In some optional implementations, “using a trained search term generation model to obtain a search term generation result based on the visual content description text” may include:

[0144] Obtaining entity organization identifiers associated with the visual content description text;

[0145] Get the expected number of search terms;

[0146] The entity organization logo, the expected number of search terms and the visual content description text are spliced ​​to obtain a data splicing result;

[0147] The trained search term generation model is used to obtain the search term generation result based on the data concatenation result.

[0148] The entity organization identifier may be a company name; the specific value of the expected number of search terms may be set according to actual business needs or application scenarios, and the embodiments of the present disclosure do not limit this. Here, the expected number of search terms is used to represent the number of search terms included in the expected search term generation results.

[0149] In one example, the entity organization identifier, the expected number of search terms and the visual content description text are spliced, and the obtained data splicing result may be: entity organization identifier-visual content description text-expected number of search terms.

[0150] Through the above method, in the embodiment of the present disclosure, the entity organization identifier related to the visual content description text and the expected number of search terms can be obtained, and the entity organization identifier, the expected number of search terms and the visual content description text are spliced ​​to obtain a data splicing result, and then the trained search term generation model is used to obtain the search term generation result based on the data splicing result. In this way, more abundant and high-value reference data can be provided for the trained search term generation model, which can further ensure the matching degree of the search term generation result with the specific visual content (for example, online advertisement).

[0151] In some optional implementations, “using a trained search term generation model to obtain a search term generation result based on the visual content description text” may include:

[0152] By using the trained search term generation model, the search term generation result is obtained based on the visual content description text according to the kernel sampling strategy and / or the repeated penalty strategy.

[0153] Among them, the core sampling strategy is the Top-p sampling strategy, which is used to limit the words considered in each step of generation to make the generated search term generation results more diverse and reasonable; the repetition penalty strategy is used to reduce the probability of selecting used words to reduce the repetitive content in the search term generation results, so as to further ensure the diversity of the search term generation results.

[0154] In addition, it should be understood that in the embodiments of the present disclosure, when "using a trained search term generation model to obtain a search term generation result based on the visual content description text" includes "obtaining an entity organization identifier related to the visual content description text, and obtaining an expected number of search terms, and splicing the entity organization identifier, the expected number of search terms, and the visual content description text to obtain a data splicing result, and then using a trained search term generation model to obtain a search term generation result based on the data splicing result", the trained search term generation model can be used to obtain a search term generation result based on the data splicing result in accordance with a core sampling strategy and / or a repetition penalty strategy.

[0155] Through the above method, in the embodiment of the present disclosure, a trained search term generation model can be used to obtain search term generation results based on visual content description text according to a core sampling strategy and / or a repeated penalty strategy to ensure the diversity and rationality of the search term generation results.

[0156] The following, combined with Figure 6 , a training method for a search term generation model and an application process of the trained search term generation model provided in an embodiment of the present disclosure are described.

[0157] First, historical visual content samples and original search term labels with corresponding relationships are obtained from the advertising delivery system. The historical visual content samples may include visual content landing pages (for example, when the visual content samples are online advertisements, the visual content landing pages may be advertisement landing pages) and entity organization identification samples. Here, the entity organization identification samples may be company names.

[0158] Using the large language model, based on the inductive prompt words, the visual content landing page is summarized and summarized to obtain the original visual content description text, and then the original visual content description text and the original search word label are used together as the original training data. Please refer to Table 1 for a visual content landing page (Table 1 only shows its text content, and the actual visual content landing page can also include image content) and the original visual content description text obtained by summarizing it.

[0159] Table 1

[0160]

[0161]

[0162] Repeat the above steps to obtain multiple original training data.

[0163] Thereafter, multiple first preliminary selected training data may be selected from the multiple original training data according to the relevance selection strategy. Specifically, the content relevance of each of the multiple original training data may be obtained, and multiple first preliminary selected training data with the greatest content relevance may be selected from the multiple original training data, or multiple first preliminary selected training data whose content relevance meets the preset relevance requirement may be selected from the multiple original training data. Each of the multiple first preliminary selected training data includes a first preliminary selected visual content description text and a first preliminary selected search term label.

[0164] Next, multiple second preliminary training data can be selected from multiple first preliminary training data according to the first diversity selection strategy, and multiple candidate training data can be selected from multiple second preliminary training data according to the second diversity selection strategy, and each candidate training data in the multiple candidate training data is used as the target training data. Among them, the first diversity selection strategy is a diversity selection strategy set for the search term performance evaluation results; each second preliminary training data in the multiple second preliminary training data includes a second preliminary visual content description text and a second preliminary search term label; the second diversity selection strategy is a diversity selection strategy set for the search term field; each candidate training data in the multiple candidate training data includes a candidate visual content description text and a candidate search term label; the target training data includes a target visual content description text and a target search term label.

[0165] In one example, the target training data may be as shown in Table 2.

[0166] Table 2

[0167]

[0168]

[0169] Finally, we can obtain samples of entity organization identification related to the target visual content description text, as well as samples of the expected number of search terms, and splice the entity organization identification samples, the expected number of search terms, and the target visual content description text to obtain data splicing result samples. We then use the search term generation model to obtain search term inference results based on the data splicing result samples, and train the search term generation model based on the target search term label and the search term inference results to obtain a trained search term generation model.

[0170] In one example, the trained search term generation model is tested to obtain the search term inference results shown in Table 3.

[0171] Table 3

[0172]

[0173]

[0174]

[0175] In the application process of the trained search term generation model, the visual content description text can be obtained, and the trained search term generation model is used to obtain the search term generation result based on the visual content description text.

[0176] For the specific functions and examples of the above steps, please refer to the relevant descriptions of the corresponding steps in the aforementioned search term generation model training method embodiment and the search term generation method embodiment, which will not be repeated here.

[0177] See also Figure 7 , which is a schematic diagram of an application scenario of a training method for a search term generation model provided in an embodiment of the present disclosure.

[0178] The training method of the search term generation model provided in the embodiments of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. The terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing device or other similar computing device.

[0179] Here, electronic devices are used to:

[0180] Obtain target training data; wherein the target training data includes target visual content description text and target search term labels;

[0181] Using the search term generation model, the text is described based on the target visual content to obtain the search term inference result;

[0182] Based on the target search term label and the search term inference result, the search term generation model is trained to obtain a trained search term generation model.

[0183] It should be noted that in the embodiments of the present disclosure, Figure 7 The application scenario diagram shown is only illustrative and not restrictive. Those skilled in the art can Figure 7 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0184] See also Figure 8 , which is a schematic diagram of an application scenario of a search term generation method provided in an embodiment of the present disclosure.

[0185] The search term generation method provided by the embodiment of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. The terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing device or other similar computing devices.

[0186] Here, electronic devices are used to:

[0187] Get the visual content description text;

[0188] Using the trained search term generation model, the text is described based on the visual content to obtain the search term generation results.

[0189] It should be noted that in the embodiments of the present disclosure, Figure 8 The application scenario diagram shown is only illustrative and not restrictive. Those skilled in the art can Figure 8 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0190] In order to better implement the aforementioned training method for the search term generation model, the embodiment of the present disclosure also provides a training device 900 for the search term generation model, which can be integrated into an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device. Fig. 9 The schematic structural block diagram shown illustrates a training device 900 for a search term generation model provided in the disclosed embodiment.

[0191] The training device 900 of the search term generation model comprises:

[0192] The training data acquisition unit 901 is used to acquire target training data; wherein the target training data includes target visual content description text and target search word label;

[0193] The inference result acquisition unit 902 is used to obtain the search word inference result based on the target visual content description text by using the search word generation model;

[0194] The model training unit 903 is used to train the search term generation model based on the target search term label and the search term inference result to obtain a trained search term generation model.

[0195] In some optional implementations, the training data acquisition unit 901 is used to:

[0196] Acquire multiple original training data; wherein each of the multiple original training data includes original visual content description text and original search term label;

[0197] According to a relevance selection strategy and / or a diversity selection strategy, multiple candidate training data are selected from multiple original training data; wherein each candidate training data in the multiple candidate training data includes a candidate visual content description text and a candidate search term label;

[0198] Based on the plurality of candidate training data, target training data is determined.

[0199] In some optional implementations, the training data acquisition unit 901 is used to:

[0200] According to the relevance selection strategy, a plurality of first preliminary selected training data are selected from the plurality of original training data; wherein each of the plurality of first preliminary selected training data comprises a first preliminary selected visual content description text and a first preliminary selected search term label;

[0201] According to the diversity selection strategy, a plurality of candidate training data are selected from a plurality of first preliminary selected training data.

[0202] In some optional implementations, the training data acquisition unit 901 is used to:

[0203] Obtaining content relevance of each original training data among the plurality of original training data;

[0204] A plurality of first preliminary selected training data with the greatest content relevance are selected from the plurality of original training data; or, a plurality of first preliminary selected training data with content relevance meeting a preset relevance requirement are selected from the plurality of original training data.

[0205] In some optional implementations, the training data acquisition unit 901 is used to:

[0206] According to the first diversity selection strategy and / or the second diversity selection strategy, multiple candidate training data are selected from multiple first preliminary training data; wherein the first diversity selection strategy is a diversity selection strategy set for the search term performance evaluation results; and the second diversity selection strategy is a diversity selection strategy set for the search term field.

[0207] In some optional implementations, the training data acquisition unit 901 is used to:

[0208] According to the first diversity selection strategy, a plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data; wherein each of the plurality of second preliminary selected training data comprises a second preliminary selected visual content description text and a second preliminary selected search term label;

[0209] According to the second diversity selection strategy, a plurality of candidate training data are selected from the plurality of second preliminary selected training data.

[0210] In some optional implementations, the training data acquisition unit 901 is used to:

[0211] Obtaining a search term performance evaluation result of each of the plurality of first preliminary selected training data;

[0212] A plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data by using a probability sampling algorithm based on a search term performance evaluation result of each of the plurality of first preliminary selected training data.

[0213] In some optional implementations, the training data acquisition unit 901 is used to:

[0214] determining a search term field for each of the plurality of second preliminary selected training data;

[0215] Grouping the plurality of second preliminary selected training data according to the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data to obtain a plurality of preliminary selected training data groups;

[0216] Selecting a plurality of high-quality training data from a plurality of preliminary training data groups respectively;

[0217] Based on the multiple high-quality training data, multiple candidate training data are obtained.

[0218] In some optional implementations, the training data acquisition unit 901 is used to:

[0219] Taking the plurality of preliminary selected training data groups as target data groups, obtaining the search term performance evaluation results of each second preliminary selected training data in the target data groups;

[0220] Select multiple high-quality training data with the best search term performance evaluation results from the target data group; or select multiple high-quality training data with the search term performance evaluation results that meet preset evaluation requirements from the target data group.

[0221] In some optional implementations, the training data acquisition unit 901 is used to:

[0222] Obtaining the inter-group selection order of multiple preliminary training data groups;

[0223] According to the order of group selection, multiple preliminary training data groups are traversed;

[0224] When traversing to one of the multiple preliminary training data groups, a target number of high-quality training data with the best search term performance evaluation results are selected from the preliminary training data group until the selected high-quality training data meets the preset number requirement.

[0225] In some optional implementations, the inference result acquisition unit 902 is used to:

[0226] Obtain entity organization logo samples related to the target visual content description text;

[0227] Get a sample of the expected number of search terms;

[0228] Splicing the entity organization logo sample, the expected search word quantity sample and the target visual content description text to obtain a data splicing result sample;

[0229] The search term generation model is used to obtain the search term inference result based on the data concatenation result.

[0230] In the embodiment of the present disclosure, the specific functions and examples of each unit in the training device 900 for the search term generation model can be found in the relevant descriptions of the corresponding steps in the aforementioned training method embodiment for the search term generation model, and will not be repeated here.

[0231] In order to better implement the aforementioned search term generation method, the embodiment of the present disclosure further provides a search term generation device 1000, which can be integrated into an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device 1000. Fig.10 The schematic structural block diagram shown illustrates a training device 1000 for a search term generation model provided in the disclosed embodiment.

[0232] The training device 1000 of the search term generation model comprises:

[0233] A description text acquisition unit 1001 is used to acquire a visual content description text;

[0234] The search word generation unit 1002 is used to obtain a search word generation result based on the visual content description text by using the trained search word generation model.

[0235] In some optional implementations, the search term generating unit 1002 is used to:

[0236] Obtaining entity organization identifiers associated with the visual content description text;

[0237] Get the expected number of search terms;

[0238] The entity organization logo, the expected number of search terms and the visual content description text are spliced ​​to obtain a data splicing result;

[0239] Using the trained search term generation model, the search term generation results are obtained based on the data splicing results.

[0240] In some optional implementations, the search term generating unit 1002 is used to:

[0241] By using the trained search term generation model, the search term generation result is obtained based on the visual content description text according to the kernel sampling strategy and / or the repeated penalty strategy.

[0242] In the embodiment of the present disclosure, the specific functions and examples of each unit in the training device 1000 of the search term generation model can be found in the relevant descriptions of the corresponding steps in the embodiment of the search term generation method, and will not be repeated here.

[0243] In the technical solution disclosed herein, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0244] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0245] Fig.11 A schematic block diagram of an example electronic device 1100 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as vehicle-mounted computing devices, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0246] like Fig.11 As shown, the electronic device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded from a storage unit 1108 to a random access memory (RAM) 1103. In RAM 1103, various programs and data required for the operation of the electronic device 1100 can also be stored. The computing unit 1101, ROM 1102, and RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0247] A number of components in the electronic device 1100 are connected to the I / O interface 1105, including: an input unit 1106, such as a keyboard, a mouse, etc.; an output unit 1107, such as various types of renderers, speakers, etc.; a storage unit 1108, such as a disk, an optical disk, etc.; and a communication unit 1109, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1109 allows the electronic device 1100 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0248] The computing unit 1101 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSP), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1101 performs the various methods and processes described above, such as the training method and / or search word generation method of the search word generation model. For example, in some embodiments, the training method and / or search word generation method of the search word generation model may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1108. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1100 via the ROM 1102 and / or the communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps in the training method of the search term generation model and / or the search term generation method described above may be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured as the training method of the search term generation model and / or the search term generation method in any other appropriate manner (e.g., by means of firmware).

[0249] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0250] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data optimization device so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine or completely on a remote machine or server.

[0251] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM) or a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0252] To provide interaction with a user, the systems and techniques described herein may be implemented on a computer having: a rendering device (e.g., a cathode ray tube (CRT) renderer or a liquid crystal display (LCD)) for rendering information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices are also used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0253] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: Local Area Network (LAN), Wide Area Network (WAN), and the Internet.

[0254] The computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0255] The embodiments of the present disclosure also provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute a training method for a search term generation model and / or a search term generation method.

[0256] The embodiments of the present disclosure also provide a computer program product, including a computer program, which, when executed by a processor, implements a training method for a search term generation model and / or a search term generation method.

[0257] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and this document does not limit this. In addition, in the present disclosure, relational terms such as "first", "second", "third", etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. In addition, "multiple" in the present disclosure can be understood as at least two.

[0258] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a search term generation model, comprising: Acquire target training data; wherein the target training data includes target visual content description text and target search term labels; Using a search term generation model, based on the target visual content description text, a search term inference result is obtained; Based on the target search term label and the search term inference result, the search term generation model is trained to obtain a trained search term generation model.

2. The method according to claim 1, wherein: The obtaining of target training data comprises: Acquire multiple original training data; wherein each of the multiple original training data includes original visual content description text and original search term label; According to a relevance selection strategy and / or a diversity selection strategy, a plurality of candidate training data are selected from the plurality of original training data; wherein each candidate training data in the plurality of candidate training data comprises a candidate visual content description text and a candidate search term label; Based on the plurality of candidate training data, target training data is determined.

3. The method according to claim 2, wherein: The selecting a plurality of candidate training data from the plurality of original training data according to the relevance selection strategy and / or the diversity selection strategy includes: According to the relevance selection strategy, a plurality of first preliminary selected training data are selected from the plurality of original training data; wherein each of the plurality of first preliminary selected training data comprises a first preliminary selected visual content description text and a first preliminary selected search term label; According to the diversity selection strategy, a plurality of candidate training data are selected from the plurality of first preliminary selected training data.

4. The method according to claim 3, wherein: The step of selecting a plurality of first preliminary selected training data from the plurality of original training data according to the relevance selection strategy comprises: Obtaining content relevance of each original training data among the plurality of original training data; A plurality of first preliminary selected training data with the greatest content relevance are selected from the plurality of original training data; or a plurality of first preliminary selected training data with content relevance meeting a preset relevance requirement are selected from the plurality of original training data.

5. The method according to claim 3, wherein: The step of selecting a plurality of candidate training data from the plurality of first preliminary selected training data according to the diversity selection strategy includes: According to the first diversity selection strategy and / or the second diversity selection strategy, multiple candidate training data are selected from the multiple first preliminary training data; wherein the first diversity selection strategy is a diversity selection strategy set for the search term performance evaluation results; and the second diversity selection strategy is a diversity selection strategy set for the search term field.

6. The method according to claim 5, wherein: The selecting a plurality of candidate training data from the plurality of first preliminary selected training data according to the first diversity selection strategy and / or the second diversity selection strategy comprises: According to the first diversity selection strategy, a plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data; wherein each second preliminary selected training data in the plurality of second preliminary selected training data comprises a second preliminary selected visual content description text and a second preliminary selected search term label; According to the second diversity selection strategy, a plurality of candidate training data are selected from the plurality of second preliminary selected training data.

7. The method according to claim 6, wherein: The selecting a plurality of second preliminary selected training data from the plurality of first preliminary selected training data according to the first diversity selection strategy comprises: Obtaining a search term performance evaluation result of each of the plurality of first preliminary selected training data; A plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data by using a probability sampling algorithm based on a search term performance evaluation result of each first preliminary selected training data in the plurality of first preliminary selected training data.

8. The method according to claim 6, wherein: The step of selecting a plurality of candidate training data from the plurality of second preliminary selected training data according to the second diversity selection strategy includes: determining a search term field for each of the plurality of second preliminary selected training data; Grouping the plurality of second preliminary selected training data according to the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data to obtain a plurality of preliminary selected training data groups; Selecting a plurality of high-quality training data from the plurality of preliminary training data groups respectively; Based on the plurality of high-quality training data, a plurality of candidate training data are obtained.

9. The method according to claim 8, wherein: The selecting a plurality of high-quality training data from the plurality of preliminary selected training data groups respectively comprises: Taking the plurality of preliminary selected training data groups as target data groups, obtaining search term performance evaluation results of each second preliminary selected training data in the target data groups; A plurality of high-quality training data with the best search term performance evaluation results are selected from the target data group; or, a plurality of high-quality training data with the search term performance evaluation results meeting preset evaluation requirements are selected from the target data group.

10. The method according to claim 8, wherein: The selecting a plurality of high-quality training data from the plurality of preliminary selected training data groups respectively comprises: Obtaining a selection order among the plurality of preliminary training data groups; According to the inter-group selection order, traverse the multiple preliminary selected training data groups; Each time a preliminary selected training data group among the plurality of preliminary selected training data groups is traversed, a target number of high-quality training data with the best search term performance evaluation results are selected from the preliminary selected training data group until the selected high-quality training data meet a preset number requirement.

11. The method according to any one of claims 1 to 10, wherein: The method of using the search word generation model to obtain the search word inference result based on the target visual content description text includes: Acquire entity organization identification samples related to the target visual content description text; Get a sample of the expected number of search terms; Splicing the entity organization identification sample, the expected search word quantity sample and the target visual content description text to obtain a data splicing result sample; The search term generation model is used to obtain the search term inference result based on the data concatenation result.

12. A method for generating a search term, comprising: Get the visual content description text; The trained search term generation model is used to obtain search term generation results based on the visual content description text.

13. The method according to claim 12, wherein: The method of using the trained search word generation model to obtain a search word generation result based on the visual content description text includes: Obtaining an entity organization identifier related to the visual content description text; Get the expected number of search terms; splicing the entity organization identifier, the expected number of search terms and the visual content description text to obtain a data splicing result; The trained search term generation model is used to obtain the search term generation result based on the data concatenation result.

14. The method according to claim 12 or 13, wherein: The method of using the trained search word generation model to obtain a search word generation result based on the visual content description text includes: The trained search term generation model is used to obtain search term generation results based on the visual content description text according to the kernel sampling strategy and / or the repetition penalty strategy.

15. A training device for a search term generation model, comprising: A training data acquisition unit, used to acquire target training data; wherein the target training data includes target visual content description text and target search term label; An inference result acquisition unit, configured to obtain a search word inference result based on the target visual content description text by using a search word generation model; A model training unit is used to train the search term generation model based on the target search term label and the search term inference result to obtain a trained search term generation model.

16. The device according to claim 15, wherein: The training data acquisition unit is used for: Acquire multiple original training data; wherein each of the multiple original training data includes original visual content description text and original search term label; According to a relevance selection strategy and / or a diversity selection strategy, a plurality of candidate training data are selected from the plurality of original training data; wherein each candidate training data in the plurality of candidate training data comprises a candidate visual content description text and a candidate search term label; Based on the plurality of candidate training data, target training data is determined.

17. The device according to claim 16, wherein: The training data acquisition unit is used for: According to the relevance selection strategy, a plurality of first preliminary selected training data are selected from the plurality of original training data; wherein each of the plurality of first preliminary selected training data comprises a first preliminary selected visual content description text and a first preliminary selected search term label; According to the diversity selection strategy, a plurality of candidate training data are selected from the plurality of first preliminary selected training data.

18. The device according to claim 17, wherein: The training data acquisition unit is used for: Obtaining content relevance of each original training data among the plurality of original training data; Selecting a plurality of first preliminary selected training data with the greatest content relevance from the plurality of original training data; Alternatively, a plurality of first preliminary selected training data whose content relevance meets a preset relevance requirement are selected from the plurality of original training data.

19. The device according to claim 17, wherein: The training data acquisition unit is used for: According to the first diversity selection strategy and / or the second diversity selection strategy, multiple candidate training data are selected from the multiple first preliminary training data; wherein the first diversity selection strategy is a diversity selection strategy set for the search term performance evaluation results; and the second diversity selection strategy is a diversity selection strategy set for the search term field.

20. The device according to claim 19, wherein The training data acquisition unit is used for: According to the first diversity selection strategy, a plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data; wherein each second preliminary selected training data in the plurality of second preliminary selected training data comprises a second preliminary selected visual content description text and a second preliminary selected search term label; According to the second diversity selection strategy, a plurality of candidate training data are selected from the plurality of second preliminary selected training data.

21. The device according to claim 20, wherein: The training data acquisition unit is used for: Obtaining a search term performance evaluation result of each of the plurality of first preliminary selected training data; A plurality of second preliminary selected training data are selected from the plurality of first preliminary selected training data by using a probability sampling algorithm based on a search term performance evaluation result of each first preliminary selected training data in the plurality of first preliminary selected training data.

22. The device according to claim 20, wherein: The training data acquisition unit is used for: determining a search term field for each of the plurality of second preliminary selected training data; Grouping the plurality of second preliminary selected training data according to the search term field of each second preliminary selected training data in the plurality of second preliminary selected training data to obtain a plurality of preliminary selected training data groups; Selecting a plurality of high-quality training data from the plurality of preliminary training data groups respectively; Based on the plurality of high-quality training data, a plurality of candidate training data are obtained.

23. The device according to claim 22, wherein: The training data acquisition unit is used for: Taking the plurality of preliminary selected training data groups as target data groups, obtaining search term performance evaluation results of each second preliminary selected training data in the target data groups; Selecting a plurality of high-quality training data with the best search term performance evaluation results from the target data group; Alternatively, a plurality of high-quality training data whose search term performance evaluation results meet preset evaluation requirements are selected from the target data group.

24. The device according to claim 22, wherein: The training data acquisition unit is used for: Obtaining a selection order among the plurality of preliminary training data groups; According to the inter-group selection order, traverse the multiple preliminary selected training data groups; Each time a preliminary selected training data group among the plurality of preliminary selected training data groups is traversed, a target number of high-quality training data with the best search term performance evaluation results are selected from the preliminary selected training data group until the selected high-quality training data meet a preset number requirement.

25. The device according to any one of claims 15 to 24, wherein: The inference result acquisition unit is used for: Acquire entity organization identification samples related to the target visual content description text; Get a sample of the expected number of search terms; Splicing the entity organization identification sample, the expected search word quantity sample and the target visual content description text to obtain a data splicing result sample; The search term generation model is used to obtain the search term inference result based on the data concatenation result.

26. A search term generating device, comprising: A description text acquisition unit, used to acquire a description text of the visual content; The search word generation unit is used to obtain a search word generation result based on the visual content description text by using a trained search word generation model.

27. The device according to claim 26, wherein: The search word generating unit is used for: Obtaining an entity organization identifier related to the visual content description text; Get the expected number of search terms; splicing the entity organization identifier, the expected number of search terms and the visual content description text to obtain a data splicing result; The trained search term generation model is used to obtain the search term generation result based on the data concatenation result.

28. The device according to claim 26 or 27, wherein The search word generating unit is used for: The trained search term generation model is used to obtain search term generation results based on the visual content description text according to the kernel sampling strategy and / or the repetition penalty strategy.

29. An electronic device comprising: at least one processor; a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 14.

30. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 14.

31. A computer program product, comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 14.

Images