Content label determination method and related device

Abstract

The invention discloses a content tag determination method and a related device, and the method comprises the steps: obtaining M tag pairs corresponding to target content, constructing corresponding input data for each tag pair according to the tag pair and the target content, and predicting the correlation between two candidate tags included in the tag pair and the target content according to the input data, and obtaining a prediction result corresponding to the label pair. It can be seen that the N candidate tags are pairwise constructed, the tag pairs and the target content are constructed into the input data, and in single prediction, prediction is carried out based on the target content and the whole input data constructed by the two different candidate tags; as a result, it is possible to analyze how the degree of correlation between the two candidate tags and the target content is relative to each other from the whole. In this way, the problem that the absolute correlation of multiple labels is not comparable due to one-sided absolute correlation comparison in the prior art is solved, and a more accurate content label is determined for the target content.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a content label determination method and related devices. Background Art

[0002] With the continuous development of Internet technology, content, as a carrier of information, data, knowledge, etc., has become a convenient way for users to obtain information, data, knowledge and other information by recommending content to users.

[0003] Accordingly, how to improve the accuracy of recommendations so that users can obtain content that meets their expectations and content providers can obtain good recommendation benefits has become the key in content recommendation scenarios. In related technologies, content tags are first determined for content, and the content tags can be used to describe the content. Then, based on the content tags, the content is recommended in a targeted manner.

[0004] However, the method of determining labels used in the related art is difficult to determine accurate content labels, which correspondingly leads to low recommendation accuracy. Summary of the invention

[0005] In order to solve the above technical problems, the present application provides a content label determination method and related devices, which are conducive to determining labels that are relatively more relevant to the content, that is, determining more accurate content labels. And, based on the content labels, it is conducive to more accurate content recommendations.

[0006] The embodiments of the present application disclose the following technical solutions:

[0007] In one aspect, an embodiment of the present application provides a method for determining a tag of content, the method comprising:

[0008] Obtain M label pairs corresponding to the target content, wherein the M label pairs are composed of two different candidate labels among N candidate labels corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1;

[0009] For each of the M label pairs, constructing input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively;

[0010] Predicting the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtaining a prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content;

[0011] According to the prediction results respectively corresponding to the M label pairs, a content label corresponding to the target content is determined from the N candidate labels, and the content label is used to recommend the target content.

[0012] In another aspect, an embodiment of the present application provides a device for determining a content label, the device comprising an acquisition unit, a construction unit, a prediction unit, and a determination unit:

[0013] The acquisition unit is used to acquire M label pairs corresponding to the target content, wherein the M label pairs are composed of two different candidate labels among the N candidate labels corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1;

[0014] The construction unit is used to construct, for each label pair of the M label pairs, input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively;

[0015] The prediction unit is used to predict the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtain the prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content;

[0016] The determining unit is used to determine a content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs respectively, and the content label is used to recommend the target content.

[0017] On the other hand, an embodiment of the present application provides a computer device, the computer device comprising a processor and a memory:

[0018] The memory is used to store a computer program and transmit the computer program to the processor;

[0019] The processor is configured to execute the method described in any one of the preceding aspects according to instructions in the computer program.

[0020] On the other hand, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program. When the computer program is executed by a computer device, the computer device executes the method described in any of the above aspects.

[0021] On the other hand, an embodiment of the present application provides a computer program product, including a computer program, which, when executed on a computer device, enables the computer device to execute the method described in any one of the aforementioned aspects.

[0022] It can be seen from the above technical solution that M label pairs corresponding to the target content are obtained, and the M label pairs are composed of two different candidate labels among the N candidate labels corresponding to the target content. For each label pair, the input data corresponding to the label pair is constructed according to the label pair and the target content, and the correlation between the two candidate labels included in the label pair and the target content is predicted based on the input data to obtain the prediction result corresponding to the label pair. It can be seen that the N candidate labels are constructed into pairs, and the label pair and the target content are constructed into input data, and in a single prediction, the input data constructed based on the target content and the two different candidate labels are predicted as a whole, so that the correlation between the two candidate labels and the target content can be analyzed from the overall perspective. Based on this, the correlation between the two candidate labels and the target content can be comprehensively analyzed. The corresponding prediction result can be used to indicate the size relationship between the correlation between the two candidate labels included in the label pair and the target content. Compared with the method in the related art that only predicts based on a single label and content and analyzes the absolute correlation between the label and the content, the correlation between the two labels and the content is analyzed as a whole, so it can solve the problem that the absolute correlation is one-sided and the absolute correlation between multiple labels is not comparable. In this way, when determining the content label for the target content from the N candidate labels based on the prediction results corresponding to the M label pairs, it is helpful to determine the label that is relatively more relevant to the target content, that is, to determine a more accurate content label. And, based on the content label, it is helpful to recommend the target content more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technical members in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0024] Figure 1 A schematic diagram of a content processing link provided in an embodiment of the present application;

[0025] Figure 2 A schematic diagram of an application scenario of a method for determining a content label provided in an embodiment of the present application;

[0026] Figure 3 A flowchart of a method for determining a content label provided in an embodiment of the present application;

[0027] Figure 4 A schematic diagram of a label processing framework for target content provided in an embodiment of the present application;

[0028] Figure 5 A structural diagram of a content label determination device provided in an embodiment of the present application;

[0029] Figure 6 A structural diagram of a terminal provided in an embodiment of the present application;

[0030] Figure 7 A structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The embodiments of the present application are described below in conjunction with the accompanying drawings.

[0032] In a related technology, for a piece of content that needs to be labeled, the correlation between each label and the content is predicted, and the correlation score corresponding to each label is obtained. However, in this method, a single prediction is only based on a single label and content, and the correlation between this label and the content is analyzed, so it is a method of analyzing the absolute correlation between a single label and content. Because of this, the absolute correlations of multiple labels may not be comparable. In practical applications, this method can also be called a prediction method based on the single point method (point-wise).

[0033] For example, for a certain content, the absolute correlation between the label "food" and it is predicted once, and the correlation score is 0.61. Also, the absolute correlation between the label "travel" and it is predicted once, and the correlation score is 0.58. The correlation scores of these two labels are not much different, and it is difficult to say that the correlation of "travel" is definitely lower than that of "food".

[0034] In addition, another related technology also provides a method of inputting all tags and content into the model together, and using the model to sort all tags as a whole. However, when this method is used, since all tags need to be sorted in a single prediction, the requirements for designing model parameters (such as the metric function for building the model, etc.) are higher, which makes optimization difficult. This also makes the results of the model output less controllable after using this method, and it is prone to omissions, duplications and other problems. Especially in practical applications, the number of tags is relatively large, so it is difficult to determine accurate content tags for the content using this method. Generally, this method can also be called a list-wise prediction method.

[0035] To this end, an embodiment of the present application provides a method and a related device for determining a content label, wherein N candidate labels corresponding to the target content are paired up, thereby obtaining M label pairs. For each label pair, the label pair and the target content are constructed into input data, and in a single prediction, the input data constructed based on the target content and two different candidate labels are predicted as a whole, so that the correlation between the two candidate labels and the target content can be analyzed as a whole. Based on this, the degree of correlation between the two candidate labels and the target content can be comprehensively analyzed. The corresponding prediction results can be used to indicate the size relationship between the two candidate labels included in the label pair and the target content.

[0036] Compared with the first related technology mentioned above, because the correlation between the two labels and the content is analyzed as a whole, it is able to solve the problem that the absolute correlation is one-sided and the absolute correlation of multiple labels is not comparable. In addition, compared with the second related technology mentioned above, two labels are predicted at a time instead of all labels, which is conducive to reducing the difficulty of optimization and avoiding omissions, duplications, etc. Correspondingly, when determining the content label for the target content from the N candidate labels based on the prediction results corresponding to the M label pairs, it is conducive to determining the label that is relatively more relevant to the target content, that is, determining a more accurate content label. Afterwards, when recommending content based on the content label, it is also conducive to improving the accuracy of the recommendation.

[0037] It can be seen that the present application constructs N candidate labels into pairs and then performs prediction, so it can also be called a prediction method based on the pair-wise method.

[0038] The content label determination method provided in the embodiment of the present application can be implemented by a computer device, which can be a terminal or a server, wherein the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminals include but are not limited to smartphones, computers, intelligent voice interaction devices, smart home appliances, car terminals, etc. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0039] The embodiments of the present application can be specifically applied to various scenarios of determining content tags for content. For example, in a content recommendation scenario, you can first determine the content tags for the content to be recommended, and then recommend the content based on the determined content tags. For example, a content recommendation scenario can include recommending content published on a content platform, and can also include determining a suitable recommendation channel for the content provided by a content provider based on the content tags, and recommending the content through this recommendation channel. In actual applications, the process of recommending content can also be called the process of distributing content, and content operations are also achieved based on this. For example, a content platform can increase the stickiness of users using the platform based on content operations.

[0040] For ease of understanding, the present application provides the following examples: Figure 1 The schematic diagram of a content processing link shown may include three main links: content production, content processing and content recommendation. Specifically:

[0041] In the content production phase, target content can be determined. For example, if a user publishes a certain content on a content platform, it can be used as target content. In the content processing phase, content tags can be determined for the target content. Finally, in the content recommendation phase, content can be recommended based on the content tags determined in the content processing phase. For example, it can be recommended on the content platform, through target recommendation channels, etc.

[0042] It should be noted that in the specific implementation of the present application, the process of determining the content label may involve relevant data such as user information. When the above embodiments of the present application are applied to specific products or technologies, the user's separate consent or separate permission is required, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0043] Figure 2 The application scenario of the method for determining the content label provided by the embodiment of the present application is shown. Figure 2 In the scenario shown, the server 200 is used as an example of the aforementioned computer device for illustration:

[0044] When it is necessary to determine a content tag for the target content, first, the server 200 may obtain M tag pairs corresponding to the target content, wherein the M tag pairs are composed of two different candidate tags among the N candidate tags corresponding to the target content. Where N is an integer greater than 1, and M is an integer greater than or equal to 1. Figure 1 In the example, N=2 and M=1 are used as an example for explanation. Specifically, the N candidate tags may be, for example, a candidate tag a and a candidate tag b.

[0045] Next, for the tag pair consisting of candidate tag a and candidate tag b, the server 200 can construct the input data corresponding to the tag pair according to the tag pair and the target content. Furthermore, the server 200 can predict the correlation between the two candidate tags, candidate tag a and candidate tag b, included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtain the prediction result corresponding to the tag pair.

[0046] Since in a single prediction, the prediction is made based on the input data constructed based on the target content and two different candidate labels, the correlation between the two candidate labels and the target content can be analyzed as a whole. Based on this, the degree of correlation between the two candidate labels and the target content can be comprehensively analyzed. The corresponding prediction result can be used to indicate the magnitude relationship between the correlation between the two candidate labels included in the label pair and the target content, that is, it can indicate which candidate label a and candidate label b has a greater correlation with the target content.

[0047] Finally, the server 200 may determine the content tag corresponding to the target content from the candidate tags a and the candidate tags b according to the prediction result. For example, the determined content tag is the candidate tag b. Afterwards, the content tag may be used to recommend the target content.

[0048] Compared with the method of predicting absolute relevance in related technologies, this application predicts the relative relevance between two candidate tags and content at a time, which is to analyze which candidate tag is more relevant from the overall perspective, which is conducive to determining a more accurate content tag. In this regard, for a better understanding, taking the target content as a travel guide to place A (including transportation to place A, routes to visit place A, photography skills, and food in place A, etc.), candidate tag a is the aforementioned "food" tag, and candidate tag b is the aforementioned "travel" tag as an example, the following example is provided:

[0049] Using the method in the related art, the absolute correlation between the label "food" and the target content was predicted once, and a correlation score of 0.61 was obtained, and the absolute correlation between the label "travel" and the target content was predicted once, and a correlation score of 0.58 was obtained. This may be because the target content includes a large amount of food in place A, and the word "travel" is mentioned less, so the above scoring results were obtained when the absolute correlation of a single label was predicted once.

[0050] After adopting this application, a single prediction is based on the target content and the two tags, so that it is possible to analyze which tag is relatively more relevant from the overall perspective. In this example, from the overall analysis, it is easier to determine that the target content focuses more on introducing matters related to traveling to place A, and food is only a part of it. Therefore, the prediction result obtained after the overall analysis can be used to indicate that the relevance of the "travel" tag is greater than the relevance of the "food" tag.

[0051] It can be seen that compared with the method of predicting absolute correlation in the related art, the method provided by the present application can analyze the correlation between two labels and the content as a whole, and can solve the problem that the absolute correlation is one-sided and causes the absolute correlations of multiple labels to be incomparable, which is conducive to determining more accurate content labels for the content.

[0052] Figure 3 This is a flowchart of a method for determining a content tag provided in an embodiment of the present application, which is described by taking a server as an example of the aforementioned computer device. The method includes S301-S304:

[0053] S301: Obtain M tag pairs corresponding to target content.

[0054] The target content may refer to the content for which the content label needs to be determined. Also, the M label pairs corresponding to the target content are composed of two different candidate labels among the N candidate labels corresponding to the target content. The N candidate labels may refer to labels that are preliminarily determined and may be used to describe the target content. Through the method provided by the present application, a label that can more accurately describe the target content can be determined from these N candidate labels (i.e., the content label of the target content is determined). Where N is an integer greater than 1, and M is an integer greater than or equal to 1.

[0055] It should be noted that this application does not make any limitation on the target content. For ease of understanding, the embodiments of this application provide the following examples:

[0056] As an example, from the source of the content, the target content can be provided by the content provider, or it can be the content published by the user on the content platform, etc. Among them, the content platform can be, for example, an application, a website, etc., which can support users to publish content, support users to obtain content published by other users, and also support users to interact based on content, etc. In actual applications, the content published by users on the content platform can be called user-generated content (UGC).

[0057] As another example, from the perspective of content type, the target content can be text-type content (i.e., pure text content), image-type content, video-type content, voice-type content, or a mixture of at least two of the above types, etc.

[0058] It should also be noted that the present application does not impose any limitation on how to obtain the M tag pairs. For ease of understanding, the present application provides the following example description:

[0059] In a possible implementation, the server may first obtain N candidate tags corresponding to the target content, and then pair every two different candidate tags among the N candidate tags to form M tag pairs.

[0060] It is understandable that this step can be implemented only when it is necessary to determine content tags for the target content, so that the tags can be paired in pairs as needed. Of course, it can also be implemented in advance, that is, M tag pairs can be pre-built, so that batch processing is convenient. For example, taking the target content as the content to be recommended published on a certain content platform, the content platform can collect all the content to be recommended published within a certain period of time and build the tag pairs corresponding to each content to be recommended. Afterwards, when enough content to be recommended is collected, the method provided in this application can be implemented to determine the corresponding content tag for each content to be recommended, and then recommend this batch of content to be recommended.

[0061] S302: For each label pair among the M label pairs, construct input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to each of the M label pairs.

[0062] For each tag pair, the server can construct the input data corresponding to the tag pair according to the tag pair and the target content. Accordingly, the input data corresponding to the M tag pairs can be obtained.

[0063] The input data corresponding to the tag pair can be used to indicate the situation of the target content and the situation of the two candidate tags included in the tag pair, that is, the overall input data is used to reflect the situations of the three.

[0064] S303: predicting the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtaining the prediction result corresponding to the tag pair.

[0065] For each tag pair, the server can predict the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtain the prediction result corresponding to the tag pair.

[0066] Since the input data corresponding to the label pair can be used to indicate the situation of the target content and the situation of the two candidate labels included in the label pair, that is, the situation of the three is reflected by using this one piece of input data as a whole. Accordingly, in a single prediction, the overall situation of the three can be understood by using this one piece of input data, and then the relative correlation between the two candidate labels and the target content can be analyzed as a whole. That is, it is possible to comprehensively analyze the overall degree of correlation between the two candidate labels and the target content. Accordingly, the prediction result corresponding to the label pair can be used to indicate the magnitude relationship of the correlation between the two candidate labels included in the label pair and the target content, that is, it can indicate which of the two candidate labels included in the label pair has a higher correlation with the target content, or the two have the same degree of correlation with the target content. That is, the prediction result can be used to indicate the order relationship of the correlation between the two candidate labels.

[0067] It can be understood that after predicting the input data corresponding to the M label pairs respectively, the prediction results corresponding to the M label pairs respectively can be obtained.

[0068] And, it should be noted that this application does not impose any restrictions on how to make predictions based on input data. In practical applications, in order to improve prediction efficiency, the prediction model can be used to implement the aforementioned S303, that is: according to the input data corresponding to the label pair, the correlation between the two candidate labels included in the label pair and the target content is predicted by the prediction model to obtain the prediction result corresponding to the label pair. Based on this, in a single prediction, the target content and two different candidate labels can be sent to the prediction model together using this input data, so that the prediction model can understand the overall situation of the three, and then analyze the relative correlation between the two candidate labels and the target content as a whole. That is, it is possible to comprehensively analyze the overall degree of correlation between the two candidate labels and the target content.

[0069] The prediction model can be used to predict the correlation between the two candidate labels in the label pair corresponding to the input data and the target content to give the aforementioned prediction results. It should be noted that this application does not make any restrictions on the setting of the prediction model. For ease of understanding, the following examples are provided in the embodiments of this application:

[0070] As an example, the aforementioned prediction model can be obtained by constructing sample input data with sample identifiers in a supervised learning manner and using the sample input data to train the neural network model. The sample input data can be constructed based on sample content and a sample label pair, the sample label pair includes two sample labels, and the sample identifier can be used to indicate the magnitude relationship between the two sample labels and the sample content.

[0071] As another example, a large model can also be selected as the aforementioned prediction model. Among them, the large model can refer to a machine learning model with powerful computing power, which has strong expression and learning capabilities in various task reasoning, and therefore can be applied to the correlation prediction task between labels and content of this application. Usually, the large model can select an existing large model (such as a general large model, a large language model, etc.), so that by directly reusing the existing model, it is beneficial to save costs. For example, reuse can be achieved by calling the application programming interface (Application Programming Interface, API) of the existing large model.

[0072] S304: Determine a content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs.

[0073] Since the prediction result corresponding to a single label pair can reflect which of the two candidate labels included in the label pair is more relevant to the target content, or the two have the same degree of relevance to the target content. Correspondingly, based on the prediction results corresponding to the M label pairs, it can be reflected which of the N candidate labels is relatively more relevant to the target content, or the two have the same degree of relevance to the target content. In this way, the relative relationship between the relevance of the N candidate labels and the target content can be reflected, that is, the order relationship between the relevance of the N candidate labels can be determined.

[0074] Therefore, the server can determine the content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs. Among them, the content label, as a label of the target content, can be used to describe the target content (such as describing what type of content the target content is, what information it contains, etc.). Specifically, the content label can refer to the candidate label that meets the relevance condition among the N candidate labels, and refers to the candidate label that can most accurately describe the target content among the N candidate labels. In addition, the content label can be used to recommend the target content.

[0075] It can be seen from the above technical solution that M label pairs corresponding to the target content are obtained, and the M label pairs are composed of two different candidate labels among the N candidate labels corresponding to the target content. For each label pair, the input data corresponding to the label pair is constructed according to the label pair and the target content, and the correlation between the two candidate labels included in the label pair and the target content is predicted based on the input data to obtain the prediction result corresponding to the label pair. It can be seen that the N candidate labels are constructed into pairs, and the label pair and the target content are constructed into input data, and in a single prediction, the input data constructed based on the target content and the two different candidate labels are predicted as a whole, so that the correlation between the two candidate labels and the target content can be analyzed from the overall perspective. Based on this, the correlation between the two candidate labels and the target content can be comprehensively analyzed. The corresponding prediction result can be used to indicate the size relationship between the correlation between the two candidate labels included in the label pair and the target content. Compared with the method in the related art that only predicts based on a single label and content and analyzes the absolute correlation between the label and the content, the correlation between the two labels and the content is analyzed as a whole, so it can solve the problem that the absolute correlation is one-sided and the absolute correlation between multiple labels is not comparable. In this way, when determining the content label for the target content from the N candidate labels based on the prediction results corresponding to the M label pairs, it is helpful to determine the label that is relatively more relevant to the target content, that is, to determine a more accurate content label. And, based on the content label, it is helpful to recommend the target content more accurately.

[0076] Through the above embodiments, the content label determination method provided by this application is described. It should also be noted that this application does not make any restrictions on how to construct the input data corresponding to the label pair, how to determine the content label from N candidate labels, and how to recommend the target content. For better understanding, this application will be illustrated one by one through the following embodiments:

[0077] (I) Regarding how to construct the input data corresponding to the tag pair, the present application embodiment provides the following example description:

[0078] In actual applications, tags are generally short, such as keywords, phrases, etc., which are intended to simply and intuitively reflect the content (such as what information is in the content, etc.). For example, tags can be "food", "travel", "movie", "xx cola", etc.

[0079] In order to improve the ability to understand the input data, specifically to improve the ability to understand the candidate labels in the input data, so as to be able to predict the relevance more accurately, in a possible implementation, the label description information of the label can also be combined to assist in better understanding the candidate labels (such as when the prediction model is used for prediction, combining the label description information can assist the prediction model to more accurately understand the candidate labels). For better understanding, the two candidate labels included in the label pair are the first candidate label and the second candidate label as an example for explanation. In specific implementation:

[0080] The server may obtain first tag description information corresponding to the first candidate tag and obtain second tag description information corresponding to the second candidate tag. Then, the server may construct input data corresponding to the tag pair according to the target content, the tag pair, the first tag description information, and the second tag description information.

[0081] Among them, the first label description information can be used to describe the situation of the first candidate label, and the second label description information can be used to describe the situation of the second candidate label. Accordingly, in a single prediction, it is possible to understand what the first candidate label is in combination with the first label description information, and to understand what the second candidate label is in combination with the second label description information. Based on this, the candidate labels can be understood more accurately. On the basis of accurately understanding the candidate labels, predicting the correlation between the two candidate labels and the target content is conducive to obtaining more accurate prediction results.

[0082] It should be noted that this application does not impose any limitation on the setting of tag description information. For ease of understanding, the following examples are provided in the embodiments of this application:

[0083] In actual applications, the tag type can reflect what type of tag it is, and is used to reflect the main theme or category of the content. For example, the tag type of the aforementioned "food", "travel", "movie", etc. can be a "theme tag", which can reflect the general theme of the content after corresponding to a certain content. For another example, the tag type of the aforementioned "xx cola" can be a "brand tag", which can reflect that the content may be a product promotion of the xx brand after corresponding to a certain content. Therefore, in a possible implementation method, the tag type can be used as tag description information.

[0084] Usually, labels are relatively short, such as keywords, phrases, etc., because the same word may have different semantics in different scenarios. Therefore, in one possible implementation, the explanatory text of the label can be used as label description information, wherein the explanatory text can be used to indicate what the current label is, and to assist in understanding the semantics of the label in the current scenario. For example, the semantics of the label "apple" can refer to a kind of fruit or a product brand. After adding the explanatory text, it is helpful to more accurately understand the meaning of the label "apple" in the current scenario, thereby improving the prediction accuracy.

[0085] Of course, in practical applications, the tag description information may also include the aforementioned tag type and explanation text to assist in understanding the tag more comprehensively. In addition, other information that is helpful in assisting in understanding the tag may also be set as the tag description information as needed, and this application does not impose any restrictions. For example, the content that is often associated with the tag, etc.

[0086] Usually, the content generally includes a lot of information, so the data volume of the content itself is relatively large. For example, an article with 10,000 words. In order to improve the prediction efficiency, in a possible implementation method, the content can be first extracted to give a concise overview of the content, and then the input data can be constructed.

[0087] In specific implementation, the server may extract a summary of the target content to obtain a content summary corresponding to the target content. Afterwards, the server may construct input data corresponding to the tag pair based on the content summary corresponding to the target content and the tag pair.

[0088] The content summary refers to a brief and comprehensive overview of the target content, which can be used to indicate the core information, main viewpoints, main ideas, etc. of the target content. Based on the content summary, the target content can be quickly understood. Based on this, extracting the content summary first is conducive to reducing the amount of data. Therefore, after constructing the input data based on the content summary, it is conducive to improving the efficiency of understanding the input data, thereby accelerating the prediction process and improving the prediction efficiency.

[0089] In this regard, it should be noted that this application does not impose any limitation on how to extract the summary. For ease of understanding, this application embodiment takes the case where the target content is multimodal content belonging to multiple modes as an example, and provides the following example description:

[0090] It is understandable that different modal contents have different abilities to convey information. In multimodal content, different modal contents can reflect the situation of the target content from different angles. Combining multiple modal contents can more accurately and comprehensively express the information that the multimodal content wants to convey. For example, in the target content of mixed pictures and texts, the content of one modality, pictures or texts, can also reflect the situation of the target content, but after combining the content of these two modalities, it can accurately reflect the information that the target content of mixed pictures and texts wants to convey.

[0091] Therefore, in order to reduce the amount of calculation while taking into account the accuracy of understanding, that is, to accurately understand the target content based on the content summary, in one possible implementation method, a summary can be extracted for the content of each modality, and then a content summary can be synthesized to represent the target content.

[0092] In specific implementation, the server can extract the summary of each modality of the multimodal content to obtain the content summary corresponding to each of the multiple modalities. Then, the server can determine the content summary corresponding to the target content according to the content summary corresponding to each of the multiple modalities.

[0093] It can be seen that for the scenario of multimodal content, a content summary is extracted for each modal content, wherein the content summary corresponding to each modal content can be used to indicate the core information of the content of this modality, etc. Afterwards, the content summaries corresponding to the contents of multiple modalities are combined to form a content summary of the target content, which is more conducive to retaining the information to be conveyed by the multimodal content. Based on this, on the basis of reducing the amount of calculation, it is conducive to taking into account the accuracy of understanding, so that the target content can also be accurately understood based on the content summary.

[0094] It is understandable that when producing the content, the content producer may mark some information as the content identifier of the content to assist the content acquirer in understanding the content. Accordingly, in order to better understand the target content, in another possible implementation, the input data may also be constructed in combination with the content identifier corresponding to the content.

[0095] For ease of understanding, the present application embodiment takes the content published on the aforementioned content platform as an example for explanation:

[0096] If the target content is the content to be recommended published by the target object on the content platform, the target object may refer to the producer of the target content or the user of the content platform, and the content platform may be the aforementioned application. In actual applications, after a user publishes a certain content on the content platform, the content platform may review the content and make targeted recommendations, so the target content can be considered as the content to be recommended.

[0097] In this regard, the server can obtain the content identifier corresponding to the target content. The content identifier can be input by the target object through the content platform for the target content, that is, the content identifier is the information marked by the target object for the target content. Then, the server can construct the input data corresponding to the tag pair according to the target content, the content identifier, and the tag pair.

[0098] Generally, content identifiers can also be considered as some labels that are subjectively considered by the content producer to be suitable for the content. Therefore, constructing input data in combination with the content identifiers that the content producer has annotated for the content can help quickly understand the target content and understand the correlation between the target content and the two candidate labels in the input data, thereby improving prediction accuracy.

[0099] Accordingly, in the recommendation phase, the server can recommend the target content in the content platform based on the content tag and content identifier. Based on this, it is possible to recommend the content in the content platform by combining the content identifier subjectively marked by the content producer and the content tag determined by analyzing the correlation, which is conducive to improving the accuracy of the recommendation. In addition, since the content identifier subjectively marked by the producer can reflect the producer's expectation of the recommendation of the content, this recommendation method is also conducive to improving the enthusiasm of the content producer.

[0100] It is understandable that different prediction models may have different ways of understanding input data. Therefore, in another possible implementation, the input data can also be constructed in combination with the specific situation of the prediction model to facilitate the prediction model to better understand the input data. For ease of understanding, the embodiment of the present application takes the prediction model as the aforementioned large model as an example for illustration:

[0101] In practical applications, the large model decodes the information carried in the input data in sequence, thereby converting it into a data form that the large model can understand (such as a feature vector, etc.). For this reason, the large model is sensitive to the order of the information carried in the input data. In the scenario where the present application is applied to predict the magnitude of the correlation between two candidate labels included in the label pair, the order of the two candidate labels in the input data may affect the final prediction result, especially when the correlation between the two candidate labels and the target content is close.

[0102] Therefore, in order to improve the stability of the large model prediction and obtain more accurate prediction results, in one possible implementation method, for each label pair, the order of the two candidate labels included in the label pair can be swapped, and the large model can be used to predict twice, and the final prediction result can be determined by combining the two prediction results.

[0103] In a specific implementation, if the aforementioned prediction model is a large model, the server can construct the first input data corresponding to the label pair and the second input data corresponding to the label pair according to the label pair and the target content. In the first input data, the two candidate labels included in the label pair are arranged in a first order, and in the second input data, the two candidate labels included in the label pair are arranged in a second order, and the second order is different from the first order. It can be seen that for a label pair, the order of the two candidate labels included therein is interchanged, thereby constructing two pieces of input data.

[0104] Correspondingly, when the aforementioned S303 is implemented, the server can predict the correlation between the two candidate tags included in the tag pair and the target content through the large model based on the first input data corresponding to the tag pair, and obtain the first prediction result corresponding to the tag pair; and predict the correlation between the two candidate tags included in the tag pair and the target content through the large model based on the second input data corresponding to the tag pair, and obtain the second prediction result corresponding to the tag pair.

[0105] Among them, the first prediction result can be used to indicate the magnitude relationship of the correlation between the two candidate tags predicted by the large model in the first order and the target content, and the second prediction result can be used to indicate the magnitude relationship of the correlation between the two candidate tags predicted by the large model in the second order and the target content. Then, the server can determine the target prediction result corresponding to the tag pair based on the first prediction result and the second prediction result. Among them, the target prediction result is the final prediction result obtained by combining the two prediction results, which is used to indicate the magnitude relationship of the correlation between the two candidate tags and the target content.

[0106] Correspondingly, when the aforementioned S304 is specifically implemented, the server may determine the content label corresponding to the target content from the N candidate labels according to the target prediction results corresponding to the M label pairs.

[0107] It can be seen that for each label pair, two input data are constructed by exchanging the order of the two candidate labels, and the final prediction result is obtained by combining the two prediction results, avoiding the problem that a single prediction may be biased due to the large model being sensitive to order.

[0108] In the related art, the labeling model is trained to determine the corresponding content label for the target content. In the specific implementation, a batch of sample contents are labeled with their corresponding content labels (this process is mostly manual labeling), so as to construct a training set to train the labeling model. This type of method relies heavily on manual labeling. In addition, the labeling model obtained in this way is essentially a classification model, which can only predict that some of the labels learned in the training stage may be able to describe the target content. In this way, when there are new labels, the related technology also needs to use the new labels to supplement the labeling model. In comparison, the large model also has a good understanding of the new labels, so there is no need for additional training. Therefore, the method of using a large model to determine the prediction results is conducive to reducing the dependence on manual labeling and saving certain labor costs.

[0109] Also, from the perspective of labels, it can be considered that each sample content labeled with a certain label is a positive sample of the label, and all other content not labeled with the label is a negative sample of the label. Generally speaking, for each label, sufficient positive samples (such as more than 50) and negative samples are required to train the model so that the model can accurately predict the label. However, in practical applications, negative samples are generally sufficient, while positive samples are relatively few. Labels with insufficient positive samples generally find it difficult to obtain good model results. In this regard, it is generally possible to filter out labels with insufficient positive samples (that is, these labels are not trained), and only use labels with sufficient positive samples for training. However, this also leads to limited classification capabilities of the model. In this regard, the ability to understand various labels is broadened after adopting a large model.

[0110] In addition, because the content of different businesses varies greatly, the labeling model is prone to overfitting in each business and is difficult to generalize to other businesses, that is, the generalization is poor, so different labeling models need to be trained for different businesses. In contrast, when using a large model to determine the prediction results, the reasoning ability of the large model can be used to achieve good prediction results for any content or any label pair. In other words, the large model has a good understanding of various content from different businesses, so it can directly determine content labels for various content from different businesses without additional training.

[0111] It should be noted that the present application does not make any limitation on how to determine the final target prediction result based on the two prediction results. For ease of understanding, the present application provides the following example description:

[0112] If the first prediction result is the same as the second prediction result, it means that for the two candidate labels in the current label pair, no matter in what order, the large model gives the same prediction result. Therefore, it is considered that the prediction of the large model is accurate. Therefore, the judgment result of the large model can be adopted. Accordingly, the first prediction result or the second prediction result can be determined as the target prediction result corresponding to the label pair.

[0113] On the contrary, if the first prediction result is different from the second prediction result, it indicates that for the two candidate tags in the current tag pair, after the order is swapped, the large model gives different prediction results, so it is believed that the large model cannot accurately judge the magnitude relationship of the correlation between the two candidate tags, which may be because the correlation between the two candidate tags and the target content is relatively close. Therefore, at this time, it can be determined that the target prediction result corresponding to the tag pair indicates that the first correlation degree is the same as the second correlation degree, wherein the first correlation degree can be used to indicate the correlation between the first candidate tag included in the tag pair and the target content, and the second correlation degree can be used to indicate the correlation between the second candidate tag included in the tag pair and the target content.

[0114] It can be seen that in the method of swapping the order of prediction twice, if the large model outputs the same prediction result twice, the judgment result of the large model is adopted, otherwise it is not adopted to avoid prediction bias. And when the two prediction results are different, the correlation between the two is determined to be consistent, and there is no relationship between which one is more relevant to the target content than the other. Based on this, on the basis of using the large model for prediction, the prediction bias problem that the large model is sensitive to the order can also be taken into account to obtain a more accurate target prediction result.

[0115] In practical applications, in order to facilitate the large model to better understand the prediction task and give more accurate prediction results, in a possible implementation, the input data can also be constructed in combination with prompts to instruct the large model to think and understand the prediction task according to the ideas set in the prompts, thereby giving better prediction results. For ease of understanding, the embodiment of the present application uses the large model as AAA and provides the following prompt example:

[0116]

[0117] Based on this, the large model AAA can be guided to think about and understand what the current prediction task is and what form the current prediction results need to be given according to the above-mentioned ideas. This is conducive to guiding the large model AAA to better understand the input data and give more accurate prediction results.

[0118] It should also be noted that this application does not make any restrictions on the setting method of the large model. As mentioned above, the large model can select an existing large model (such as a general large model, a large language model, etc.). In this way, by directly reusing the existing model, it is beneficial to save costs. For better understanding, the following examples are provided in the embodiments of this application:

[0119] Different big models may have different understanding capabilities for different types of data. For example, some big models have stronger performance in text understanding, while some big models have stronger performance in image understanding. Therefore, in one possible implementation, a big model can be flexibly selected according to the type of target content.

[0120] In practical applications, the target content mostly includes text-type content. Therefore, as an example, the large model can be a large language model (LLM). Among them, LLM is obtained by training with a large amount of text data, based on which it can fully learn the statistical laws and semantic information of the language, so as to predict the next word or sentence. LLM has a wide range of applications in the field of natural language processing (Natural Language Processing, NLP), such as machine translation, speech recognition, text generation, etc., and with the continuous expansion of the model scale and the increase of training data, its performance and application scope are also constantly expanding. LLM can adopt different training strategies and model structures, such as pre-training and fine-tuning.

[0121] Pre-training can refer to large-scale training of the model in an unsupervised or self-supervised manner, so that it has a general understanding and expression ability of language data. Fine-tuning is to optimize the model for specific application scenarios or tasks through supervised learning based on the pre-trained model. For example, in the content recommendation scenario targeted in this application, it can be fine-tuned to improve the model's ability to predict the size of the correlation between two candidate labels in the input data and the target content.

[0122] In some scenarios, the target content may be multimodal content (such as mixed text and image content), so as another example, the large model may be a multimodal large language model (MLLM). Based on LLM, data of other modalities (such as images, videos, audio, etc.) are integrated to enable simultaneous processing of information of different modalities, better understanding and expression of semantics, and thus improving the ability to understand the target content, the corresponding prediction effect and accuracy.

[0123] Generally, when MLLM recognizes multimodal content, optical character recognition (OCR) technology can be used to detect and recognize text content from images to understand the information carried by the images. In addition, automatic speech recognition (ASR) technology can be used, which is a technology that converts speech into text. Based on this, text content can be detected and recognized from video or audio to understand the information carried by the video or audio.

[0124] (II) Regarding how to determine a content tag from N candidate tags, the present application provides the following example description:

[0125] It is understandable that the prediction result corresponding to a single tag pair can reflect which of the two candidate tags included in the tag pair is more relevant to the target content, or the two have the same degree of relevance to the target content. Correspondingly, based on the prediction results corresponding to the M tag pairs, it can be reflected that among the N candidate tags, which of the two candidate tags is relatively more relevant to the target content, or the two have the same degree of relevance to the target content. In this way, the relative relationship between the N candidate tags and the target content can be reflected.

[0126] Therefore, in a possible implementation, the N candidate tags can be fully sorted based on the prediction results corresponding to the M tag pairs, so as to determine the relative relevance of the N candidate tags, and then select several tags that can better represent the target content as content tags. Based on this, the full sorting method can intuitively reflect the size relationship between the candidate tags to support downstream applications.

[0127] Considering that the prediction is made in pairs, the correlation degree can reflect the relative correlation between the two candidate tags, and in most cases, N may be greater than 2. In order to screen out the N content tags that are relatively more relevant as a whole, in another possible implementation, a single candidate tag can be evaluated from the N candidate tags as a whole, and the relative correlation degree among the N tags can be determined before determining the content tag.

[0128] It is understandable that when N is greater than 2, a single candidate label will exist in at least two label pairs, and multiple prediction results related to the candidate label will be obtained. Therefore, based on this, the relative relevance of the candidate label in the N wholes can be comprehensively evaluated.

[0129] In a specific implementation, if N is an integer greater than 2, and M is an integer greater than 2, the server can determine the correlation score between the candidate tag and the target content for each of the N candidate tags according to the prediction results corresponding to the R target tag pairs associated with the candidate tag. Among them, the R target tag pairs refer to the tag pairs including the candidate tag among the M tag pairs, and R is an integer greater than 1 and less than M. Then, the server can determine the content tag corresponding to the target content from the N candidate tags according to the correlation scores corresponding to the N candidate tags.

[0130] In this way, for each candidate tag, the prediction results of all tag pairs including the candidate tag can be integrated to evaluate the relevance of the candidate tag to the target content from the overall perspective, that is, the relevance score corresponding to the candidate tag can be used to indicate the relative relevance of the candidate tag among the N overall ones, which can better represent how relevant the candidate tag is to the target content among the current N candidate tags, so it is more accurate and intuitive. Accordingly, when determining the content tag based on this, it is helpful to select the candidate tag that is more relevant from the overall perspective as the content tag.

[0131] It should also be noted that the present application does not make any limitation on how to determine the relevance score and how to determine the content tag based on the relevance score. For ease of understanding, the present application provides the following example description:

[0132] First, the following example explains how the relevance score is determined:

[0133] In a possible implementation, for any candidate tag, its corresponding relevance score can be determined by the following formula:

[0134]

[0135] In the above formula, d i Can be used to indicate the i-th candidate tag among N candidate tags, d j It can be used to indicate the jth candidate tag among N candidate tags, that is, i and j are both integers greater than or equal to 1 and less than or equal to N. Since every two different candidate tags are paired, i and j are different numbers.

[0136] And, in the above formula, S i It can be used to indicate the relevance score corresponding to the i-th candidate tag. δ() can be a symbolic function. When the condition in the brackets is true, the value is 1, otherwise, the value is 0.

[0137] Correspondingly, the above formula can be understood as:

[0138] For the i-th candidate label, pair it with other candidate labels and get the corresponding prediction results. If the relevance of the i-th candidate label is greater than that of the j-th candidate label, it can be recorded as d i >d j , and accumulate 1 point. If the correlation between the two is the same, it can be recorded as d i =d j , and accumulate 0.5 points, otherwise accumulate 0 points. Since for any candidate tag, the number of target tag pairs associated with it can be N-1, it can be divided by N-1 in the end to normalize the cumulative score based on N-1 target tag pairs, and map the final correlation score to between 0 and 1, so as to more intuitively reflect the degree of correlation of each candidate tag.

[0139] In practical applications, time complexity can be used to describe the relationship between the time required for the algorithm to execute and the size of the input data. It is usually expressed using Big O notation, that is, O(f(n)), where n is the size of the input data, and f(n) is the relationship between the algorithm running time and the input data size. In the above example, for the size of N candidate labels, the time complexity of this method of converting to relevance scores is O(N 2 ).

[0140] Next, the following example explains how to determine content labels based on relevance scores:

[0141] As an example, the N candidate tags can be fully sorted based on their corresponding relevance scores, and then content tags can be filtered from the N candidate tags based on the results of the full sorting, such as filtering several candidate tags with the highest relevance scores as content tags. Based on this, the relative size of the N candidate tags can be intuitively reflected through the full sorting method.

[0142] As another example, considering that in actual downstream applications, only the candidate tags with the highest relevance scores need to be used to recommend the target content, the Top K (K is an integer greater than or equal to 1 and less than N) candidate tags can be directly selected as content tags based on the relevance scores corresponding to the N candidate tags.

[0143] In practical applications, a min heap is a special complete binary tree, which is a type of heap structure and is often used to implement a priority queue, where the minimum element can be found and removed efficiently. Also, in a min heap, the value of each parent node is less than or equal to the values ​​of all its child nodes, which means that the root node at the top of the binary tree (the top of the tree) is the smallest element in the entire heap.

[0144] Therefore, in the present application, a binary tree such as a small top heap can be used to screen the Top K candidate tags from N candidate tags. In a specific implementation, if the number of content tags is K, the server can construct a binary tree including K nodes according to the correlation scores corresponding to the K candidate tags among the N candidate tags. Among them, the binary tree has the following characteristics: K nodes are used to indicate the correlation scores corresponding to the K candidate tags respectively, different nodes are used to indicate the correlation scores corresponding to different candidate tags among the K candidate tags, the correlation score indicated by the parent node is less than or equal to the correlation score indicated by the child node corresponding to the parent node, the root node can be used to indicate the minimum value of the K correlation scores, the parent node can refer to a node with child nodes, and the root node can be a parent node located at the top level of the binary tree. That is, a small top heap including K nodes is constructed.

[0145] Next, the server can traverse the W candidate tags except the K candidate tags among the N candidate tags in turn, where W = NK. When traversing to the rth candidate tag, the relevance score corresponding to the rth candidate tag is compared with the relevance score indicated by the root node to obtain a comparison result, where r is an integer greater than or equal to 1 and less than or equal to W.

[0146] If the comparison result indicates that the relevance score of the rth candidate label is greater than the relevance score indicated by the root node, it indicates that the K relevance scores stored in the current binary tree are not Top K, so the relevance score indicated by the root node can be updated to the relevance score of the rth candidate label to obtain the pending tree. In addition, the pending tree can be updated until the pending tree meets the characteristics of a binary tree and the pending tree is determined to be a binary tree.

[0147] If the comparison result indicates that the relevance score of the rth candidate tag is less than or equal to the relevance score indicated by the root node, it means that among the K relevance scores stored in the current binary tree and the relevance score of the rth candidate tag, the K relevance scores stored in the current binary tree are the Top K, so the rth candidate tag can be eliminated at this time.

[0148] Finally, the server may determine the K candidate tags corresponding to the K correlation scores stored in the target binary tree as the K content tags corresponding to the target content. The target binary tree refers to the binary tree obtained when the Wth candidate tag is traversed.

[0149] Based on this, by using a binary tree such as a mini-heap, the top K relevance scores among N relevance scores can be found and removed quickly and efficiently, that is, the top K candidate tags can be found quickly and efficiently, which is conducive to improving the efficiency of determining content tags.

[0150] In practical applications, the time complexity of using a small top heap to select the Top K from N candidate tags is O(N log(K)).

[0151] For a better understanding, the process of updating the above-mentioned undetermined tree until the undetermined tree satisfies the characteristics of a binary tree is described by way of example:

[0152] In practical applications, the operations of the mini-heap mainly include insertion and deletion. Insertion can mean that when a new element is inserted, the new element is placed at the end of the array, and then "floats up" by comparing with the parent node and exchanging positions until the heap order characteristic is met, that is, the characteristic of the aforementioned binary tree. And, deletion can mean that when deleting the smallest element (that is, the element stored in the root node), the root node is usually replaced with the element at the end of the array (that is, the element at the end of the array is greater than the element at the root node), and then "sinks" by comparing with the child nodes and exchanging positions until the heap order characteristic is met, that is, the characteristic of the aforementioned binary tree. Whether it is insertion or deletion, it depends on a function that compares the size of nodes to determine whether to float up or sink.

[0153] In the task of selecting Top K from N candidate tags in this application, the above-mentioned deletion operation and sinking method can be used to update the pending tree. Specifically:

[0154] In the pending tree, the relevance score indicated by the root node is the relevance score of the rth candidate label. It can be compared with the child node corresponding to the root node, and the child node with a smaller relevance score than the root node can be exchanged to the root node (if there are two child nodes, and both child nodes are smaller than the root node, then exchange them with the smallest child node). Next, the node storing the relevance score of the rth candidate label is compared with its child nodes, and the child node with a smaller relevance score than it can be exchanged (similarly, if there are two child nodes, and both are smaller than the current node, then exchange them with the smallest child node). This is done until the aforementioned characteristics of the binary tree are met.

[0155] When the binary tree property is satisfied, the root node stores the smallest relevance score, and the next candidate tag can be traversed until the last candidate tag (i.e., the Wth candidate tag) is traversed, and the Top K is selected.

[0156] (III) Regarding how to recommend target content, the present application provides the following examples:

[0157] As in the above-mentioned embodiment, in some scenarios, the target content may be recommended in the content platform, for example, the target content may be recommended to those users in the content platform who may want to obtain the target content, thereby achieving targeted recommendation.

[0158] In addition, in some other scenarios, it is also possible to determine a suitable recommendation channel for the target content based on the content tag, and recommend the content through this recommendation channel. Among them, the recommendation channel can be used to display the target content, such as an application (such as some social applications), a search website, etc. In this regard, for ease of understanding, this application also provides the following example description:

[0159] In specific implementation, the server can determine the target object group corresponding to the target content from multiple object groups based on the content tag corresponding to the target content and the group characteristics corresponding to the multiple object groups. Among them, the group characteristics corresponding to the object groups can reflect the situation of this object group (specifically, they can reflect the demand and preference of this object group for obtaining content, etc.), and the determined target object group can be considered as the potential audience of the target content. Therefore, the server can then determine the target recommendation channel from multiple recommendation channels based on the group characteristics of the target object group and the channel object characteristics corresponding to the multiple recommendation channels, and recommend the target content through the target recommendation channel.

[0160] Among them, the channel object characteristics corresponding to the recommendation channel can reflect the situation of the users in this recommendation channel (specifically, it can reflect the needs and preferences of these users for obtaining content, etc.). The channel object characteristics corresponding to the target recommendation channel and the group characteristics of the target object group meet the object characteristic similarity conditions, that is, it is considered that the users in the target recommendation channel are relatively close to the potential audience of the target content, so recommendations can be made in the target recommendation channel to improve the accuracy of recommendations.

[0161] Through the above embodiments, the content label determination method provided by the present application is described in detail, specifically, determining a content label for the target content from the N candidate labels corresponding to the target content, and recommending the target content based on the determined content label. It should also be noted that the present application does not make any limitation on how to determine the N candidate labels corresponding to the target content. For ease of understanding, the present application embodiment provides the following example description:

[0162] It is understandable that similar content usually has similar content tags. Therefore, in a possible implementation, the tags of content similar to the target content can be used as candidate tags for the target content. In specific implementation, the aforementioned N candidate tags can be determined in the following manner:

[0163] First, the server may determine the content features corresponding to the target content through a feature encoder according to the target content. The content features of the target content may be used to indicate the situation of the target content (such as a feature vector corresponding to the target content). Next, the server may determine the target reference content features that meet the content feature similarity condition with the content features of the target content from the multiple reference content features according to the content features corresponding to the target content and the multiple reference content features stored in the search library.

[0164] The reference content feature is a content feature corresponding to the reference content determined by the feature encoder, and can be used to indicate the situation of the corresponding reference content (such as a feature vector corresponding to the reference content, etc.). Since the target reference content feature and the content feature of the target content meet the content feature similarity condition, it can be considered that the reference content corresponding to the target reference content feature is similar to the target content. Therefore, the server can determine N candidate tags based on the tags of the target reference content feature stored in the search library.

[0165] It can be seen that the candidate tags of the target content can be determined based on the tags of the content similar to the target content. Based on this, for any target content, N candidate tags can be initially determined for it by using the reference content with tags stored in the search library, which is more convenient.

[0166] It is understandable that in actual applications, the number of target reference content features can be multiple (for example, the Top N reference content features can be screened), and these multiple features may have the same labels. Therefore, the labels of multiple target reference content features can be deduplicated, and the deduplicated labels can be used as candidate labels for the target content.

[0167] It should be noted that this application does not impose any limitations on the setting of the retrieval library and the setting of the reference content. For example, the reference content may be content that is manually labeled. After the content features are extracted from the reference content by the feature encoder, they can be written into the retrieval library in the form of <reference content features, label list> for subsequent use. When there is a new label, it is only necessary to collect a sufficient number of reference contents containing the label, process them, and write them into the retrieval library, so new labels can be supported. Usually, the construction of the retrieval library can be completed by offline calculation, thereby accelerating the efficiency of subsequently determining candidate labels for the target content.

[0168] During the use phase, after determining the characteristics of the target reference content, the label list corresponding to the target reference content can be directly extracted, and after deduplication and other processing, the candidate labels corresponding to the target content can be obtained.

[0169] It should also be noted that the feature encoder can be used to extract features from content, and this application does not impose any restrictions on the setting method of the feature encoder. For ease of understanding, the embodiments of this application provide the following example description:

[0170] In practical applications, you can flexibly select the appropriate feature encoder according to the type of content. For example, for plain text content, you can choose a pre-trained model based on a transformer (BidirectionalEncoder Representations from Transformers, BERT), and for multimodal content (such as video content), you can choose a multimodal Transformer model.

[0171] Among them, Transformer is a network structure model suitable for processing sequence data. It uses a full attention structure instead of a recurrent neural network (RNN) or a convolutional neural network (CNN). Its core idea is to model the relationship between each element in the input sequence and other elements as an attention weight matrix, and calculate the representation of each element through a self-attention mechanism. This structure enables Transformer to have a strong ability to capture long-distance dependencies in the sequence, thus achieving good performance in natural language processing (NLP) tasks. The Transformer model is widely used in the field of NLP, including machine translation, text summarization, question-answering systems and other tasks. It introduced the self-attention mechanism for the first time and surpassed the sequence-to-sequence model (Sequence-to-Sequence, Seq2Seq), such as the Long Short-Term Memory Network (Long Short-Term Memory, LSTM) and the Gated Recurrent Unit (GRU) in many aspects. The success of Transformer has also spawned many variant models based on the self-attention mechanism, such as the aforementioned BERT, Generative Pre-trained Transformer (GPT), etc. These models have achieved good performance in NLP tasks.

[0172] In practical applications, when extracting features from content, such as extracting features from text content, the Transformer-based feature encoder can process words or subwords in the text as basic units. In Transformer, words or subwords in the text can be recorded as tokens. In practical applications, the relevance score can be used to indicate the probability of generating a certain token.

[0173] In practical applications, the process of determining candidate tags for target content based on a retrieval library can also be called tag recall, and the subsequent process of processing candidate tags to determine content tags can also be called tag sorting. Among them, for tag recall, a tag recall module can be constructed for implementation. For example, a framework from content to content to tag (item-to-item-to-tag, i2i2t) can be adopted, and a tag recall module can be designed using a non-parametric retriever. Specifically, it can include the aforementioned construction of a retrieval library, combined with a feature encoder, to find a controllable number of candidate tags for the target content from a large number of tags, so that they can be sorted using a predictive model later. In practical applications, for example, a search framework for artificial intelligence similarity search can be used for implementation.

[0174] In actual applications, the number of candidate tag libraries (i.e., the number of tags in the search library) ranges from thousands to millions (depending on business needs). Therefore, the amount of data processing required to sort all tags is too large, and downstream applications do not need the sorting results of all tags. Therefore, based on this tag recall method, N possible candidate tags are preliminarily screened out, so that the number of tags entering the sorting stage is more reasonable. Based on this, it is beneficial to improve the overall processing efficiency.

[0175] For a more comprehensive understanding, the present application also provides the following embodiments: Figure 4 A schematic diagram of a label processing framework for target content is shown, specifically:

[0176] For the target content, in the label recall stage, the content features of the target content can be extracted through the feature encoder, and N candidate labels can be recalled from the retrieval library using the content features.

[0177] Next, in the tag ranking stage, the N candidate tags can be ranked to determine the full ranking or the Top K. Specifically, the N candidate tags can be compared in pairs, that is, every two different candidate tags constitute a tag pair, such as Figure 4In the example, (d1, d2), (d2, d3), etc. are label pairs. For each label pair, the corresponding input data is constructed for the target content. For example, the input data corresponding to the label pair (d1, d2) can be recorded as (target content, d1, d2). Then, the input data can be input into the big model, and the corresponding prediction results can be determined by the big model.

[0178] Finally, the prediction results corresponding to all label pairs can be converted into relevance scores, that is, the relevance score of each candidate label can be obtained. After that, the N candidate labels can be fully sorted, or the Top K can be screened.

[0179] Subsequently, content tags can be determined for the target content based on the full sorting results, or the top K candidate tags can be directly determined as the content tags of the target content. In addition, targeted recommendations for target content based on content tags are helpful to improve content distribution efficiency and reduce the cost of manual content review.

[0180] Through the above embodiments, the present application is described in detail. In summary, a sorting method based on pairwise comparison is provided, for the recalled N candidate tags, two by two are paired, and then the correlation between them and the target content is predicted relatively, so as to determine a more accurate content tag for the target content.

[0181] It should be noted that, based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods.

[0182] based on Figure 3 Corresponding to the content label determination method provided in the embodiment, the embodiment of the present application further provides a content label determination device 500, the content label determination device 500 includes an acquisition unit 501, a construction unit 502, a prediction unit 503 and a determination unit 504:

[0183] The acquisition unit 501 is used to acquire M tag pairs corresponding to the target content, wherein the M tag pairs are composed of two different candidate tags among the N candidate tags corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1;

[0184] The construction unit 502 is used to construct, for each label pair of the M label pairs, input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively;

[0185] The prediction unit 503 is used to predict the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtain the prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content;

[0186] The determining unit 504 is configured to determine a content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs respectively, and the content label is used to recommend the target content.

[0187] In a possible implementation, if the two candidate tags included in the tag pair are a first candidate tag and a second candidate tag, the construction unit is further configured to:

[0188] Acquire first tag description information corresponding to the first candidate tag, and acquire second tag description information corresponding to the second candidate tag;

[0189] Input data corresponding to the tag pair is constructed according to the target content, the tag pair, the first tag description information, and the second tag description information.

[0190] In a possible implementation, the construction unit is further used to:

[0191] Extracting a summary of the target content to obtain a content summary corresponding to the target content;

[0192] According to the content summary corresponding to the target content and the label pair, input data corresponding to the label pair is constructed.

[0193] In a possible implementation, if the target content is multimodal content belonging to multiple modes, the construction unit is further configured to:

[0194] Extracting summaries of the contents belonging to each mode in the multimodal content respectively, to obtain content summaries corresponding to the contents of the multiple modes respectively;

[0195] The content summary corresponding to the target content is determined according to the content summaries respectively corresponding to the contents of the multiple modes.

[0196] In a possible implementation, if the target content is content to be recommended published by the target object on the content platform, the construction unit is further configured to:

[0197] Acquire a content identifier corresponding to the target content, where the content identifier is input by the target object through the content platform for the target content;

[0198] Constructing input data corresponding to the tag pair according to the target content, the content identifier, and the tag pair;

[0199] The determining unit is further configured to recommend the target content in the content platform according to the content tag and the content identifier.

[0200] In a possible implementation manner, if N is an integer greater than 2, and M is an integer greater than 2, the determining unit is further configured to:

[0201] For each candidate tag among the N candidate tags, determine the correlation score between the candidate tag and the target content according to the prediction results corresponding to the R target tag pairs associated with the candidate tag, where the R target tag pairs refer to the tag pairs including the candidate tag among the M tag pairs, and R is an integer greater than 1 and less than M;

[0202] According to the relevance scores respectively corresponding to the N candidate tags, a content tag corresponding to the target content is determined from the N candidate tags.

[0203] In a possible implementation, if the number of the content tags is K, where K is an integer greater than or equal to 1 and less than N, the determining unit is further configured to:

[0204] According to the relevance scores respectively corresponding to K candidate tags among the N candidate tags, a binary tree including K nodes is constructed, wherein the binary tree has the following characteristics: the K nodes are respectively used to indicate the relevance scores corresponding to the K candidate tags, different nodes are used to indicate the relevance scores corresponding to different candidate tags among the K candidate tags, the relevance score indicated by the parent node is less than or equal to the relevance score indicated by the child node corresponding to the parent node, the root node is used to indicate the minimum value among the K relevance scores, the parent node refers to a node having child nodes, and the root node is the parent node located at the top level of the binary tree;

[0205] For the W candidate tags other than the K candidate tags among the N candidate tags, traverse the W candidate tags in sequence, and when traversing to the r-th candidate tag, compare the relevance score corresponding to the r-th candidate tag with the relevance score indicated by the root node to obtain a comparison result, where r is an integer greater than or equal to 1 and less than or equal to W;

[0206] If the comparison result indicates that the relevance score of the r-th candidate tag is greater than the relevance score indicated by the root node, the relevance score indicated by the root node is updated to the relevance score of the r-th candidate tag to obtain a pending tree;

[0207] Updating the pending tree until the pending tree satisfies the characteristics of the binary tree, and determining the pending tree as the binary tree;

[0208] If the comparison result indicates that the relevance score of the r-th candidate tag is less than or equal to the relevance score indicated by the root node, the r-th candidate tag is eliminated;

[0209] The K candidate tags corresponding to the K correlation scores stored in the target binary tree are determined as the K content tags corresponding to the target content, and the target binary tree refers to the binary tree obtained when the Wth candidate tag is traversed.

[0210] In a possible implementation, the construction unit is further used to:

[0211] According to the label pair and the target content, construct first input data corresponding to the label pair and second input data corresponding to the label pair, wherein in the first input data, two candidate labels included in the label pair are arranged in a first order, and in the second input data, the two candidate labels included in the label pair are arranged in a second order, and the second order is different from the first order;

[0212] The prediction unit is also used for:

[0213] According to the first input data corresponding to the label pair, the correlation between the two candidate labels included in the label pair and the target content is predicted by a large model to obtain a first prediction result corresponding to the label pair;

[0214] According to the second input data corresponding to the label pair, the correlation between the two candidate labels included in the label pair and the target content is predicted by the large model to obtain a second prediction result corresponding to the label pair;

[0215] Determining a target prediction result corresponding to the label pair according to the first prediction result and the second prediction result;

[0216] The determining unit is further configured to determine a content label corresponding to the target content from the N candidate labels according to the target prediction results respectively corresponding to the M label pairs.

[0217] In a possible implementation manner, the prediction unit is further configured to:

[0218] If the first prediction result is the same as the second prediction result, determining the first prediction result or the second prediction result as the target prediction result corresponding to the label pair;

[0219] If the first prediction result is different from the second prediction result, it is determined that the target prediction result corresponding to the label pair indicates that the first degree of correlation is the same as the second degree of correlation, the first degree of correlation is used to indicate the correlation between the first candidate label included in the label pair and the target content, and the second degree of correlation is used to indicate the correlation between the second candidate label included in the label pair and the target content.

[0220] In a possible implementation manner, the determining unit is further configured to:

[0221] Determining a target object group corresponding to the target content from the multiple object groups according to the content label corresponding to the target content and the group characteristics corresponding to the multiple object groups;

[0222] Determine a target recommendation channel from the plurality of recommendation channels according to the group characteristics of the target object group and the channel object characteristics corresponding to the plurality of recommendation channels, wherein the channel object characteristics corresponding to the target recommendation channel and the group characteristics of the target object group satisfy an object characteristic similarity condition;

[0223] The target content is recommended through the target recommendation channel.

[0224] In a possible implementation manner, the acquiring unit is further configured to:

[0225] According to the target content, determining content features corresponding to the target content by a feature encoder;

[0226] According to the content feature corresponding to the target content and a plurality of reference content features stored in a search library, determining a target reference content feature that satisfies a content feature similarity condition with the content feature of the target content from the plurality of reference content features, wherein the reference content feature is a content feature corresponding to the reference content determined by the feature encoder;

[0227] The N candidate tags are determined based on the tags of the target reference content features stored in the search library.

[0228] It can be seen from the above technical solution that M label pairs corresponding to the target content are obtained, and the M label pairs are composed of two different candidate labels among the N candidate labels corresponding to the target content. For each label pair, the input data corresponding to the label pair is constructed according to the label pair and the target content, and the correlation between the two candidate labels included in the label pair and the target content is predicted based on the input data to obtain the prediction result corresponding to the label pair. It can be seen that the N candidate labels are constructed into pairs, and the label pair and the target content are constructed into input data, and in a single prediction, the input data constructed based on the target content and the two different candidate labels are predicted as a whole, so that the correlation between the two candidate labels and the target content can be analyzed from the overall perspective. Based on this, the correlation between the two candidate labels and the target content can be comprehensively analyzed. The corresponding prediction result can be used to indicate the size relationship between the correlation between the two candidate labels included in the label pair and the target content. Compared with the method in the related art that only predicts based on a single label and content and analyzes the absolute correlation between the label and the content, the correlation between the two labels and the content is analyzed as a whole, so it can solve the problem that the absolute correlation is one-sided and the absolute correlation between multiple labels is not comparable. In this way, when determining the content label for the target content from the N candidate labels based on the prediction results corresponding to the M label pairs, it is helpful to determine the label that is relatively more relevant to the target content, that is, to determine a more accurate content label. And, based on the content label, it is helpful to recommend the target content more accurately.

[0229] The embodiment of the present application further provides a computer device, which may be a terminal. For example, the terminal is a smart phone:

[0230] Figure 6 The block diagram shows a partial structure of a smart phone provided in an embodiment of the present application. Figure 6 The smartphone includes: a radio frequency (RF) circuit 610, a memory 620, an input unit 630, a display unit 640, a sensor 650, an audio circuit 660, a wireless fidelity (WiFi) module 670, a processor 680, and a power supply 690. The input unit 630 may include a touch panel 631 and other input devices 632, the display unit 640 may include a display panel 641, and the audio circuit 660 may include a speaker 661 and a microphone 662. Those skilled in the art will appreciate that Figure 6 The structure of the smartphone shown in the figure does not constitute a limitation of the smartphone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0231] The memory 620 can be used to store software programs and modules. The processor 680 executes various functional applications and data processing of the smartphone by running the software programs and modules stored in the memory 620. The memory 620 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, a phone book, etc.), etc. In addition, the memory 620 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0232] The processor 680 is the control center of the smartphone, which uses various interfaces and lines to connect various parts of the entire smartphone, and executes various functions of the smartphone and processes data by running or executing software programs and / or modules stored in the memory 620, and calling data stored in the memory 620. Optionally, the processor 680 may include one or more processing units; preferably, the processor 680 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 680.

[0233] In this embodiment, the steps performed by the processor 680 in the smartphone may be based on Figure 6 The structure shown is implemented.

[0234] The computer device provided in the embodiment of the present application may also be a server. Figure 7 As shown, Figure 7 The structural diagram of the server 700 provided in the embodiment of the present application, the server 700 may have relatively large differences due to different configurations or performances, and may include one or more processors, such as a central processing unit (CPU) 722, and a memory 732, one or more storage media 730 (such as one or more mass storage devices) storing application programs 742 or data 744. Among them, the memory 732 and the storage medium 730 can be temporary storage or permanent storage. The program stored in the storage medium 730 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the server. Furthermore, the central processing unit 722 can be configured to communicate with the storage medium 730 and execute a series of instruction operations in the storage medium 730 on the server 700.

[0235] The server 700 may also include one or more power supplies 726, one or more wired or wireless network interfaces 750, one or more input and output interfaces 758, and / or one or more operating systems 741, such as Windows Server 2000. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM etc.

[0236] In this embodiment, the central processor 722 in the server 700 may perform the following steps:

[0237] Obtain M label pairs corresponding to the target content, wherein the M label pairs are composed of two different candidate labels among N candidate labels corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1;

[0238] For each of the M label pairs, constructing input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively;

[0239] Predicting the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtaining a prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content;

[0240] According to the prediction results respectively corresponding to the M label pairs, a content label corresponding to the target content is determined from the N candidate labels, and the content label is used to recommend the target content.

[0241] According to one aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium is used to store a computer program. When the computer program is executed by a computer device, the computer device executes the content label determination method described in the aforementioned embodiments.

[0242] According to one aspect of the present application, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in various optional implementations of the above-mentioned embodiments.

[0243] The descriptions of the processes or structures corresponding to the above-mentioned figures have different emphases. For parts that are not described in detail in a certain process or structure, please refer to the relevant descriptions of other processes or structures.

[0244] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0245] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0246] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0247] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0248] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the relevant technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store program codes.

[0249] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0250] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, ordinary technical members in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for determining a content label, characterized in that: The method comprises: Obtain M label pairs corresponding to the target content, wherein the M label pairs are composed of two different candidate labels among N candidate labels corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1; For each of the M label pairs, constructing input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively; Predicting the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtaining a prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content; According to the prediction results respectively corresponding to the M label pairs, a content label corresponding to the target content is determined from the N candidate labels, and the content label is used to recommend the target content.

2. The method according to claim 1, characterized in that If the two candidate tags included in the tag pair are the first candidate tag and the second candidate tag, constructing input data corresponding to the tag pair according to the tag pair and the target content includes: Acquire first tag description information corresponding to the first candidate tag, and acquire second tag description information corresponding to the second candidate tag; Input data corresponding to the tag pair is constructed according to the target content, the tag pair, the first tag description information, and the second tag description information.

3. The method according to claim 1, characterized in that The step of constructing input data corresponding to the tag pair according to the tag pair and the target content includes: Extracting a summary of the target content to obtain a content summary corresponding to the target content; According to the content summary corresponding to the target content and the label pair, input data corresponding to the label pair is constructed.

4. The method according to claim 3, characterized in that If the target content is multimodal content belonging to multiple modes, extracting a summary of the target content to obtain a content summary corresponding to the target content includes: Extracting summaries of the contents belonging to each mode in the multimodal content respectively, to obtain content summaries corresponding to the contents of the multiple modes respectively; The content summary corresponding to the target content is determined according to the content summaries respectively corresponding to the contents of the multiple modes.

5. The method according to claim 1, characterized in that If the target content is content to be recommended published by the target object on the content platform, constructing input data corresponding to the tag pair according to the tag pair and the target content includes: Acquire a content identifier corresponding to the target content, where the content identifier is input by the target object through the content platform for the target content; Constructing input data corresponding to the tag pair according to the target content, the content identifier, and the tag pair; The method further comprises: The target content is recommended in the content platform according to the content tag and the content identifier.

6. The method according to claim 1, characterized in that If N is an integer greater than 2, and M is an integer greater than 2, determining the content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs respectively includes: For each candidate tag among the N candidate tags, determine the correlation score between the candidate tag and the target content according to the prediction results corresponding to the R target tag pairs associated with the candidate tag, where the R target tag pairs refer to the tag pairs including the candidate tag among the M tag pairs, and R is an integer greater than 1 and less than M; According to the relevance scores respectively corresponding to the N candidate tags, a content tag corresponding to the target content is determined from the N candidate tags.

7. The method according to claim 6, characterized in that If the number of the content tags is K, where K is an integer greater than or equal to 1 and less than N, determining the content tag corresponding to the target content from the N candidate tags according to the relevance scores respectively corresponding to the N candidate tags includes: According to the relevance scores respectively corresponding to K candidate tags among the N candidate tags, a binary tree including K nodes is constructed, wherein the binary tree has the following characteristics: the K nodes are respectively used to indicate the relevance scores corresponding to the K candidate tags, different nodes are used to indicate the relevance scores corresponding to different candidate tags among the K candidate tags, the relevance score indicated by the parent node is less than or equal to the relevance score indicated by the child node corresponding to the parent node, the root node is used to indicate the minimum value among the K relevance scores, the parent node refers to a node having child nodes, and the root node is the parent node located at the top level of the binary tree; For the W candidate tags other than the K candidate tags among the N candidate tags, traverse the W candidate tags in sequence, and when traversing to the r-th candidate tag, compare the relevance score corresponding to the r-th candidate tag with the relevance score indicated by the root node to obtain a comparison result, where r is an integer greater than or equal to 1 and less than or equal to W; If the comparison result indicates that the relevance score of the r-th candidate tag is greater than the relevance score indicated by the root node, the relevance score indicated by the root node is updated to the relevance score of the r-th candidate tag to obtain a pending tree; Updating the pending tree until the pending tree satisfies the characteristics of the binary tree, and determining the pending tree as the binary tree; If the comparison result indicates that the relevance score of the r-th candidate tag is less than or equal to the relevance score indicated by the root node, the r-th candidate tag is eliminated; The K candidate tags corresponding to the K correlation scores stored in the target binary tree are determined as the K content tags corresponding to the target content, and the target binary tree refers to the binary tree obtained when the Wth candidate tag is traversed.

8. The method according to claim 1, characterized in that The step of constructing input data corresponding to the tag pair according to the tag pair and the target content includes: According to the label pair and the target content, construct first input data corresponding to the label pair and second input data corresponding to the label pair, wherein in the first input data, two candidate labels included in the label pair are arranged in a first order, and in the second input data, the two candidate labels included in the label pair are arranged in a second order, and the second order is different from the first order; The step of predicting the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair to obtain the prediction result corresponding to the tag pair includes: According to the first input data corresponding to the label pair, the correlation between the two candidate labels included in the label pair and the target content is predicted by a large model to obtain a first prediction result corresponding to the label pair; According to the second input data corresponding to the label pair, the correlation between the two candidate labels included in the label pair and the target content is predicted by the large model to obtain a second prediction result corresponding to the label pair; Determining a target prediction result corresponding to the label pair according to the first prediction result and the second prediction result; The determining, according to the prediction results respectively corresponding to the M label pairs, a content label corresponding to the target content from the N candidate labels comprises: According to the target prediction results respectively corresponding to the M label pairs, a content label corresponding to the target content is determined from the N candidate labels.

9. The method according to claim 8, characterized in that The determining, according to the first prediction result and the second prediction result, a target prediction result corresponding to the label pair includes: If the first prediction result is the same as the second prediction result, determining the first prediction result or the second prediction result as the target prediction result corresponding to the label pair; If the first prediction result is different from the second prediction result, it is determined that the target prediction result corresponding to the label pair indicates that the first degree of correlation is the same as the second degree of correlation, the first degree of correlation is used to indicate the correlation between the first candidate label included in the label pair and the target content, and the second degree of correlation is used to indicate the correlation between the second candidate label included in the label pair and the target content.

10. The method according to claim 1, characterized in that The method further comprises: Determining a target object group corresponding to the target content from the multiple object groups according to the content label corresponding to the target content and the group characteristics corresponding to the multiple object groups; Determine a target recommendation channel from the plurality of recommendation channels according to the group characteristics of the target object group and the channel object characteristics corresponding to the plurality of recommendation channels, wherein the channel object characteristics corresponding to the target recommendation channel and the group characteristics of the target object group satisfy an object characteristic similarity condition; The target content is recommended through the target recommendation channel.

11. The method according to any one of claims 1 to 10, characterized in that The N candidate tags are determined in the following manner: According to the target content, determining content features corresponding to the target content by a feature encoder; According to the content feature corresponding to the target content and a plurality of reference content features stored in a search library, determining a target reference content feature that satisfies a content feature similarity condition with the content feature of the target content from the plurality of reference content features, wherein the reference content feature is a content feature corresponding to the reference content determined by the feature encoder; The N candidate tags are determined based on the tags of the target reference content features stored in the search library.

12. A device for determining a content label, characterized in that: The device comprises an acquisition unit, a construction unit, a prediction unit and a determination unit: The acquisition unit is used to acquire M label pairs corresponding to the target content, wherein the M label pairs are composed of two different candidate labels among the N candidate labels corresponding to the target content, where N is an integer greater than 1 and M is an integer greater than or equal to 1; The construction unit is used to construct, for each label pair of the M label pairs, input data corresponding to the label pair according to the label pair and the target content, so as to obtain input data corresponding to the M label pairs respectively; The prediction unit is used to predict the correlation between the two candidate tags included in the tag pair and the target content according to the input data corresponding to the tag pair, and obtain the prediction result corresponding to the tag pair, wherein the prediction result is used to indicate the magnitude relationship of the correlation between the two candidate tags included in the tag pair and the target content; The determining unit is used to determine a content label corresponding to the target content from the N candidate labels according to the prediction results corresponding to the M label pairs respectively, and the content label is used to recommend the target content.

13. A computer device, characterized in that: The computer device comprises a processor and a memory: The memory is used to store a computer program and transmit the computer program to the processor; The processor is configured to execute the method according to any one of claims 1 to 11 according to the instructions in the computer program.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and when the computer program is executed by a computer device, the computer device executes the method according to any one of claims 1 to 11.

15. A computer program product comprising a computer program, characterized in that When the method is executed on a computer device, the computer device is enabled to execute the method according to any one of claims 1 to 11.

Images