Recommended resource processing method and device, equipment and storage medium

Abstract

The invention provides a resource processing model training method, a recommended resource processing method and device, equipment and a storage medium, and relates to the technical field of data processing, in particular to the technical fields of artificial intelligence, deep learning, big data, big models and the like. According to the specific implementation scheme, a target sorting result of N pieces of target sample data is obtained by utilizing a large language model; wherein the target sorting result can represent the similarity between the target sample data and the target resource; based on the target sorting result, obtaining M positive and negative sample pairs corresponding to the target resource; m and N are integers greater than 1; and performing similarity comparison learning training on the initial resource processing model by using the target resource and the M positive and negative sample pairs corresponding to the target resource to obtain a target resource processing model.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to technical fields such as artificial intelligence, deep learning, big data and big models. Background Art

[0002] In the current recommendation field, although there have been a variety of recommendation strategies with explicit features or implicit features of various granularities, the recommendation results are still somewhat different from human sensory perception, which affects the user experience. Summary of the invention

[0003] The present invention provides a training method for a resource processing model, a recommended resource processing method, and a device, equipment, and storage medium thereof.

[0004] According to one aspect of the present disclosure, a method for training a resource processing model is provided, comprising:

[0005] Using the large language model, a target ranking result of N target sample data is obtained; wherein the target ranking result can represent the similarity between the target sample data and the target resource;

[0006] Based on the target sorting results, M positive and negative sample pairs corresponding to the target resources are obtained; M and N are both integers greater than 1;

[0007] The target resource and M positive and negative sample pairs corresponding to the target resource are used to perform similarity comparison learning training on the initial resource processing model to obtain the target resource processing model.

[0008] According to another aspect of the present disclosure, a method for processing recommended resources is provided, comprising:

[0009] Using the target resource processing model, a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historical recommended resources are obtained, wherein the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using the target sorting result; the target sorting result is obtained by sorting N target sample data based on the similarity between the target sample data and the target resource using the large language model; N is an integer greater than 1;

[0010] Based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each historically recommended resource, obtaining a first similarity between the resource to be processed and each historically recommended resource;

[0011] Based at least on the first similarity, the recommendation weight of the resource to be processed is adjusted.

[0012] According to another aspect of the present disclosure, there is provided a training device for a resource processing model, comprising:

[0013] A sorting unit, used to obtain a target sorting result of N target sample data using a large language model; wherein the target sorting result can represent the similarity between the target sample data and the target resource;

[0014] A sample determination unit, used to obtain M positive and negative sample pairs corresponding to the target resource based on the target sorting result; M and N are both integers greater than 1;

[0015] The training unit is used to perform similarity comparison learning training on the initial resource processing model using the target resource and M positive and negative sample pairs corresponding to the target resource to obtain the target resource processing model.

[0016] According to another aspect of the present disclosure, there is provided a recommended resource processing device, comprising:

[0017] The feature determination unit is used to obtain a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historical recommended resources using a target resource processing model, wherein the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using a target sorting result; the target sorting result is obtained by sorting N target sample data using a large language model and based on the similarity between the target sample data and the target resource; N is an integer greater than 1; based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each of the historical recommended resources, a first similarity between the resource to be processed and each of the historical recommended resources is obtained;

[0018] The recommendation adjustment unit is used to adjust the recommendation weight of the resource to be processed based on at least the first similarity.

[0019] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0020] at least one processor; and

[0021] a memory communicatively connected to the at least one processor; wherein,

[0022] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any method in the embodiments of the present disclosure.

[0023] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute any method according to the embodiments of the present disclosure.

[0024] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements any method according to the embodiments of the present disclosure when executed by a processor.

[0025] In this way, the disclosed solution first uses the reasoning ability of the large language model to obtain the target ranking results of N target sample data, and then performs similarity comparison learning training on the initial resource processing model based on the target ranking results. In this way, the knowledge of the large language model is transferred to the target resource processing model, that is, knowledge distillation is completed, which lays the foundation for subsequent accurate recommendations and improved user experience.

[0026] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0028] Figure 1 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 1 ;

[0029] Figure 2 This is a comparative learning diagram of the initial resource processing model according to an embodiment of the present application. Figure 1 ;

[0030] Figure 3 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 2 ;

[0031] Figure 4 is a schematic diagram of the structure of an initial resource processing model according to an embodiment of the present application;

[0032] Figure 5 This is a comparative learning diagram of the initial resource processing model according to an embodiment of the present application. Figure 2 ;

[0033] Figure 6 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 3 ;

[0034] FIG. 7( a ) is a schematic diagram of grouping results after target sample data are grouped according to an embodiment of the present application;

[0035] FIG. 7( b ) is a schematic diagram of a data group constructed based on target resources and target sample data according to an embodiment of the present application;

[0036] FIG8( a ) is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 4 ;

[0037] FIG8( b ) is a schematic diagram illustrating sliding to determine multiple windows according to an embodiment of the present application;

[0038] FIG8( c ) is a schematic diagram illustrating a target ranking result obtained based on a sliding window method according to an embodiment of the present application;

[0039] Fig. 9 is a flowchart of a training method for a resource processing model according to an embodiment of the present application in a specific example;

[0040] Fig.10 is a schematic flow chart of a resource processing method recommended according to an embodiment of the present application;

[0041] Fig.11 is a structural diagram of a training device for a resource processing model according to an embodiment of the present application;

[0042] Fig.12 is a structural diagram of a recommended resource processing device according to an embodiment of the present application;

[0043] Fig.13 It is a block diagram of an electronic device used to implement the training method of the resource processing model or the recommended resource processing method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0044] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0045] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The term "at least one" in this article means any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C. The terms "first" and "second" in this article refer to multiple similar technical terms and distinguish them. They do not mean to limit the order or to limit them to only two. For example, the first feature and the second feature refer to two types / two features. The first feature can be one or more, and the second feature can also be one or more.

[0046] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the specific embodiments below. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the main purpose of the present disclosure.

[0047] In order to further improve the implicit expression ability of diversity, the present disclosure provides a training method for a resource processing model. With the help of the world knowledge and reasoning and induction ability of a large language model (LLM), samples of resources that are perceived to be similar by users are annotated. For example, the large language model is used to sort N target sample data, and then the sorting results are used to fine-tune the small model (corresponding to the initial resource processing model). In this way, the knowledge of the large language model specializing in the recommendation field is transferred to the initial resource processing model, that is, knowledge distillation is realized, thereby effectively enhancing the adaptation ability of the small model to the recommendation vertical domain, so that the recommended resources can be more in line with the user's real interests, thereby improving the user experience.

[0048] Specifically, Figure 1 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 1 The method may be optionally applied to electronic devices, such as personal computers, servers, server clusters, and other electronic devices.

[0049] Further, the method includes at least part of the following contents. Figure 1 As shown, including:

[0050] Step S101: using a large language model, obtaining target ranking results of N (N is an integer greater than 1) target sample data.

[0051] Here, the target ranking result can characterize the similarity between the target sample data and the target resource.

[0052] Furthermore, in one example, the target ranking result is obtained based on the similarity between the target resource and each target sample data. For example, the similarity between the target resource and each target sample data is first obtained, and then the N similarities are arranged in descending order to obtain the target ranking result.

[0053] Step S102: Based on the target sorting result, M (M is an integer greater than 1) positive and negative sample pairs corresponding to the target resource are obtained.

[0054] Step S103: using the target resource and the M positive and negative sample pairs corresponding to the target resource, similarity comparison learning training is performed on the initial resource processing model to obtain the target resource processing model.

[0055] In this way, the disclosed solution first uses the reasoning ability of the large language model to obtain the target ranking results of N target sample data, and then performs similarity comparison learning training on the initial resource processing model based on the target ranking results. In this way, the knowledge of the large language model is transferred to the target resource processing model, that is, knowledge distillation is completed, which lays the foundation for subsequent accurate recommendations and improved user experience.

[0056] In one example, the amount of adjustable parameters in the initial resource processing model is smaller than the amount of adjustable parameters in the large language model. Furthermore, in a specific example, the amount of adjustable parameters in the initial resource processing model is smaller than the amount of adjustable parameters in the large language model, such as the order of magnitude of the former is smaller than the latter. In this way, in the scenario of using the target resource processing model for resource recommendation, on the one hand, because the knowledge of the large language model is transferred to the target resource processing model through comparative learning, the accuracy of the recommendation results is effectively ensured.

[0057] On the other hand, since the number of parameters of the target resource processing model is small, the reasoning speed is effectively ensured. At the same time, it is easier to deploy in resource-constrained environments, thus reducing dependence on computing resources.

[0058] In a specific example, the target resource in the above example may be specifically a historical browsing resource of the target object, or may be any sample data among N target sample data, etc., and the disclosed solution does not impose any specific limitation on this.

[0059] Furthermore, in a specific example, the target resource may be specifically graphic data, video data, etc.; similarly, the target sample data may be specifically graphic data, video data, etc., and the present disclosure does not limit the specific forms of the target resource and the target sample data.

[0060] Furthermore, in a specific example, the method for obtaining the N target sample data may be obtained in the following manner, and the specific steps include:

[0061] Aggregate according to the granularity of the target object to obtain multiple initial similar samples. For example, first obtain a recommendation list of multiple target objects, which contains multiple resources. Then, based on the recommendation list of multiple target objects, filter out similar resources, and use the resources whose number of similar resources is greater than a threshold as initial similar samples.

[0062] A representation vector (eg, a semantic representation vector or a user collaborative collinear vector) is obtained for each initial similar sample, and based on the representation vector of each initial similar sample, the similarity between the initial similar samples is obtained; based on the similarity between the initial similar samples, N target sample data are screened and obtained.

[0063] It should be noted that the above is only an example. In actual applications, other methods can be used to obtain N target sample data, and the present disclosure does not limit this.

[0064] Further, the initial resource processing model may be trained by similarity comparison learning in the following manner; specifically, the above-mentioned similarity comparison learning training of the initial resource processing model using the target resource and the M positive and negative sample pairs corresponding to the target resource (for example, the above-mentioned step S103) may specifically include:

[0065] Fine-tune the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and / or to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair.

[0066] For example, in one example, the adjustable parameters of the initial resource processing model are fine-tuned to increase the similarity between the target resource and the positive sample in the positive-negative sample pair; or, in another example, the adjustable parameters of the initial resource processing model are fine-tuned to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair; or, in yet another example, the adjustable parameters of the initial resource processing model are fine-tuned to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair.

[0067] For example, if Figure 2 As shown, the target resource and the positive samples and negative samples corresponding to the target resource can be input into the initial resource processing model, and by comparing the loss values, the adjustable parameters of the initial resource processing model can be fine-tuned to increase the similarity between the target resource and its positive samples, and / or reduce the similarity between the target resource and its negative samples.

[0068] In this way, the disclosed scheme provides a refined scheme for similarity contrast learning training. Since the scheme adopts the contrast learning training method, the training effect can be achieved without a large amount of labeled data, and the training process is more efficient. At the same time, the model obtained after training can better capture the essential characteristics of the data. Furthermore, the optimization target in the training process of the disclosed scheme is clearer and can more intuitively reflect the changes in sample similarity, making the model training process and results easier to understand, and therefore more interpretable. In addition, since the data used for the training model is obtained based on the target ranking results obtained by the large language model, in the contrast learning training process, the capabilities of the large language model can be efficiently transferred to the resource processing model to achieve knowledge distillation, thereby making the resources recommended by the target resource processing model closer to the user's real perception (that is, better matching the user's real interests), thereby laying the foundation for improving user experience.

[0069] It should be pointed out that, for one training in similarity comparison learning training, the number of negative samples (or positive samples) input to the initial resource processing model may be one or more, and the present disclosure does not impose any specific limitation on this.

[0070] Figure 3 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 2 The method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, and other electronic devices. It can be understood that the above Figure 1 and Figure 2 The relevant contents of the method shown can also be applied to this example, and the relevant contents will not be described in detail in this example.

[0071] Further, the method includes at least part of the following contents. Figure 3 As shown, including:

[0072] Step S301: using a large language model, obtaining target ranking results of N (N is an integer greater than 1) target sample data.

[0073] Here, the target ranking result can characterize the similarity between the target sample data and the target resource.

[0074] Step S302: Based on the target sorting result, M (M is an integer greater than 1) positive and negative sample pairs corresponding to the target resource are obtained.

[0075] Step S303: using the initial resource processing model, obtain a first implicit vector corresponding to the target resource, a second implicit vector corresponding to the positive sample in the positive-negative sample pair, and a third implicit vector corresponding to the negative sample in the positive-negative sample pair.

[0076] It should be noted that the relevant contents about the target ranking results, target resources, and initial resource processing model can be referred to the above examples and will not be repeated here.

[0077] Furthermore, in one example, the initial resource processing model is used to encode the input sample and output an implicit vector corresponding to the input sample. In other words, the initial resource processing model can characterize the implicit features of the input content.

[0078] Here, it should be noted that since the disclosed solution can transfer the world knowledge of the large language model to the initial resource processing model, the implicit features expressed by the target resource processing model obtained after the transfer are more accurate, which makes it easier to mine potential recommended resources based on the implicit features in the recommendation scenario, and thus makes the resources recommended by the target resource processing model closer to the user's real interests, thereby effectively improving the user experience.

[0079] For example, Figure 4 A structural diagram of the initial resource processing model is provided, such as Figure 4 As shown, the initial resource processing model may include an embedding layer (which can be recorded as an Embedding Layer) and a network layer; using the embedding layer in the initial resource processing model, the input sample can be mapped to a fixed embedding space, and then the implicit vector corresponding to the input sample can be obtained after processing the features in the embedding space using the network layer in the initial resource processing model.

[0080] Step S304: fine-tune the adjustable parameters of the initial resource processing model to reduce the distance between the first implicit vector and the second implicit vector, thereby improving the similarity between the target resource and the positive sample, and / or to increase the distance between the first implicit vector and the third implicit vector, thereby reducing the similarity between the target resource and the negative sample.

[0081] That is to say, in one example, fine-tuning the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive and negative sample pairs can specifically include: fine-tuning the adjustable parameters of the initial resource processing model to reduce the distance between the first implicit vector and the second implicit vector, thereby improving the similarity between the target resource and the positive sample.

[0082] Alternatively, in another example, fine-tuning the adjustable parameters of the initial resource processing model to reduce the similarity between the target resource and the negative sample in the positive and negative sample pairs can specifically include: fine-tuning the adjustable parameters of the initial resource processing model to increase the distance between the first implicit vector and the third implicit vector, thereby reducing the similarity between the target resource and the negative sample.

[0083] In this way, the disclosed scheme further provides a refined scheme for similarity contrast learning training, namely, optimizing the adjustable parameters of the initial resource processing model by utilizing the distance between the first implicit vector of the target resource and the second implicit vector corresponding to the positive sample, and / or utilizing the distance between the first implicit vector of the target resource and the third implicit vector corresponding to the negative sample. In this way, the implicit features expressed by the trained target resource processing model are more accurate. In other words, they can better express the essential characteristics of the data, thereby facilitating accurate similarity judgments from a large number of resources based on the obtained implicit features, laying the foundation for the subsequent recommended resources to match the user's real interests and improve the user experience.

[0084] It should be noted that since the disclosed solution can use distillation to transfer the knowledge of a large language model with a large number of parameters (which can be understood as a teacher model) to a target resource processing model with a smaller number of parameters (also understood as a student model), in subsequent recommendation scenarios, the implicit features extracted by the target resource processing model are more accurate. At the same time, it can also ensure that the recommended resources obtained based on the implicit features are close to the recommended resources obtained using the large language model. In this way, the user experience is effectively improved. At the same time, compared with using a large language model for recommendation, the reasoning speed of the target resource processing model is also effectively improved.

[0085] Further, in a specific example, the first implicit vector, the second implicit vector, and the third implicit vector may be obtained in the following manner; specifically, the above-described method of using the initial resource processing model to obtain the first implicit vector corresponding to the target resource, the second implicit vector corresponding to the positive sample in the positive-negative sample pair, and the third implicit vector corresponding to the negative sample in the positive-negative sample pair (for example, step S303) may specifically include:

[0086] Using the first network in the initial resource processing model, a first implicit vector corresponding to the target resource is obtained;

[0087] Using the second network in the initial resource processing model, a second implicit vector corresponding to the positive sample in the positive and negative sample pairs is obtained;

[0088] The third network in the initial resource processing model is used to obtain a third implicit vector corresponding to the negative sample in the positive and negative sample pairs.

[0089] Here, the values ​​of the weight parameters in the first network, the weight parameters in the second network, and the weight parameters in the third network are associated. For example, in one example, the values ​​of the weight parameters in the three networks are the same.

[0090] For example, in one example, the twin network can be networked in a paired manner (also called a pairwise manner), so that by processing paired input data, the similarity between the input data can be effectively learned.

[0091] For example, Figure 5 As shown, the initial resource processing model includes network 1, network 2 and network 3, wherein network 1, network 2 and network 3 share the same weight parameter (for example, represented by "w"). At this time, using network 1 in the initial resource processing model, the implicit vector 1 of the target resource can be obtained, using network 2 in the initial resource processing model, the implicit vector 2 corresponding to the positive sample can be obtained, and using network 3 in the initial resource processing model, the implicit vector 3 corresponding to the negative sample can be obtained. Furthermore, based on the implicit vectors obtained above, the contrast loss value can be obtained, and then based on the contrast loss value, the adjustable parameters of the initial resource processing model, such as weight parameters, can be fine-tuned.

[0092] It should be pointed out that Figure 5 The initial resource processing model shown also includes an embedding layer. For example, an embedding layer is set before each network. At this time, the embedding layer 1 in the initial resource processing model is used to map the target resource to a fixed embedding space, and then the network 1 in the initial resource processing model is used to process the features in the embedding space, so as to obtain the implicit vector 1 corresponding to the target resource; similarly, the embedding layer 2 in the initial resource processing model is used to map the positive sample to a fixed embedding space, and then the network 2 in the initial resource processing model is used to process the features in the embedding space, so as to obtain the implicit vector 2 corresponding to the positive sample; the embedding layer 3 in the initial resource processing model is used to map the negative sample to a fixed embedding space, and then the network 3 in the initial resource processing model is used to process the features in the embedding space, so as to obtain the implicit vector 3 corresponding to the negative sample.

[0093] Furthermore, during the model training phase, the network parameters of the above-mentioned embedding layers can be further fine-tuned.

[0094] It should be noted that the network structure of the above initial resource processing model is only an exemplary description. In actual applications, other methods can be used to construct the initial resource processing model, and the present disclosure does not impose any specific restrictions on this.

[0095] In this way, the disclosed scheme provides a specific scheme for obtaining implicit vectors. The scheme can utilize the twin network in the initial resource processing model to obtain implicit vectors (i.e., implicit features) of the target resources, positive samples, and negative samples. For example, the implicit vectors corresponding to the target resources are extracted by the first network, the implicit vectors corresponding to the positive samples are extracted by the second network, and the implicit vectors corresponding to the negative samples are extracted by the third network. Therefore, while ensuring the accurate extraction of features, the complexity and operational difficulty of implicit vector extraction are effectively reduced, and the data processing logic is more reasonable and efficient, thereby providing strong support for subsequent efficient comparative learning training and the realization of knowledge distillation.

[0096] Further, in a specific example, the initial resource processing model may be optimized by the following optimization method to reduce the distance between the first implicit vector and the second implicit vector, or to increase the distance between the first implicit vector and the third implicit vector, specifically including:

[0097] Mode 1: The above-mentioned fine-tuning of the adjustable parameters of the initial resource processing model to reduce the distance between the first implicit vector and the second implicit vector may specifically include:

[0098] Performing an inner product operation on the first implicit vector and the second implicit vector to obtain a first inner product result;

[0099] Fine-tune the adjustable parameters of the initial resource processing model to reduce the difference between the first inner product result and the first threshold. Here, the first threshold is a theoretical similarity extreme value. For example, in one example, the first threshold can be specifically set to 1.

[0100] It can be understood that the first inner product result in this example can be used to characterize the similarity between the target resource and the positive sample in the positive and negative sample pair. Accordingly, the smaller the difference between the first inner product result and the first threshold value (for example, the closer the first inner product result is to the first threshold value), the greater the similarity between the target resource and the positive sample in the positive and negative sample pair.

[0101] Mode 2: namely, fine-tuning the adjustable parameters of the initial resource processing model as described above to increase the distance between the first implicit vector and the third implicit vector, which may specifically include:

[0102] Performing an inner product operation on the first implicit vector and the third implicit vector to obtain a second inner product result;

[0103] Fine-tune the adjustable parameters of the initial resource processing model to reduce the difference between the second inner product result and the second threshold. Here, the second threshold is a theoretical difference extreme value. For example, in one example, the second threshold can be specifically set to 0.

[0104] It can be understood that the second inner product result in this example can be used to characterize the similarity between the target resource and the negative sample in the positive-negative sample pair. Accordingly, the smaller the difference between the second inner product result and the second threshold value (for example, the closer the second inner product result is to the second threshold value), the smaller the similarity between the target resource and the negative sample in the positive-negative sample pair.

[0105] Mode 3: namely, fine-tuning the adjustable parameters of the initial resource processing model as described above to reduce the distance between the first implicit vector and the second implicit vector, and to increase the distance between the first implicit vector and the third implicit vector, may specifically include:

[0106] Performing an inner product operation on the first implicit vector and the second implicit vector to obtain a first inner product result;

[0107] Performing an inner product operation on the first implicit vector and the third implicit vector to obtain a second inner product result;

[0108] The adjustable parameters of the initial resource processing model are fine-tuned to reduce the difference between the first inner product result and the first threshold value, and to reduce the difference between the second inner product result and the second threshold value.

[0109] It can be understood that, in practical applications, the target loss function used to train the initial resource processing model may include a first part for measuring the similarity between the first implicit vector and the second implicit vector, and / or a second part for measuring the similarity between the first implicit vector and the third implicit vector.

[0110] In this way, the disclosed scheme provides a refined scheme for similarity contrast learning training, that is, the scheme provides a refined scheme for reducing the distance between the first implicit vector and the second implicit vector, in which the first implicit vector and the second implicit vector are firstly inner-producted, and then the difference between the result of the inner-product operation (that is, the first inner-product result) and the theoretical similarity extreme value is reduced; or, a refined scheme for increasing the distance between the first implicit vector and the third implicit vector is also provided, for example, the first implicit vector and the third implicit vector are first inner-producted, and then the difference between the result of the inner-product operation (that is, the second inner-product result) and the theoretical difference extreme value is reduced; the data processing logic of the above two refined schemes is more concise and clear, and at the same time, it is also easier to calculate, so it can further improve the optimization efficiency of similarity contrast learning training, and at the same time, make the entire training process easier to understand and explain, in other words, it is also highly interpretable.

[0111] In addition, the target resource processing model obtained after the above training can express the implicit characteristics of the input information more accurately. In other words, it can better capture the essential characteristics of the input information. Therefore, the resources recommended by the subsequent target resource processing model are closer to the user's real perception (that is, they can better match the user's real interests), thereby laying the foundation for improving the user experience.

[0112] Figure 6 This is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 3 The method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, and other electronic devices. It can be understood that the above Figures 1 to 5 The relevant contents of the method shown can also be applied to this example, and the relevant contents will not be described in detail in this example.

[0113] Further, the method includes at least part of the following contents. Figure 6 As shown, including:

[0114] Step S601: using a large language model, obtaining target ranking results of N (N is an integer greater than 1) target sample data.

[0115] Here, the target ranking result can characterize the similarity between the target sample data and the target resource.

[0116] Step S602: based on the target sorting result, group the N target sample data to obtain at least two groups.

[0117] Here, the similarity between the target sample data and the target resource included in the first group of the at least two groups is greater than the similarity between the target sample data and the target resource included in the other groups except the first group. In other words, the minimum similarity among the similarities between the target sample data and the target resource included in the first group is greater than the maximum similarity among the similarities between the target sample data and the target resource included in the other groups except the first group.

[0118] Step S603: Select positive samples from the target sample data included in the first group, and select negative samples from groups other than the first group, to obtain M positive and negative sample pairs.

[0119] Step S604: using the target resource and the M positive and negative sample pairs corresponding to the target resource, perform similarity comparison learning training on the initial resource processing model to obtain the target resource processing model.

[0120] It should be noted that the relevant contents about the target ranking results, target resources, initial resource processing model, and similarity comparison learning training can be referred to the above examples, which will not be repeated here.

[0121] In this way, the disclosed scheme provides a scheme for quickly constructing positive and negative sample pairs, namely, first grouping the target ranking results obtained based on the large language model, then selecting positive samples from the group with high similarity, and selecting negative samples from other groups, so as to quickly construct positive and negative sample pairs, and then using the target resources and the constructed positive and negative sample pairs to perform similarity comparison learning training to quickly perform knowledge distillation, so that the implicit features expressed by the target resource processing model obtained after knowledge distillation are more accurate, and then it is convenient to make accurate similarity judgments from a large number of resources based on the obtained implicit features, so that the subsequently recommended resources can better match the user's real interests, thereby improving the user experience.

[0122] Further, in a specific example, the N target sample data may be grouped in the following manner; specifically, the N target sample data may be grouped based on the target sorting result described above to obtain at least two groups (for example, step S602), which may specifically include:

[0123] Based on the target sorting result, the N target sample data are grouped into a head group, a middle group and a tail group.

[0124] Here, the similarity between the target sample data contained in the head group and the target resource is greater than the similarity between the target sample data contained in the middle group and the target resource; the similarity between the target sample data contained in the middle group and the target resource is greater than the similarity between the target sample data contained in the tail group and the target resource.

[0125] Here, it should be noted that the samples in the middle group can be understood as hard samples. The above grouping method is more conducive to the model's learning of hard samples, thereby providing strong support for improving the model training effect.

[0126] For example, the large language model is used to sort the 12 target sample data in descending order, and the target sorting result shown in Figure 7(a) is obtained, that is:

[0127] {I0,I2,I1,I8,I3,I6,I4,I7,I5,I9,I10,I11};

[0128] Furthermore, {I0, I2, I1, I8} among the 12 target sample data are taken as the head group, {I3, I6, I4, I7} among the 12 target sample data are taken as the middle group, and {I5, I9, I10, I11} among the 12 target sample data are taken as the tail group.

[0129] Further, the above-mentioned selecting positive samples from the target sample data included in the first group, and selecting negative samples from groups other than the first group to obtain M positive and negative sample pairs (for example, the above-mentioned step S603), may specifically include:

[0130] Positive samples are selected from the target sample data included in the head group, and negative samples are selected from the target sample data included in the middle group or the tail group to obtain M positive and negative sample pairs.

[0131] For example, continuing to take the head group, middle group and tail group shown in Figure 7(a) as an example, at this time, as shown in Figure 7(b), positive samples can be selected from I0, I2, I1, I8 contained in the head group to obtain the data group of <target resource, head> constructed by the target resource and the positive sample. And, negative samples can be selected from I3, I6, I4, I7 contained in the middle group to obtain the data group of <target resource, middle> constructed by the target resource and the negative sample, or negative samples can be selected from I5, I9, I10, I11 contained in the tail group to obtain the data group of <target resource, tail> constructed by the target resource and the negative sample. In this way, it is convenient to quickly perform similarity comparison learning and training based on the above-constructed data group.

[0132] Furthermore, in one example, in order to improve the model's learning ability for sample data, for example, to improve the learning ability for samples in the middle group (the samples in the middle group can be understood as difficult samples), a positive sample can be selected from the target sample data contained in the head group, and a negative sample can be selected from the target sample data contained in the middle group and the tail group, so as to construct a sample group of <target resource, head, middle, and tail>. In this way, similarity comparison learning training can be performed to enable the model to better learn the sample characteristics of the middle group.

[0133] In this way, the disclosed scheme provides a refined scheme for constructing positive and negative sample pairs, namely, based on the target ranking results obtained using a large language model, N target sample data are divided into a head group, a middle group and a tail group. In this way, a large number of positive and negative sample pairs can be quickly constructed, providing sufficient data support for subsequent similarity comparison learning and training.

[0134] FIG8( a ) is a schematic flow chart of a method for training a resource processing model according to an embodiment of the present application. Figure 4 The method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, and other electronic devices. It can be understood that the above Figure 1 The related contents of the method shown in FIG. 7 can also be applied to this example, and the related contents will not be described in detail in this example.

[0135] Further, the method includes at least part of the following contents. As shown in FIG8(a), it includes:

[0136] Step S801: using a large language model, obtaining the core content of each target sample data among N target sample data.

[0137] Here, the core content of the target sample data refers to the key information and key points extracted from the target sample data, such as the summary of the target sample data, etc. This provides strong support for the subsequent rapid sorting based on the similarity between the core contents of the target sample data.

[0138] In practical applications, if the target sample data is a graphic resource, optical character recognition (OCR) technology can be used for recognition first, and then the recognized content can be processed using a large language model to obtain a summary of the graphic resource; or, in another example, if the target sample data is a video resource, a large language model can be used directly for processing to obtain a summary of the video resource. It should be noted that the above processing examples are only exemplary. In practical applications, for different data forms, other methods can be used for preprocessing, and then a large language model can be used to extract a summary, etc. The disclosed solution does not impose specific restrictions on this.

[0139] Step S802: using the large language model and based on the core content of each target sample data, obtaining a target ranking result.

[0140] It should be noted that the disclosed solution first obtains the core content of each target sample data, and then uses the large language model to quickly obtain the similarity between the core contents of each target sample data, and then quickly obtains the target ranking result. In other words, the disclosed solution makes full use of the world knowledge and reasoning and induction capabilities of the large language model to obtain the target ranking result, thus providing strong support for knowledge distillation.

[0141] Step S803: Based on the target sorting result, M positive and negative sample pairs corresponding to the target resource are obtained.

[0142] Step S804: using the target resource and the M positive and negative sample pairs corresponding to the target resource, perform similarity comparison learning training on the initial resource processing model to obtain the target resource processing model.

[0143] In this way, the disclosed solution realizes the transfer of knowledge from the large language model to the target resource processing model, that is, completes knowledge distillation, thus laying the foundation for subsequent accurate recommendations and improving user experience.

[0144] Here, it should be noted that in actual applications, a large number of samples are required for model training. At this time, if all samples are input into the large language model at one time, the accuracy of the sorting results will be reduced. Based on this, in order to make the target sorting results more accurate and better reflect the world knowledge capabilities of the large language model, a sliding window method can be used to sort the target sample data.

[0145] Specifically, in a specific example, the above-mentioned use of the large language model and obtaining the target ranking result based on the core content of each target sample data (for example, step S802) may specifically include:

[0146] Step S802-1: adopt a sliding window method, and based on the initial sorting of N target sample data, slide to determine multiple windows.

[0147] Here, each window contains n target sample data. In other words, n represents a preset window length. Further, n is an integer less than or equal to N and greater than 1.

[0148] For example, in one example, the above-described method of using a sliding window and sliding to determine multiple windows based on the initial sorting of N target sample data (for example, step S802-1) may also specifically include:

[0149] Step a: Based on a preset window length, n consecutive target sample data are determined from an initial sort of N target sample data to determine a window.

[0150] Step b: Slide the window based on a preset step length and along a preset sliding direction to obtain the next window, and determine multiple windows by sliding.

[0151] For example, as shown in FIG8(b), the initial order of the 11 target sample data is {I0, I1, I2, I3, I4, I5, I6, I7, I8, I9, I10}, the preset window length n is 6, the preset step size is 3, and the preset sliding direction of the window (i.e., the gray area in the figure) is from the tail to the head of the initial sequence. Based on this, according to the preset window length, 6 consecutive target sample data (i.e., I5, I6, I7, I8, I9, I10) are determined from the initial order of the 11 target sample data. ) to determine the first window; further, according to the preset step size, slide the window toward the head by 3 units, at this time, obtain new continuous 6 target sample data (i.e., I2, I3, I4, I5, I6, I7) to determine the second window; further, when it is determined that the number of steps sliding toward the head (i.e., step number 2) is less than the preset step size, slide the window toward the head by 2 units to obtain continuous new 6 target sample data (i.e., I0, I1, I2, I3, I4, I5, I6) to determine the third window. So far, 3 windows for 11 target sample data are obtained.

[0152] Here, the above is only an example. The sliding direction of the window in this example can also be from the head to the tail of the sequence, etc., and the present disclosure does not impose any specific limitation on this.

[0153] In this way, the disclosed solution utilizes a sliding window approach to quickly determine multiple "windows" from N target sample data, thereby providing strong support for subsequently accurately obtaining the target sorting results of the N target sample data. Moreover, this method is simple and efficient, thereby laying the foundation for subsequent rapid distillation to obtain the target resource processing model.

[0154] Step S802-2: using the large language model and the core content of each target sample data, determine the sub-sorting result of each window.

[0155] For example, for a window, a prompt template (Prompt) can be constructed according to the core content of each target sample data framed by the window and the core content of the target resource, and the prompt template can be input into the large language model to use the reasoning ability of the large language model to determine the similarity between the core content of the target resource and the core content of the target sample data, and based on the similarity, score and sort the target sample data framed by the window, and then output the sub-sort result of the target sample data framed by the current window.

[0156] Step S802-3: Obtain a target sorting result based on the sub-sorting results of each window.

[0157] For example, continuing to take the method of determining the window in Figure 8(b) as an example, as shown in Figure 8(c), after the first window is determined, the large language model is used, and based on the core content of each target sample data in the six target sample data framed by the first window, the sub-sorting result of the six target sample data framed by one window is obtained; further, after obtaining the sub-sorting result corresponding to one window, the above method is used to slide, and the second window is determined, and then the large language model is used, and based on the core content of each target sample data in the six target sample data framed by the second window, the sub-sorting result of the six target sample data framed by the second window is obtained; similarly, the above sliding method is continued to be used to determine the third window, and then the large language model is used, and based on the core content of each target sample data in the six target sample data framed by the third window, the sub-sorting result of the six target sample data framed by the third window is determined. In this way, after determining that all sortings are completed, the target sorting result is obtained.

[0158] Specifically, as shown in FIG8(c), the step of obtaining the target ranking result may specifically include:

[0159] Step a: The n consecutive target sample data selected by the current window are used as the n window sample data of the current window.

[0160] For example, when sorting for the first time, n consecutive target sample data can be determined from the tail or head of the initial sorting to obtain n window sample data of the current window. For example, from the tail, consecutive I5, I6, I7, I8, I9, I10 are used as the 6 window sample data that need to be sorted for the first time.

[0161] Step b: using the large language model, based on the core content of each window sample data in the current window, determine the sub-sorting results of n window sample data in the current window.

[0162] Step c: based on the sub-sorting results of the n window sample data in the current window, update the sorting of the N target sample data to obtain a new current sorting of the N target sample data.

[0163] For example, after obtaining the sub-sorting results of the 6 window sample data that need to be sorted for the first time (i.e., {I8, I6, I7, I5, I9, I10}), the sorting of the 11 target sample data is updated to obtain the current sorting of the 11 target sample data, i.e., {I0, I1, I2, I3, I4, I8, I6, I7, I5, I9, I10}.

[0164] Step d: Determine whether the number of sliding steps of the current window is less than the preset step size. If so, proceed to step e; otherwise, proceed to step g.

[0165] Step e: Determine whether the number of sliding steps of the current window is greater than 0; if so, proceed to step f; otherwise, output the current sorting of the N target sample data (ie, the target sorting result).

[0166] Step f: Slide the window along the preset sliding direction according to the number of sliding steps to obtain a new current window, and then return to step a.

[0167] For example, when the number of sliding steps is 2, the window is slid 2 steps along the preset sliding direction according to the number of sliding steps to obtain a new current window. The 6 consecutive target sample data selected by the new current window are I0, I1, I2, I8, I3, and I6.

[0168] Step g: Slide the window based on a preset step size and along a preset sliding direction to obtain a new current window, and then return to step a.

[0169] For example, after completing the first sorting, it is determined that the number of sliding steps (such as 5) is greater than the preset step size. At this time, the window is slid according to the preset step size and along the preset sliding direction to obtain a new current window. The 6 consecutive target sample data selected by the new current window are I2, I3, I4, I8, I6, and I7, and return to step a.

[0170] In this way, the disclosed solution can utilize a sliding window approach and combine it with a large language model to obtain accurate target sorting results. This effectively improves the sorting efficiency while saving the computing resources required for sorting. At the same time, it also improves the accuracy of the sorting results, thereby providing data support for subsequent similarity comparison learning and training, and efficiently realizing knowledge distillation.

[0171] The following combination Fig. 9 The disclosed solution is further described in detail, specifically, Fig. 9 As shown in FIG. 1 , in the knowledge extraction stage, a sliding window method is adopted and a large language model is used to obtain a target ranking result of 12 target sample data (i.e., {I0, I2, I1, I8, I3, I6, I4, I7, I5, I9, I10, I11}); in the knowledge processing stage, according to the target ranking result, the 12 target sample data are divided, for example, into a head group, a middle group and a tail group, and the above-mentioned construction method is adopted to construct the following: Fig. 9Multiple data groups shown. Furthermore, in the knowledge transfer stage, the implicit vector 1 corresponding to the target resource is obtained by using network 1 in the initial resource processing model, the implicit vector 2 corresponding to the positive sample is obtained by using network 2 in the initial resource processing model, and the implicit vector 3 corresponding to the negative sample is obtained by using network 3 in the initial resource processing model; further, according to each implicit vector, a contrast loss value is obtained, and then according to the contrast loss value, the adjustable parameters of the initial resource processing model are fine-tuned to obtain the target resource processing model by distillation. In this way, the knowledge of the large language model is transferred to the target resource processing model, that is, the knowledge distillation is completed, which lays the foundation for subsequent accurate recommendations and improved user experience.

[0172] In this way, the disclosed scheme effectively utilizes the knowledge distillation method to obtain a target resource processing model that expresses implicit information more accurately. Specifically, the disclosed scheme first uses the world knowledge and reasoning and induction capabilities of the large language model to obtain the target ranking results, and then uses the target ranking results to perform similarity comparison learning training on the initial resource processing model, so that the initial resource processing model can better learn the order relationship, and then transfer the knowledge of the large language model specializing in the recommendation field to the initial resource processing model. In this way, knowledge distillation is achieved, thereby effectively enhancing the adaptability of the target resource processing model to the recommendation vertical domain, truly realizing the diversity modeling of users' real interests, making the recommended resources closer to users' real interests, thereby improving user experience.

[0173] The disclosed solution also provides a detailed solution for resource recommendation using the target resource processing model obtained above, specifically, Fig.10 As shown, the recommended resource processing method may specifically include:

[0174] Step S1001: using the target resource processing model, obtaining a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historically recommended resources.

[0175] Here, the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using the target ranking result.

[0176] Furthermore, the target ranking result is obtained by ranking N target sample data using a large language model based on the similarity between the target sample data and the target resource, where N is an integer greater than 1.

[0177] Furthermore, in one example, the target resource processing model is obtained by using the resource processing model training method described above.

[0178] Here, it should be noted that the multiple historical recommendation resources can be the latest multiple historical recommendation resources browsed by the target object. In other words, they can be the most recent multiple historical data. This makes it easy to quickly optimize the recommended resources based on the latest browsing content of the target object, thereby laying the foundation for further improving the user experience.

[0179] In addition, it should be noted that the specific number of historical recommended resources can be preset based on specific recommendation scenarios, recommendation requirements, etc., and the present disclosure does not impose specific restrictions on this.

[0180] Step S1002: based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each historically recommended resource, obtain a first similarity between the resource to be processed and each historically recommended resource.

[0181] Step S1003: adjusting the recommendation weight of the resource to be processed based at least on the first similarity.

[0182] In this way, the disclosed scheme can first use the target resource processing model obtained after knowledge distillation to obtain a more accurate implicit vector of implicit information, for example, to obtain the implicit vector of the resource to be processed and the implicit vector of the historically recommended resource, and then use the obtained implicit vector to perform similarity matching to obtain a more accurate similarity measurement value, and finally use the more accurately expressed similarity to optimize the recommendation weight of the resource to be processed. In this way, it provides quantifiable data support for the subsequent recommended resources to better match the user's real interests, thereby laying the foundation for further improving the user experience.

[0183] Further, in a specific example, the following method can be used to optimize the recommendation weight of the resource to be processed. Specifically, the above-mentioned adjustment of the recommendation weight of the resource to be processed based on at least each first similarity (for example, step S1003) includes:

[0184] Step S1003 - 1 : Determine a first number of historically recommended resources whose first similarity is greater than a similarity threshold.

[0185] Step S1003-2: Determine whether the first number is greater than or equal to the preset number; if so, that is, when the first number is greater than or equal to the preset number, execute step S1003-3; otherwise, that is, when the first number is less than the preset number, execute step S1003-4.

[0186] Step S1003-3: Increase the recommendation weight of the resource to be processed.

[0187] Here, it should be noted that when the first number is greater than or equal to the preset number, it can be understood that the current resource to be processed is highly similar to the historically recommended resource. At this time, the recommendation weight of the processed resource can be increased to achieve a moderate addition of recommended resources of short-term interest to the target object, thereby achieving a more personalized recommendation rhythm and further improving the user experience.

[0188] In addition, it should be noted that the degree of increase may be related to the first number, further, the degree of increase may be proportional to the first number, or the degree of increase may be related to the proportion of the first number in the total number of historically recommended resources. It is understandable that in actual applications, other ways of increasing the number may also be used, and the disclosed solution does not impose specific limitations on this.

[0189] Step S1003 - 4 : when the first number is less than the preset number, reduce the recommendation weight of the resource to be processed.

[0190] Here, it should be noted that when the first number is less than the preset number, it can be understood that the current resource to be processed has a weaker similarity with the historically recommended resource. At this time, the recommendation weight of the resource to be processed can be appropriately reduced. In this way, a more personalized recommendation rhythm is achieved. At the same time, it can also effectively avoid the recommended resources being far away from the user's real interests, thereby laying the foundation for improving user experience.

[0191] Further, in another example, when the first number is less than the preset number, it is also possible to further determine how to optimize the current resource to be processed based on the recommended candidate resources in the current recommendation list; specifically, when the first number is less than the preset number, reducing the recommendation weight of the resource to be processed (for example, the step S1003-4 described above) may specifically include:

[0192] Step S1003-4-1: when the first number is less than the preset number, using the target resource processing model, obtain a sixth implicit vector of each of the multiple recommended candidate resources.

[0193] Here, it should be noted that the multiple recommended candidate resources can be understood as the multiple resources to be recommended contained in the recommendation list, in other words, the resources that will be recommended to the target object in the future. Further, the above-mentioned pending resources can be understood as the resources that have not entered the recommendation list, in other words, the resources that need to be determined based on the recommendation weight whether to enter the recommendation list.

[0194] Step S1003-4-2: based on the fourth implicit vector of the resource to be processed and the sixth implicit vector of each recommended candidate resource, obtain a second similarity between the resource to be processed and each recommended candidate resource.

[0195] Step S1003-4-3: Determine a second number of recommended candidate resources whose second similarity is greater than a similarity threshold.

[0196] Step S1003-4-4: when the second number is greater than or equal to the preset number, reduce the recommendation weight of the resource to be processed.

[0197] That is to say, when the second number is greater than or equal to the preset number, it can be understood that the resource to be processed has a weaker similarity with the historical recommended resources, but a stronger similarity with the recommended candidate resources in the recommendation list. At this time, in order to avoid the information cocoon problem, the recommendation weight of the resource to be processed can be reduced. In this way, the recommended resources that have been insensitive to the target object for a long time are broken up, thereby achieving a more personalized recommendation rhythm, providing quantifiable data support for the subsequent recommended resources to better match the user's real interests, and thus laying the foundation for further improving the user experience.

[0198] In addition, it should be noted that the degree of reduction may be related to the second number, and further, the degree of reduction may be proportional to the second number; or the degree of reduction may be related to the proportion of the second number in the total number of recommended candidate resources. It is understandable that in actual applications, other reduction methods may be used, and the present disclosure does not impose specific restrictions on this.

[0199] The disclosed solution provides a training device for a resource processing model, such as Fig.11 As shown, including:

[0200] The sorting unit 1101 is used to obtain a target sorting result of N target sample data using a large language model; wherein the target sorting result can represent the similarity between the target sample data and the target resource;

[0201] The sample determination unit 1102 is used to obtain M positive and negative sample pairs corresponding to the target resource based on the target sorting result; M and N are both integers greater than 1;

[0202] The training unit 1103 is used to perform similarity comparison learning training on the initial resource processing model using the target resource and M positive and negative sample pairs corresponding to the target resource to obtain the target resource processing model.

[0203] In a specific example of the disclosed solution, the training unit is specifically used to:

[0204] Fine-tune the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and / or to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair.

[0205] In a specific example of the disclosed solution, the initial resource processing model is used to encode the input sample and output the implicit vector corresponding to the input sample;

[0206] Wherein, the training unit is specifically used for:

[0207] Using the initial resource processing model, obtain a first implicit vector corresponding to the target resource, a second implicit vector corresponding to the positive sample in the positive-negative sample pair, and a third implicit vector corresponding to the negative sample in the positive-negative sample pair;

[0208] The adjustable parameters of the initial resource processing model are fine-tuned to reduce the distance between the first implicit vector and the second implicit vector, and / or to increase the distance between the first implicit vector and the third implicit vector.

[0209] In a specific example of the disclosed solution, the training unit is specifically used for:

[0210] Performing an inner product operation on the first implicit vector and the second implicit vector to obtain a first inner product result;

[0211] Fine-tune the adjustable parameters of the initial resource processing model to reduce the difference between the first inner product result and the first threshold; wherein the first threshold is a theoretical similarity extreme value.

[0212] In a specific example of the disclosed solution, the training unit is specifically used for:

[0213] Performing an inner product operation on the first implicit vector and the third implicit vector to obtain a second inner product result;

[0214] The adjustable parameters of the initial resource processing model are fine-tuned to reduce the difference between the second inner product result and a second threshold, wherein the second threshold is a theoretical difference extreme value.

[0215] In a specific example of the disclosed solution, the training unit is specifically used to:

[0216] Using the first network in the initial resource processing model, a first implicit vector corresponding to the target resource is obtained;

[0217] Using the second network in the initial resource processing model, a second implicit vector corresponding to the positive sample in the positive and negative sample pairs is obtained;

[0218] Using the third network in the initial resource processing model, a third implicit vector corresponding to the negative sample in the positive and negative sample pairs is obtained;

[0219] Among them, the values ​​of the weight parameters in the first network, the weight parameters in the second network, and the weight parameters in the third network are associated.

[0220] In a specific example of the disclosed solution, the sample determination unit is specifically used to:

[0221] Based on the target sorting result, the N target sample data are grouped to obtain at least two groups; wherein the similarity between the target sample data included in the first group of the at least two groups and the target resource is greater than the similarity between the target sample data included in other groups except the first group and the target resource;

[0222] Positive samples are selected from the target sample data included in the first group, and negative samples are selected from groups other than the first group to obtain M positive and negative sample pairs.

[0223] In a specific example of the disclosed solution, the sample determination unit is specifically used to:

[0224] Based on the target sorting result, N target sample data are grouped to obtain a head group, a middle group and a tail group; wherein the similarity between the target sample data contained in the head group and the target resource is greater than the similarity between the target sample data contained in the middle group and the target resource; the similarity between the target sample data contained in the middle group and the target resource is greater than the similarity between the target sample data contained in the tail group and the target resource;

[0225] Positive samples are selected from the target sample data included in the head group, and negative samples are selected from the target sample data included in the middle group or the tail group to obtain M positive and negative sample pairs.

[0226] In a specific example of the disclosed solution, the sorting unit is specifically used to:

[0227] Using the large language model, the core content of each target sample data among the N target sample data is obtained;

[0228] Using a large language model and based on the core content of each target sample data, the target ranking results are obtained.

[0229] In a specific example of the disclosed solution, the sorting unit is specifically used to:

[0230] A sliding window method is adopted, and based on the initial sorting of N target sample data, multiple windows are slidingly determined; wherein each window contains n target sample data, and n represents a preset window length, which is an integer less than or equal to N and greater than 1;

[0231] Using the large language model and the core content of each target sample data, the sub-sorting results of each window are determined;

[0232] Based on the sub-sorting results of each window, the target sorting result is obtained.

[0233] In a specific example of the disclosed solution, the sorting unit is specifically used to:

[0234] Based on a preset window length, determining n consecutive target sample data from an initial sort of the N target sample data to determine a window;

[0235] The window is slid based on a preset step length and along a preset sliding direction to obtain a next window, and a plurality of windows are determined by sliding.

[0236] The disclosed solution also provides a recommended resource processing device, such as Fig.12 As shown, including:

[0237] The feature determination unit 1201 is used to obtain a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historical recommended resources using a target resource processing model, wherein the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using a target sorting result; the target sorting result is obtained by sorting N target sample data using a large language model and based on the similarity between the target sample data and the target resource; N is an integer greater than 1; based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each of the historical recommended resources, a first similarity between the resource to be processed and each of the historical recommended resources is obtained;

[0238] The recommendation adjustment unit 1202 is configured to adjust the recommendation weight of the resource to be processed based at least on the first similarity.

[0239] In a specific example of the disclosed solution, the recommendation adjustment unit is specifically used to:

[0240] Determine a first number of historically recommended resources having a first similarity greater than a similarity threshold;

[0241] When the first number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is increased.

[0242] In a specific example of the disclosed solution, the recommendation adjustment unit is further used to:

[0243] When the first number is less than the preset number, the recommendation weight of the resource to be processed is reduced.

[0244] In a specific example of the disclosed solution, the recommendation adjustment unit is specifically used to:

[0245] When the first number is less than the preset number, using the target resource processing model to obtain a sixth implicit vector of each of the plurality of recommended candidate resources;

[0246] Based on the fourth implicit vector of the resource to be processed and the sixth implicit vector of each recommended candidate resource, obtaining a second similarity between the resource to be processed and each recommended candidate resource;

[0247] Determining a second number of recommended candidate resources whose second similarity is greater than a similarity threshold;

[0248] When the second number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is reduced.

[0249] For the description of the specific functions and examples of each unit of the device in the embodiment of the present disclosure, reference can be made to the relevant description of the corresponding steps in the above method embodiment, which will not be repeated here.

[0250] In the technical solution disclosed herein, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0251] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0252] Fig.13 A schematic block diagram of an example electronic device 1300 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0253] like Fig.13 As shown, the device 1300 includes a computing unit 1301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1302 or a computer program loaded from a storage unit 1308 into a random access memory (RAM) 1303. In the RAM 1303, various programs and data required for the operation of the device 1300 can also be stored. The computing unit 1301, the ROM 1302, and the RAM 1303 are connected to each other via a bus 1304. An input / output (I / O) interface 1305 is also connected to the bus 1304.

[0254] A number of components in the device 1300 are connected to the I / O interface 1305, including: an input unit 1306, such as a keyboard, a mouse, etc.; an output unit 1307, such as various types of displays, speakers, etc.; a storage unit 1308, such as a disk, an optical disk, etc.; and a communication unit 1309, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1309 allows the device 1300 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0255] The computing unit 1301 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1301 performs the various methods and processes described above, such as a training method for a resource processing model or a recommended resource processing method. For example, in some embodiments, the training method for a resource processing model or a recommended resource processing method may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 1308. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1300 via the ROM 1302 and / or the communication unit 1309. When the computer program is loaded into the RAM 1303 and executed by the computing unit 1301, one or more steps of the training method for the resource processing model or the recommended resource processing method described above may be executed. Alternatively, in other embodiments, the computing unit 1301 may be configured to execute a training method for a resource processing model or a recommended resource processing method in any other appropriate manner (eg, by means of firmware).

[0256] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0257] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0258] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0259] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0260] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0261] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0262] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0263] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a resource processing model, comprising: Using the large language model, a target ranking result of N target sample data is obtained; wherein the target ranking result can represent the similarity between the target sample data and the target resource; Based on the target sorting results, M positive and negative sample pairs corresponding to the target resources are obtained; M and N are both integers greater than 1; The target resource and M positive and negative sample pairs corresponding to the target resource are used to perform similarity comparison learning training on the initial resource processing model to obtain the target resource processing model.

2. The method according to claim 1, wherein: The method of using the target resource and the M positive and negative sample pairs corresponding to the target resource to perform similarity comparison learning training on the initial resource processing model includes: Fine-tune the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and / or to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair.

3. The method according to claim 2, wherein: The initial resource processing model is used to encode the input sample and output the implicit vector corresponding to the input sample; The step of fine-tuning the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and / or to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair includes: Using the initial resource processing model, obtain a first implicit vector corresponding to the target resource, a second implicit vector corresponding to the positive sample in the positive-negative sample pair, and a third implicit vector corresponding to the negative sample in the positive-negative sample pair; The adjustable parameters of the initial resource processing model are fine-tuned to reduce the distance between the first implicit vector and the second implicit vector, and / or to increase the distance between the first implicit vector and the third implicit vector.

4. The method according to claim 3, wherein: Fine-tuning adjustable parameters of the initial resource processing model to reduce the distance between the first implicit vector and the second implicit vector includes: Performing an inner product operation on the first implicit vector and the second implicit vector to obtain a first inner product result; Fine-tune the adjustable parameters of the initial resource processing model to reduce the difference between the first inner product result and the first threshold; wherein the first threshold is a theoretical similarity extreme value.

5. The method according to claim 3, wherein: Fine-tuning an adjustable parameter of the initial resource processing model to increase the distance between the first implicit vector and the third implicit vector includes: Performing an inner product operation on the first implicit vector and the third implicit vector to obtain a second inner product result; The adjustable parameters of the initial resource processing model are fine-tuned to reduce the difference between the second inner product result and a second threshold, wherein the second threshold is a theoretical difference extreme value.

6. The method according to any one of claims 3 to 5, wherein: The method of using the initial resource processing model to obtain a first implicit vector corresponding to the target resource, a second implicit vector corresponding to the positive sample in the positive-negative sample pair, and a third implicit vector corresponding to the negative sample in the positive-negative sample pair includes: Using the first network in the initial resource processing model, a first implicit vector corresponding to the target resource is obtained; Using the second network in the initial resource processing model, a second implicit vector corresponding to the positive sample in the positive and negative sample pairs is obtained; Using the third network in the initial resource processing model, a third implicit vector corresponding to the negative sample in the positive and negative sample pairs is obtained; Among them, the values ​​of the weight parameters in the first network, the weight parameters in the second network, and the weight parameters in the third network are associated.

7. The method according to any one of claims 1 to 6, wherein: The method of obtaining M positive and negative sample pairs corresponding to the target resource based on the target sorting result includes: Based on the target sorting result, the N target sample data are grouped to obtain at least two groups; wherein the similarity between the target sample data included in the first group of the at least two groups and the target resource is greater than the similarity between the target sample data included in other groups except the first group and the target resource; Positive samples are selected from the target sample data included in the first group, and negative samples are selected from groups other than the first group to obtain M positive and negative sample pairs.

8. The method according to claim 7, wherein: The N target sample data are grouped based on the target sorting result to obtain at least two groups, including: Based on the target sorting result, N target sample data are grouped to obtain a head group, a middle group and a tail group; wherein the similarity between the target sample data contained in the head group and the target resource is greater than the similarity between the target sample data contained in the middle group and the target resource; the similarity between the target sample data contained in the middle group and the target resource is greater than the similarity between the target sample data contained in the tail group and the target resource; The step of selecting positive samples from the target sample data included in the first group, and selecting negative samples from groups other than the first group, to obtain M positive and negative sample pairs, includes: Positive samples are selected from the target sample data included in the head group, and negative samples are selected from the target sample data included in the middle group or the tail group to obtain M positive and negative sample pairs.

9. The method according to any one of claims 1 to 8, wherein: The method of using the large language model to obtain the target ranking results of N target sample data includes: Using the large language model, the core content of each target sample data among the N target sample data is obtained; Using a large language model and based on the core content of each target sample data, the target ranking results are obtained.

10. The method according to claim 9, wherein: The target ranking results are obtained by using the large language model and based on the core content of each target sample data, including: A sliding window method is adopted, and based on the initial sorting of N target sample data, multiple windows are slidingly determined; wherein each window contains n target sample data, and n represents a preset window length, which is an integer less than or equal to N and greater than 1; Using the large language model and the core content of each target sample data, the sub-sorting results of each window are determined; Based on the sub-sorting results of each window, the target sorting result is obtained.

11. The method according to claim 10, wherein: The sliding window method is adopted, and based on the initial sorting of N target sample data, multiple windows are slidingly determined, including: Based on a preset window length, determining n consecutive target sample data from an initial sort of the N target sample data to determine a window; The window is slid based on a preset step length and along a preset sliding direction to obtain a next window, and a plurality of windows are determined by sliding.

12. A method for processing recommended resources, comprising: Using the target resource processing model, a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historical recommended resources are obtained, wherein the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using the target sorting result; the target sorting result is obtained by sorting N target sample data based on the similarity between the target sample data and the target resource using the large language model; N is an integer greater than 1; Based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each historically recommended resource, obtaining a first similarity between the resource to be processed and each historically recommended resource; Based at least on the first similarity, the recommendation weight of the resource to be processed is adjusted.

13. The method according to claim 12, wherein: The step of adjusting the recommendation weight of the resource to be processed based at least on the first similarity includes: Determine a first number of historically recommended resources having a first similarity greater than a similarity threshold; When the first number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is increased.

14. The method according to claim 13, further comprising: When the first number is less than the preset number, the recommendation weight of the resource to be processed is reduced.

15. The method according to claim 14, wherein: When the first number is less than the preset number, reducing the recommended weight of the to-be-processed resources includes: When the first number is less than the preset number, using the target resource processing model to obtain a sixth implicit vector of each of the plurality of recommended candidate resources; Based on the fourth implicit vector of the resource to be processed and the sixth implicit vector of each recommended candidate resource, obtaining a second similarity between the resource to be processed and each recommended candidate resource; Determining a second number of recommended candidate resources whose second similarity is greater than a similarity threshold; When the second number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is reduced.

16. A training device for a resource processing model, comprising: A sorting unit, used to obtain a target sorting result of N target sample data using a large language model; wherein the target sorting result can represent the similarity between the target sample data and the target resource; A sample determination unit, used to obtain M positive and negative sample pairs corresponding to the target resource based on the target sorting result; M and N are both integers greater than 1; The training unit is used to perform similarity comparison learning training on the initial resource processing model using the target resource and M positive and negative sample pairs corresponding to the target resource to obtain the target resource processing model.

17. The device according to claim 16, wherein: The training unit is specifically used for: Fine-tune the adjustable parameters of the initial resource processing model to increase the similarity between the target resource and the positive sample in the positive-negative sample pair, and / or to reduce the similarity between the target resource and the negative sample in the positive-negative sample pair.

18. The device according to claim 17, wherein: The initial resource processing model is used to encode the input sample and output the implicit vector corresponding to the input sample; the training unit is specifically used to: Using the initial resource processing model, obtain a first implicit vector corresponding to the target resource, a second implicit vector corresponding to the positive sample in the positive-negative sample pair, and a third implicit vector corresponding to the negative sample in the positive-negative sample pair; The adjustable parameters of the initial resource processing model are fine-tuned to reduce the distance between the first implicit vector and the second implicit vector, and / or to increase the distance between the first implicit vector and the third implicit vector.

19. The device according to claim 18, wherein: The training unit is specifically used for: Performing an inner product operation on the first implicit vector and the second implicit vector to obtain a first inner product result; Fine-tune the adjustable parameters of the initial resource processing model to reduce the difference between the first inner product result and the first threshold; wherein the first threshold is a theoretical similarity extreme value.

20. The device according to claim 18, wherein The training unit is specifically used for: Performing an inner product operation on the first implicit vector and the third implicit vector to obtain a second inner product result; The adjustable parameters of the initial resource processing model are fine-tuned to reduce the difference between the second inner product result and a second threshold, wherein the second threshold is a theoretical difference extreme value.

21. The device according to any one of claims 18 to 20, wherein: The training unit is specifically used for: Using the first network in the initial resource processing model, a first implicit vector corresponding to the target resource is obtained; Using the second network in the initial resource processing model, a second implicit vector corresponding to the positive sample in the positive and negative sample pairs is obtained; Using the third network in the initial resource processing model, a third implicit vector corresponding to the negative sample in the positive and negative sample pairs is obtained; Among them, the values ​​of the weight parameters in the first network, the weight parameters in the second network, and the weight parameters in the third network are associated.

22. The device according to any one of claims 16 to 21, wherein: The sample determination unit is specifically used for: Based on the target sorting result, the N target sample data are grouped to obtain at least two groups; wherein the similarity between the target sample data included in the first group of the at least two groups and the target resource is greater than the similarity between the target sample data included in other groups except the first group and the target resource; Positive samples are selected from the target sample data included in the first group, and negative samples are selected from groups other than the first group to obtain M positive and negative sample pairs.

23. The device according to claim 22, wherein: The sample determination unit is specifically used for: Based on the target sorting result, N target sample data are grouped to obtain a head group, a middle group and a tail group; wherein the similarity between the target sample data contained in the head group and the target resource is greater than the similarity between the target sample data contained in the middle group and the target resource; the similarity between the target sample data contained in the middle group and the target resource is greater than the similarity between the target sample data contained in the tail group and the target resource; Positive samples are selected from the target sample data included in the head group, and negative samples are selected from the target sample data included in the middle group or the tail group to obtain M positive and negative sample pairs.

24. The device according to any one of claims 16 to 23, wherein: The sorting unit is specifically used for: Using the large language model, the core content of each target sample data among the N target sample data is obtained; Using a large language model and based on the core content of each target sample data, the target ranking results are obtained.

25. The device according to claim 24, wherein: The sorting unit is specifically used for: A sliding window method is adopted, and based on the initial sorting of N target sample data, multiple windows are slidingly determined; wherein each window contains n target sample data, and n represents a preset window length, which is an integer less than or equal to N and greater than 1; Using the large language model and the core content of each target sample data, the sub-sorting results of each window are determined; Based on the sub-sorting results of each window, the target sorting result is obtained.

26. The device according to claim 25, wherein The sorting unit is specifically used for: Based on a preset window length, determining n consecutive target sample data from an initial sort of the N target sample data to determine a window; The window is slid based on a preset step length and along a preset sliding direction to obtain a next window, and a plurality of windows are determined by sliding.

27. A recommended resource processing device, comprising: The feature determination unit is used to obtain a fourth implicit vector of the resource to be processed and a fifth implicit vector of each of the multiple historical recommended resources using a target resource processing model, wherein the target resource processing model is obtained by performing similarity comparison learning and training on the initial resource processing model using a target sorting result; the target sorting result is obtained by sorting N target sample data using a large language model and based on the similarity between the target sample data and the target resource; N is an integer greater than 1; based on the fourth implicit vector of the resource to be processed and the fifth implicit vector of each of the historical recommended resources, a first similarity between the resource to be processed and each of the historical recommended resources is obtained; The recommendation adjustment unit is used to adjust the recommendation weight of the resource to be processed based on at least the first similarity.

28. The device according to claim 27, wherein The recommendation adjustment unit is specifically used for: Determine a first number of historically recommended resources having a first similarity greater than a similarity threshold; When the first number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is increased.

29. The device according to claim 28, wherein The recommendation adjustment unit is further used for: When the first number is less than the preset number, the recommendation weight of the resource to be processed is reduced.

30. The device according to claim 29, wherein: The recommendation adjustment unit is specifically used for: When the first number is less than the preset number, using the target resource processing model to obtain a sixth implicit vector of each of the plurality of recommended candidate resources; Based on the fourth implicit vector of the resource to be processed and the sixth implicit vector of each recommended candidate resource, obtaining a second similarity between the resource to be processed and each recommended candidate resource; Determining a second number of recommended candidate resources whose second similarity is greater than a similarity threshold; When the second number is greater than or equal to the preset number, the recommendation weight of the resource to be processed is reduced.

31. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 15.

32. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-15.

33. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 15.

Images