Recommendation degree obtaining method, recommendation model training method, recommendation degree obtaining device, recommendation model training device and electronic equipment

Abstract

The invention provides a recommendation degree obtaining method, a recommendation model training method and device and electronic equipment, and relates to the technical field of computers, in particular to the application fields of machine learning, large models, generative search, search engines, information retrieval, knowledge questions and answers and the like. The specific implementation scheme is as follows: acquiring target data; wherein the target data comprises the following contents spliced according to a specific sequence: a target search content, a target reply content and a target recommended search content; fusing the spliced target search content, target reply content and target recommendation search content by using a feature fusion module included in the trained recommendation model to obtain a feature fusion result; determining a target feature unit corresponding to the target recommended search content from the feature fusion result based on a specific sequence; and converting the target feature unit by utilizing a recommendation degree acquisition module included in the recommendation model to obtain a model recommendation degree for the target recommendation search content.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to application fields such as machine learning, large models, generative search, search engines, information retrieval, and knowledge question and answer, and specifically to a method for obtaining a recommendation degree, a training method for a recommendation model, a device, and an electronic device. Background Art

[0002] When using search applications, after a user enters a search term and receives a response from the system regarding the search term, the system may also recommend a series of related search terms to the user, aiming to guide the user to further explore related topics. Summary of the invention

[0003] The present disclosure provides a method for obtaining a recommendation degree, a method for training a recommendation model, a device, and an electronic device.

[0004] According to a first aspect of the present disclosure, a method for obtaining a recommendation degree is provided, comprising:

[0005] Obtaining target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content, and target recommended search content;

[0006] Using the feature fusion module included in the trained recommendation model, the spliced ​​target search content, the target reply content and the target recommendation search content are fused to obtain a feature fusion result;

[0007] Based on a specific order, determine a target feature unit corresponding to the target recommended search content from the feature fusion results;

[0008] The target feature unit is converted by using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

[0009] According to a second aspect of the present disclosure, a training method for a recommendation model is provided, comprising:

[0010] Obtaining a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and a target recommended search content sample;

[0011] Using the feature fusion module included in the recommendation model, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are fused to obtain a feature fusion result sample;

[0012] Based on a specific order, determine a target feature unit sample corresponding to a target recommended search content sample from the feature fusion result sample;

[0013] Using the recommendation degree acquisition module included in the recommendation model, the target feature unit sample is converted to obtain the model recommendation degree for the target recommended search content sample;

[0014] The recommendation model is trained based on the recommendation label and model recommendation degree.

[0015] According to a third aspect of the present disclosure, a device for obtaining a recommendation degree is provided, comprising:

[0016] Obtaining target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content, and target recommended search content;

[0017] Using the feature fusion module included in the trained recommendation model, the spliced ​​target search content, the target reply content and the target recommendation search content are fused to obtain a feature fusion result;

[0018] Based on a specific order, determine a target feature unit corresponding to the target recommended search content from the feature fusion results;

[0019] The target feature unit is converted by using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

[0020] According to a fourth aspect of the present disclosure, a training device for a recommendation model is provided, comprising:

[0021] Obtaining a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and a target recommended search content sample;

[0022] Using the feature fusion module included in the recommendation model, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are fused to obtain a feature fusion result sample;

[0023] Based on a specific order, determine a target feature unit sample corresponding to a target recommended search content sample from the feature fusion result sample;

[0024] Using the recommendation degree acquisition module included in the recommendation model, the target feature unit sample is converted to obtain the model recommendation degree for the target recommended search content sample;

[0025] Based on the recommendation degree label and the model recommendation degree, the recommendation model is trained to obtain a trained recommendation model.

[0026] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0027] at least one processor;

[0028] a memory communicatively coupled to the at least one processor;

[0029] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method provided in the first aspect of the present disclosure.

[0030] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method provided according to the first aspect of the present disclosure.

[0031] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program, which implements the method provided according to the first aspect of the present disclosure when executed by a processor.

[0032] The present disclosure can ensure the accuracy of the final recommendation score for the target recommended search content, thereby improving the user experience.

[0033] 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

[0034] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0035] Figure 1 A flowchart of a training method for a recommendation model provided in an embodiment of the present disclosure;

[0036] Figure 2 An auxiliary illustration of a historical search event provided by an embodiment of the present disclosure;

[0037] Figure 3 A diagram illustrating a process of obtaining a feature fusion result sample provided in an embodiment of the present disclosure;

[0038] Figure 4 A diagram illustrating a process of obtaining a target feature unit sample provided by an embodiment of the present disclosure;

[0039] Figure 5 A diagram illustrating another process of obtaining a target feature unit sample provided by an embodiment of the present disclosure;

[0040] Figure 6 A diagram illustrating a process of obtaining another target feature unit sample provided in an embodiment of the present disclosure;

[0041] Figure 7 A flowchart of a method for obtaining a recommendation degree provided in an embodiment of the present disclosure;

[0042] Figure 8 A schematic diagram of an application scenario of a training method for a recommendation model provided in an embodiment of the present disclosure;

[0043] Fig. 9 A schematic diagram of an application scenario of a method for obtaining a recommendation degree provided in an embodiment of the present disclosure;

[0044] Fig.10 A schematic structural block diagram of a training device for a recommendation model provided in an embodiment of the present disclosure;

[0045] Fig.11 A schematic structural block diagram of a device for obtaining a recommendation degree provided in an embodiment of the present disclosure;

[0046] Fig.12 A schematic structural block diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0047] 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.

[0048] As described in the background technology, during the use of search applications, when the user enters the search content (i.e., Query content) and receives the system's reply information for the search content (i.e., Answer information), the system may also recommend a series of related search content to the user (hereinafter referred to as recommended search content, i.e., Suggesting content) to guide the user to further explore related topics. However, the inventors have found that, at present, after the system recalls multiple candidate recommended search content, each of the multiple candidate recommended search content is usually used as the target recommended search content, and based on the popularity of the target recommended search content in the search of related topics, the final recommendation score for the target recommended search content is obtained. In this process, the user's behavioral characteristics and preferences are not taken into account, and therefore, the accuracy of the final recommendation score for the target recommended search content will be affected, thereby reducing the user experience.

[0049] In view of the above problems, the embodiments of the present disclosure provide a training method for a recommendation model, which can be applied to an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing unit or other similar computing devices. Figure 1 The flowchart diagram shown illustrates a training method for a recommendation model provided by an embodiment of the present disclosure. It should be noted that although a logical order is shown in the flowchart diagram, in some cases, the steps shown or described in the flowchart can also be performed in other orders.

[0050] Step S101, obtaining a target sample.

[0051] Among them, the target sample may include the recommendation degree label of the target recommended search content sample, and the following content spliced ​​in a specific order: target search content sample, target reply content sample and target recommended search content sample. Here, the specific order can be any arrangement order of the target search content sample, the target reply content sample and the target recommended search content sample. For example, the specific order can be: target search content sample—target reply content sample—target recommended search content sample; for another example, the specific order can be: target reply content sample—target search content sample—target recommended search content sample; for another example, the specific order can be: target recommended search content sample—target search content sample—target reply content sample.

[0052] In addition, in the embodiments of the present disclosure, the target search content samples, the target reply content samples and the target recommended search content samples, as well as the recommendation degree labels of the target recommended search content samples can all be obtained based on historical search events.

[0053] Step S102, using the feature fusion module included in the recommendation model, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are fused to obtain a feature fusion result sample.

[0054] The recommendation model may be a neural network model constructed using a feature fusion module and a recommendation degree acquisition module. Here, the feature fusion module is used to fuse the spliced ​​target search content sample, the target reply content sample, and the target recommendation search content sample to obtain a feature fusion result sample. For example, for the target recommendation search content sample, the feature weighting may be performed based on the target search content sample and the target reply content sample to obtain a target feature unit sample corresponding to the target recommendation search content sample. The target feature unit sample is included in the feature fusion result sample.

[0055] Step S103: determining a target feature unit sample corresponding to the target recommended search content sample from the feature fusion result samples based on a specific order.

[0056] Since the feature fusion module included in the recommendation model is used to fuse the spliced ​​target search content sample, target reply content sample and target recommended search content sample to obtain the feature fusion result sample, in the feature fusion result sample, the arrangement order of the three feature unit samples corresponding one by one to the target search content sample, the target reply content sample and the target recommended search content sample is the same as the specific order. Therefore, based on the specific order, the target feature unit sample corresponding to the target recommended search content sample can be determined from the feature fusion result sample.

[0057] Step S104: using the recommendation degree acquisition module included in the recommendation model, convert the target feature unit sample to obtain the model recommendation degree for the target recommended search content sample.

[0058] The recommendation degree acquisition module is used to convert the target feature unit sample to obtain the model recommendation degree for the target recommended search content sample. Here, the model recommendation degree for the target recommended search content sample can be a recommendation score in the interval [0, 1].

[0059] Step S105: training the recommendation model based on the recommendation degree label and the model recommendation degree.

[0060] In one example, the loss value between the recommendation degree label of the target recommended search content sample and the model recommendation degree for the target recommended search content sample can be calculated, and based on the loss value between the recommendation degree label of the target recommended search content sample and the model recommendation degree for the target recommended search content sample, the parameter of the recommendation model is adjusted to obtain a trained recommendation model. The loss function used when calculating the loss value between the recommendation degree label of the target recommended search content sample and the model recommendation degree for the target recommended search content sample can be a cross entropy loss function.

[0061] Using the training method of the recommendation model provided by the embodiment of the present disclosure, on the one hand, after obtaining the target search content sample, the target reply content sample and the target recommended search content sample spliced ​​in a specific order, the feature fusion module included in the recommendation model can be used to fuse the spliced ​​target search content sample, the target reply content sample and the target recommended search content sample. In this way, the user's behavioral characteristics and preferences can be well captured, so that the obtained feature fusion result sample is more in line with the user's interests. On the other hand, after obtaining the feature fusion result sample, the target feature unit sample corresponding to the target recommended search content sample can be determined from the feature fusion result sample based on a specific order, and the target feature unit sample can be converted using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content sample, and then the recommendation model is trained based on the recommendation degree label of the target recommended search content sample and the model recommendation degree for the target recommended search content sample. That is to say, in the process of using the recommendation model for feature processing, the recommendation model will only focus on the target feature unit samples related to the target recommended search content samples, and this targeted feature processing will effectively reduce the noise interference caused by other features, thereby improving the training efficiency and accuracy of the recommendation model.

[0062] Based on the above two points, when the model recommendation degree for the target recommended search content is obtained through the trained recommendation module, the accuracy of the model recommendation degree for the target recommended search content can be ensured, and the model recommendation degree for the target recommended search content can be used as the final recommendation score for the target recommended search content, that is, the accuracy of the final recommendation score for the target recommended search content can be ensured, thereby improving the user experience.

[0063] In addition, it should be noted that in the embodiment of the present disclosure, step S101, step S102, step S103, step S104 and step S105 can be executed in a loop multiple times until the recommendation model meets the convergence condition to obtain the final trained recommendation model. The convergence condition can be that the loss value between the recommendation degree label of the target recommended search content sample and the model recommendation degree for the target recommended search content sample is less than or equal to the preset loss value. Here, the preset loss value can be set according to actual business needs or application scenarios, and the embodiment of the present disclosure does not limit this.

[0064] The target samples in the embodiments of the present disclosure are further described below in combination with specific examples.

[0065] For example, Figure 2 The historical search events shown may include:

[0066] Historical search content: Winter wear 20 words:

[0067] Historical reply content: A warm and fashionable outfit for cold winter, down sweater with long boots, simple yet stylish;

[0068] Historical recommended search content: (1) Are there any other 20-word suggestions for winter outfits? (2) Describe winter outfits more vividly? (3) Are there any other 20-word phrases for winter outfits?

[0069] Then, based on the historical search content of "winter outfit in 20 words", we can get the target search content sample, which can be "winter outfit in 20 words"; based on the historical reply content of "warm and fashionable outfit in cold winter, down sweater with long boots, simple but stylish", we can get the target reply content sample, which can be "warm and fashionable outfit in cold winter, down sweater with long boots, simple but stylish"; based on the three historical recommended search contents of "are there any other suggestions for winter outfit in 20 words; describe winter outfit more vividly; are there any other 20-word phrases recommended for winter outfit", we can get three groups of "target recommended search content samples, and recommendation degree labels of target recommended search content samples". Specifically, assuming that the user clicks on the historical recommended content of "are there any other suggestions for winter outfit in 20 words", but does not click on the two historical recommended contents of "describe winter outfit more vividly; are there any other 20-word phrases recommended for winter outfit", then we can get:

[0070] The first group of “target recommendation search content samples and recommendation degree labels of target recommendation search content samples”: the target recommendation search content sample “Are there any other 20-word suggestions for winter outfits?” and the recommendation degree label for the target recommendation search content sample “Are there any other 20-word suggestions for winter outfits?”, which may be “1”;

[0071] The second group of “target recommended search content samples, and recommendation degree labels of the target recommended search content samples”: the target recommended search content sample “more vividly describe winter outfits”, and the recommendation degree label for the target recommended search content sample “more vividly describe winter outfits”, which may specifically be “0”;

[0072] The third group of “target recommendation search content samples, and recommendation degree labels of target recommendation search content samples”: the target recommendation search content sample “Are there any other 20-word phrases recommended for winter wear?”, and the recommendation degree label for the target recommendation search content sample “Are there any other 20-word phrases recommended for winter wear?”, which can specifically be “0”.

[0073] Finally, three groups of target samples can be obtained:

[0074] The first set of target samples:

[0075] Sample target search content: 20 words about winter outfits;

[0076] Sample target response: A stylish outfit for keeping warm in the cold winter, a down sweater with long boots, simple yet stylish;

[0077] Samples of target recommended search content, and recommendation labels of the sample of target recommended search content: Are there any other 20-word suggestions on winter outfits? 1.

[0078] The second set of target samples:

[0079] Sample target search content: 20 words about winter outfits;

[0080] Sample target response: A stylish outfit for keeping warm in the cold winter, a down sweater with long boots, simple yet stylish;

[0081] Target recommendation search content samples, and recommendation degree labels of target recommendation search content samples: more vividly describe winter outfits; 0.

[0082] The third group of target samples:

[0083] Sample target search content: 20 words about winter outfits;

[0084] Sample target response: A stylish outfit for keeping warm in the cold winter, a down sweater with long boots, simple yet stylish;

[0085] Target recommendation search content samples, and recommendation degree labels of target recommendation search content samples: Are there any other 20-word phrases for winter wear recommendations? 0.

[0086] Further, in some optional implementations, step S101, i.e., “obtaining a target sample” may include:

[0087] Determine the click ratio of multiple historical recommended search contents displayed in the real recommendation scenario;

[0088] Obtain multiple candidate samples according to the click ratio; wherein the ratio of positive and negative samples in the multiple candidate samples is the same as the click ratio;

[0089] Based on multiple candidate samples, a target sample is obtained.

[0090] Continue with Figure 2Taking the historical search event shown as an example, the multiple historical recommended search contents displayed in the real recommendation scene may include: (1) Are there any other 20-word suggestions for winter outfits? (2) Describe winter outfits more vividly? (3) Are there any other 20-word phrases for winter outfits recommended? Among them, the user clicked on the historical recommendation content "Are there any other 20-word suggestions for winter outfits?", but did not click on the two historical recommendation contents "Describe winter outfits more vividly; Are there any other 20-word phrases for winter outfits recommended?". Therefore, the click ratio of the multiple historical recommended search contents displayed in the real recommendation scene is 1 / 2.

[0091] In addition, it should be noted that in the disclosed embodiment, the specific number of multiple candidate samples can be 3.5 million, and the ratio of positive and negative samples in these 3.5 million candidate samples is the same as the click ratio, that is, there can be 1.17 million positive samples and 2.33 million negative samples in these 3.5 million candidate samples. Among them, the positive sample can be a candidate sample with a recommendation label of "1" for the target recommended search content sample; the negative sample can be a candidate sample with a recommendation label of "0" for the target recommended search content sample.

[0092] In one example, after obtaining multiple candidate samples, each of the multiple candidate samples can be used as a target sample. In another example, after obtaining multiple candidate samples, the target sample can also be obtained by:

[0093] Based on multiple candidate samples, samples to be tested are obtained; wherein the samples to be tested include search content samples to be tested, reply content samples to be tested, and recommended search content samples to be tested;

[0094] Obtaining a sample of an object to be inspected; wherein the sample of an object to be inspected includes one of a search content sample to be inspected, a reply content sample to be inspected, and a recommended search content sample to be inspected;

[0095] Determine the second character quantity requirement corresponding to the sample of the object to be tested;

[0096] When the number of internal characters of the sample of the object to be detected meets the second character number requirement, obtaining a target sample based on the sample of the object to be detected;

[0097] When the number of internal characters of the sample of the object to be detected does not meet the second character number requirement, the number of characters of the sample of the object to be detected is adjusted to obtain an adjusted sample of the object to be detected, and a target sample is obtained based on the adjusted sample of the object to be detected.

[0098] In a specific example, after obtaining multiple candidate samples, each of the multiple candidate samples can be used as a sample to be checked, wherein the sample to be checked includes a search content sample to be checked, a reply content sample to be checked, and a recommended search content sample to be checked.

[0099] Thereafter, one of the search content sample to be checked, the reply content sample to be checked, and the recommended search content sample to be checked may be used as the sample of the object to be checked, and the second character quantity requirement corresponding to the sample of the object to be checked may be determined. Exemplarily, when the search content sample to be checked is used as the sample of the object to be checked, the second character quantity requirement corresponding to the object to be checked may be that the number of internal characters of the sample of the object to be checked is less than 180 characters; when the reply content sample to be checked is used as the sample of the object to be checked, the second character quantity requirement corresponding to the object to be checked may be that the number of internal characters of the sample of the object to be checked is less than 100 characters; when the recommended search content sample to be checked is used as the sample of the object to be checked, the second character quantity requirement corresponding to the object to be checked may be that the number of internal characters of the sample of the object to be checked is less than 20 characters.

[0100] Then, when the number of internal characters of the sample of the object to be tested meets the second number of characters requirement, a target sample can be obtained based on the sample of the object to be tested; or, when the number of internal characters of the sample of the object to be tested does not meet the second number of characters requirement, the number of characters of the sample of the object to be tested is adjusted to obtain the adjusted sample of the object to be tested, so as to obtain the target sample based on the adjusted sample of the object to be tested. Among them, "adjusting the number of characters of the sample of the object to be tested" can be truncating the sample of the object to be tested to ensure that the number of internal characters of the sample of the object to be tested meets the second number of characters requirement. For example, the first part of the content of the sample of the object to be tested can be retained, and the second part of the content of the sample of the object to be tested can be deleted to achieve truncation of the sample of the object to be tested.

[0101] Through the above method, in the embodiment of the present disclosure, the click ratio of multiple historical recommended search contents displayed in the real recommendation scene can be determined, and multiple candidate samples can be obtained according to the click ratio, and then the target sample can be obtained based on the multiple candidate samples, and the ratio of positive and negative samples in the multiple candidate samples is the same as the click ratio. In other words, in the embodiment of the present disclosure, it can ensure that there is a relatively reasonable ratio between positive and negative samples, thereby reflecting more realistic user behavior characteristics, and at the same time, it can prompt the recommendation module to learn a more reliable decision boundary, thereby further improving the training efficiency and accuracy of the recommendation model.

[0102] In addition, when obtaining the target sample based on multiple candidate samples, the sample to be tested including the search content sample to be tested, the reply content sample to be tested and the recommended search content sample to be tested is obtained based on the multiple candidate samples, and one of the search content sample to be tested, the reply content sample to be tested and the recommended search content sample to be tested is used as the sample to be tested to determine the second character quantity requirement corresponding to the sample to be tested, and then, when the second internal character quantity of the sample to be tested meets the character quantity requirement, the target sample is obtained based on the sample to be tested; or, when the internal character quantity of the sample to be tested does not meet the second character quantity requirement, the character quantity of the sample to be tested is adjusted to obtain the adjusted sample to be tested, and the target sample is obtained based on the adjusted sample to be tested. In this way, the standardization of the internal character quantity can avoid the situation where any component of the target sample is too long and causes feature overload, making it easier for the recommendation module to extract and understand features, thereby further improving the training efficiency and accuracy of the recommendation model.

[0103] In some optional implementations, step S102, i.e., “using a feature fusion module included in the recommendation model to fuse the spliced ​​target search content sample, the target reply content sample, and the target recommendation search content sample to obtain a feature fusion result sample” may include:

[0104] Using the content decomposition module included in the feature fusion module, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are decomposed to obtain a text unit sequence sample;

[0105] Using the vectorization module included in the feature fusion module, the text unit sequence samples are vectorized to obtain vectorized feature samples;

[0106] The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized feature samples to obtain feature fusion result samples.

[0107] Based on this, it can be understood that in the embodiment of the present disclosure, the recommendation model may include a feature fusion module and a recommendation degree acquisition module connected in sequence. The feature fusion module may include a content decomposition module, a vectorization module and a feature encoding module connected in sequence.

[0108] Please combine Figure 3In the feature fusion module, the content decomposition module (i.e., Tokenizer) can be used to decompose (i.e., word segmentation) the spliced ​​target search content sample, target reply content sample and target recommendation search content sample to obtain text unit sequence samples; the vectorization module can be a feature embedding module (i.e., Embedding module), which can be used to vectorize the text unit sequence samples to obtain vectorized feature samples; the feature encoding module can be a pre-trained language model (Multilingual T5, MT5), which can be used to perform fusion encoding on the vectorized feature samples to obtain feature fusion result samples.

[0109] Through the above method, in the embodiment of the present disclosure, the spliced ​​target search content sample, target reply content sample and target recommendation search content sample can be decomposed by a feature fusion module with a simple model structure (content decomposition module, vectorization module and feature encoding module connected in sequence) to obtain text unit sequence samples, and the text unit sequence samples are vectorized to obtain vectorized feature samples, and then the vectorized feature samples are fusion-encoded to obtain feature fusion result samples. In this process, no complex data processing flow is involved, so the efficiency of obtaining feature fusion result samples can be improved, thereby further improving the training efficiency of the recommendation model.

[0110] In some optional implementations, step S103, that is, "determining a target feature unit sample corresponding to the target recommended search content sample from the feature fusion result sample based on a specific order" may include:

[0111] Determining the number of second text units corresponding to the target recommended search content sample from the text unit sequence sample;

[0112] Based on the specific order and the number of the second text units, a target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

[0113] As mentioned above, when executing step S102, the spliced ​​target search content samples, target reply content samples and target recommendation search content samples can be decomposed to obtain text unit sequence samples, and the text unit sequence samples can be vectorized to obtain vectorized feature samples, and then the vectorized feature samples can be fusion-encoded to obtain feature fusion result samples.

[0114] Based on this, it can be understood that in the embodiment of the present disclosure, the number of second text units, the number of second vectorized feature units, and the number of second fusion feature units are all equal. Among them, the number of second text units can be the number of second text units corresponding to the target recommended search content sample determined from the text unit sequence sample; the number of second vectorized feature units can be the number of second vectorized feature units corresponding to the target recommended search content sample determined from the vectorized feature sample; the number of second fusion feature units can be the number of second fusion feature units corresponding to the target recommended search content sample determined from the feature fusion result sample. Therefore, in the embodiment of the present disclosure, the number of second text units corresponding to the target recommended search content sample can be determined from the text unit sequence sample, and based on the specific order and the number of second text units, the target feature unit sample corresponding to the target recommended search content sample can be determined from the feature fusion result sample.

[0115] Please combine Figure 4 , for example, there are target samples:

[0116] Sample target search content: 20 words about winter outfits;

[0117] Sample target response: A stylish outfit for keeping warm in the cold winter, a down sweater with long boots, simple yet stylish;

[0118] Samples of target recommended search content, and recommendation labels of the sample of target recommended search content: Are there any other 20-word suggestions on winter outfits? 1.

[0119] The specific order of the target search content sample, the target reply content sample and the target recommendation search content sample is: target search content sample—target reply content sample—target recommendation search content sample.

[0120] Then, after decomposing the spliced ​​target search content sample, target reply content sample and target recommendation search content sample to obtain a text unit sequence sample, the second text unit corresponding to the target recommendation search content sample in the text unit sequence sample may include: "and", "other", "winter", "outfit", "20", "word", "of", "suggestion", "?". In other words, the number of second text units can be 9. Therefore, based on a specific order (that is, target search content sample—target reply content sample—target recommendation search content sample) and the number of second text units (that is, 9), 9 second fusion feature units corresponding to the target recommendation search content sample (that is, the 9 second fusion feature units located at the last position in the feature fusion result sample) can be determined from the feature fusion result sample as the target feature unit sample.

[0121] In some other optional implementations, step S103, that is, "determining a target feature unit sample corresponding to the target recommended search content sample from the feature fusion result sample based on a specific order" may also include:

[0122] Determine a second reference number of characters corresponding to the target recommended search content sample; wherein the second reference number of characters is a second internal number of characters of the target recommended search content, or the second reference number of characters is determined based on a character number requirement corresponding to the target recommended search content sample;

[0123] Based on the specific order and the second reference character quantity, a target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

[0124] Wherein, when the second reference number of characters is determined based on the second number of characters requirement corresponding to the target recommended search content sample, the second reference number of characters may be an upper limit of the number of characters of the target recommended search content sample determined based on the second number of characters requirement corresponding to the target recommended search content sample.

[0125] As mentioned above, when executing step S102, the spliced ​​target search content samples, target reply content samples and target recommendation search content samples can be decomposed to obtain text unit sequence samples, and the text unit sequence samples can be vectorized to obtain vectorized feature samples, and then the vectorized feature samples can be fusion-encoded to obtain feature fusion result samples.

[0126] Since the most fragmented decomposition method is that one character corresponds to one second text unit when decomposing the spliced ​​target search content sample, target reply content sample and target recommended search content sample, in the embodiment of the present disclosure, the number of second text units, the number of second vectorized feature units and the number of second fusion feature units must be less than or equal to the second internal number of characters, and less than or equal to the upper limit of the number of characters of the target recommended search content sample. Among them, the second internal number of characters can be the number of internal characters of the target recommended search content sample; the second text unit number can be the number of second text units corresponding to the target recommended search content sample determined from the text unit sequence sample; the second vectorized feature unit number can be the number of second vectorized feature units corresponding to the target recommended search content sample determined from the vectorized feature sample; the second fusion feature unit number can be the number of second fusion feature units corresponding to the target recommended search content sample determined from the feature fusion result sample. Therefore, in an embodiment of the present disclosure, the second internal character number of the target recommended search content or the upper limit of the character number of the target recommended search content sample determined based on the second character number requirement corresponding to the target recommended search content sample can be used as the second reference character number corresponding to the target recommended search content sample, and based on the specific order and the second reference character number, the target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

[0127] Please combine Figure 5 and Figure 6 , for example, there are target samples:

[0128] Sample target search content: 20 words about winter outfits;

[0129] Sample target response: A stylish outfit for keeping warm in the cold winter, a down sweater with long boots, simple yet stylish;

[0130] Samples of target recommended search content, and recommendation labels of the sample of target recommended search content: Are there any other 20-word suggestions on winter outfits? 1.

[0131] The specific order of the target search content sample, the target reply content sample and the target recommendation search content sample is: target search content sample—target reply content sample—target recommendation search content sample.

[0132] Among them, the number of internal characters of the target recommendation search content sample is 14, that is, the second number of internal characters is 14. Therefore, based on a specific order (that is, target search content sample—target reply content sample—target recommendation search content sample) and the second number of internal characters (that is, 14), 14 second fusion feature units corresponding to the target recommendation search content sample (that is, the 14 second fusion feature units located at the last position in the feature fusion result sample) can be determined from the feature fusion result sample and used together as the target feature unit sample. Specifically, it can be as follows Figure 5 shown.

[0133] Assuming that the number of characters corresponding to the target recommended search content sample is required to be less than 20 characters in the target recommended search content sample, that is, the upper limit of the number of characters of the target recommended search content sample is 20, then 20 second fusion feature units corresponding to the target recommended search content sample (that is, the 20 second fusion feature units located at the last position in the feature fusion result sample) can be determined from the feature fusion result sample and used together as the target feature unit sample. Specifically, Figure 6 shown.

[0134] Through the above two methods, in the embodiment of the present disclosure, the number of second text units corresponding to the target recommended search content sample can be determined from the text unit sequence sample, and based on the specific order and the second number of text units, the target feature unit sample corresponding to the target recommended search content sample can be determined from the feature fusion result sample, or the second reference character number corresponding to the target recommended search content sample can be determined (the second reference character number is the second internal character number of the target recommended search content, or the second reference character number is determined based on the second character number requirement corresponding to the target recommended search content sample), and based on the specific order and the second reference character number, the target feature unit sample corresponding to the target recommended search content sample can be determined from the feature fusion result sample. In this way, it is possible to more accurately focus on the target feature unit sample related to the target recommended search content sample, so as to reduce the noise interference caused by other features to a greater extent, thereby further improving the training efficiency and accuracy of the recommendation model.

[0135] Please combine Figure 4 , Figure 5 and Figure 6 After obtaining the target feature unit sample, the target feature unit sample can be converted using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content sample. The recommendation degree acquisition module can be a classifier, such as a multilayer perceptron (MLP), which is not limited in the embodiments of the present disclosure.

[0136] The disclosed embodiment provides a method for obtaining a recommendation degree, which can be applied to an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a car computer, etc.), a personal digital assistant or other similar computing device. Figure 7 The flowchart diagram shown illustrates a method for obtaining a recommendation degree provided by an embodiment of the present disclosure. It should be noted that although a logical sequence is shown in the flowchart diagram, in some cases, the steps shown or described in the flowchart can also be performed in other sequences.

[0137] Step S701, obtaining target data.

[0138] The target data includes the following contents spliced ​​in a specific order: target search content, target reply content, and target recommended search content. Here, the specific order can be any arrangement order of the target search content, the target reply content, and the target recommended search content. For example, the specific order can be: target search content—target reply content—target recommended search content; for another example, the specific order can also be: target reply content—target search content—target recommended search content; for another example, the specific order can be: target recommended search content—target search content—target reply content.

[0139] Step S702, using the feature fusion module included in the trained recommendation model, the spliced ​​target search content, the target reply content and the target recommendation search content are fused to obtain a feature fusion result.

[0140] The trained recommendation model may be a trained neural network model built using a feature fusion module and a recommendation degree acquisition module. Here, the feature fusion module is used to fuse the spliced ​​target search content, target reply content, and target recommended search content to obtain a feature fusion result. For example, for the target recommended search content, feature weighting may be performed based on the target search content and the target reply content to obtain a target feature unit corresponding to the target recommended search content. The target feature unit is included in the feature fusion result.

[0141] Step S703: determining a target feature unit corresponding to the target recommended search content from the feature fusion result based on a specific order.

[0142] Since the feature fusion module included in the trained recommendation model is used to fuse the spliced ​​target search content, target reply content and target recommended search content to obtain the feature fusion result, in the feature fusion result, the arrangement order of the three feature units corresponding to the target search content, target reply content and target recommended search content is the same as the specific order. Therefore, based on the specific order, the target feature unit corresponding to the target recommended search content can be determined from the feature fusion result.

[0143] Step S704: using the recommendation degree acquisition module included in the recommendation model, the target feature unit is converted to obtain the model recommendation degree for the target recommended search content.

[0144] The recommendation degree acquisition module is used to convert the target feature unit to obtain the model recommendation degree for the target recommended search content. Here, the model recommendation degree for the target recommended search content can be a recommendation score in the interval [0, 1].

[0145] By adopting the recommendation degree acquisition method provided by the embodiment of the present disclosure, on the one hand, after obtaining the target search content, target reply content and target recommended search content spliced ​​in a specific order, the spliced ​​target search content, target reply content and target recommended search content can be fused using the feature fusion module included in the trained recommendation model. In this way, the user's behavioral characteristics and preferences can be well captured, so that the obtained feature fusion result samples are more in line with the user's interests. On the other hand, after obtaining the feature fusion result, the target feature unit corresponding to the target recommended search content can be determined from the feature fusion result based on a specific order, and the target feature unit can be converted using the recommendation degree acquisition module included in the trained recommendation model to obtain the model recommendation degree for the target recommended search content. That is to say, in the process of using the recommendation model for feature processing, the recommendation model will only focus on the target feature unit related to the target recommended search content, and this targeted feature processing will effectively reduce the noise interference caused by other features, thereby improving the accuracy of the output result of the trained recommendation model (that is, the model recommendation degree for the target recommended search content).

[0146] The recommendation model may be obtained by training the recommendation module using the aforementioned recommendation model training method.

[0147] Further, in some optional implementations, step S701, i.e., "obtaining target data", may include:

[0148] Obtaining an object to be inspected; wherein the object to be inspected includes one of an initial search content, an initial reply content, and an initial recommended search content;

[0149] Determine the first character quantity requirement corresponding to the object to be inspected;

[0150] When the number of internal characters of the object to be inspected meets the first character number requirement, acquiring target data based on the object to be inspected;

[0151] When the number of internal characters of the object to be inspected does not meet the first character number requirement, the number of characters of the object to be inspected is adjusted to obtain an adjusted object to be inspected, so as to obtain target data based on the adjusted object to be inspected.

[0152] The initial reply content may be obtained by the big model based on the initial search content. The big model may be a pre-trained neural network model that has universal and massive language knowledge, world knowledge, professional knowledge in various fields, etc. In one example, the big model may be an autoregressive generative model of the Transformer architecture.

[0153] In addition, in the embodiment of the present disclosure, the initial recommended search content may be obtained by: obtaining multiple candidate recommended search content recalled by the system; and taking each candidate recommended search content in the multiple candidate recommended content as the initial recommended search content.

[0154] In one example, one of the initial search content, the initial reply content, and the initial recommended search content may be used as an object to be checked, and the first character quantity requirement corresponding to the object to be checked may be determined. Exemplarily, when the initial search content is used as the object to be checked, the first character quantity requirement corresponding to the object to be checked may be that the internal character quantity of the object to be checked is less than 180 characters; when the initial reply content is used as the object to be checked, the first character quantity requirement corresponding to the object to be checked may be that the internal character quantity of the object to be checked is less than 100 characters; when the initial recommended search content is used as the object to be checked, the first character quantity requirement corresponding to the object to be checked may be that the internal character quantity of the object to be checked is less than 20 characters.

[0155] Then, if the number of internal characters of the object to be inspected meets the first character number requirement, the target data can be obtained based on the object to be inspected; or, if the number of internal characters of the object to be inspected does not meet the first character number requirement, the number of characters of the object to be inspected is adjusted to obtain the adjusted object to be inspected, so as to obtain the target data based on the adjusted object to be inspected. Among them, "adjusting the number of characters of the object to be inspected" can be truncating the object to be inspected to ensure that the number of internal characters of the object to be inspected meets the first character number requirement. For example, the first part of the content of the object to be inspected can be retained, and the second part of the content of the object to be inspected can be deleted to achieve truncation of the object to be inspected.

[0156] Through the above method, in the embodiment of the present disclosure, the initial search content, the initial reply content and the initial recommended search content can be obtained, and at least one of the initial search content, the initial reply content and the initial recommended search content can be used as the object to be inspected to determine the first character quantity requirement corresponding to the object to be inspected, and then, when the internal character quantity of the object to be inspected meets the first character quantity requirement, the target data is constructed based on the object to be inspected; or, when the internal character quantity of the object to be inspected does not meet the first character quantity requirement, the character quantity of the object to be inspected is adjusted to obtain the object to be inspected after the character quantity is adjusted, and the target data is constructed based on the object to be inspected after the character quantity is adjusted. In this way, the standardization of the internal character quantity can avoid the situation where any component of the target data is too long and causes feature overload, so that the trained recommendation module can more easily extract and understand features, thereby further improving the accuracy of the output results of the trained recommendation model (that is, the model recommendation degree for the target recommended search content).

[0157] In an optional implementation, step S702, i.e., “using a feature fusion module included in a trained recommendation model to fuse the spliced ​​target search content, target reply content, and target recommendation search content to obtain a feature fusion result” may include:

[0158] Using the content decomposition module included in the feature fusion module, the spliced ​​target search content, target reply content and target recommendation search content are decomposed to obtain a text unit sequence;

[0159] Using the vectorization module included in the feature fusion module, the text unit sequence is vectorized to obtain vectorized features;

[0160] The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized features to obtain the feature fusion result.

[0161] Based on this, it can be understood that in the embodiment of the present disclosure, the trained recommendation model may include a feature fusion module and a recommendation degree acquisition module connected in sequence. The feature fusion module may include a content decomposition module, a vectorization module and a feature encoding module connected in sequence.

[0162] Here, the content decomposition module (i.e., Tokenizer) can be used to decompose (i.e., word segmentation processing) the concatenated target search content, target reply content and target recommended search content to obtain a text unit sequence; the vectorization module can be a feature embedding module (i.e., Embedding module), which can be used to vectorize the text unit sequence to obtain vectorized features; the feature encoding module can be MT5, which can be used to fusion encode the vectorized features to obtain feature fusion results.

[0163] Through the above method, in the embodiment of the present disclosure, the spliced ​​target search content, target reply content and target recommendation search content can be decomposed by a feature fusion module with a simple model structure (a content decomposition module, a vectorization module and a feature encoding module connected in sequence) to obtain a text unit sequence, and the text unit sequence is vectorized to obtain vectorized features, and then the vectorized features are fusion-encoded to obtain feature fusion results. In this process, no complex data processing flow is involved, so the efficiency of obtaining feature fusion results can be improved, thereby improving the efficiency of obtaining the output results of the trained recommendation model (that is, the model recommendation degree for the target recommendation search content).

[0164] In some optional implementations, step S703, i.e., “determining a target feature unit corresponding to the target recommended search content from the feature fusion result based on a specific order” may include:

[0165] Determining, from the text unit sequence, the number of first text units corresponding to the target recommended search content;

[0166] Based on the specific order and the number of the first text units, a target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

[0167] As mentioned above, when executing step S702, the spliced ​​target search content, target reply content and target recommended search content can be decomposed to obtain a text unit sequence, and the text unit sequence can be vectorized to obtain vectorized features, and then the vectorized features can be fusion-encoded to obtain feature fusion results.

[0168] Based on this, it can be understood that in the embodiment of the present disclosure, the number of first text units, the number of first vectorized feature units, and the number of first fusion feature units are all equal. Among them, the number of first text units can be the number of first text units corresponding to the target recommended search content determined from the text unit sequence; the number of first vectorized feature units can be the number of first vectorized feature units corresponding to the target recommended search content determined from the vectorized features; the number of first fusion feature units can be the number of first fusion feature units corresponding to the target recommended search content determined from the feature fusion results. Therefore, in the embodiment of the present disclosure, the number of first text units corresponding to the target recommended search content can be determined from the text unit sequence, and based on the specific order and the number of first text units, the target feature unit corresponding to the target recommended search content can be determined from the feature fusion results.

[0169] For example, there are target data:

[0170] Target search content: 20 words about winter outfits;

[0171] Target reply content: Keep warm and fashionable in cold winter, down sweater with long boots, simple but stylish;

[0172] Target recommended search content: Are there any other 20-word suggestions for winter outfits?

[0173] The specific order of target search content, target reply content and target recommended search content is: target search content—target reply content—target recommended search content.

[0174] Then, after decomposing the spliced ​​target search content, target reply content and target recommended search content to obtain a text unit sequence, the first text unit corresponding to the target recommended search content in the text unit sequence may include: "and", "other", "winter", "outfit", "20", "word", "of", "suggestion", "?". In other words, the number of first text units can be 9. Therefore, based on a specific order (that is, target search content—target reply content—target recommended search content) and the number of first text units (that is, 9), 9 first fused feature units corresponding to the target recommended search content (that is, the 9 first fused feature units located at the last position in the feature fusion result) can be determined from the feature fusion result as the target feature unit.

[0175] In some other optional implementations, step S703, that is, "determining the target feature unit corresponding to the target recommended search content from the feature fusion result based on the specific order" may include:

[0176] Determine a first reference number of characters corresponding to the target recommended search content; wherein the first reference number of characters is a first internal number of characters of the target recommended search content, or the first reference number of characters is determined based on a character number requirement corresponding to the target recommended search content;

[0177] Based on the specific order and the first reference character quantity, a target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

[0178] When the first reference number of characters is determined based on the first character number requirement corresponding to the target recommended search content, the first reference number of characters may be an upper limit on the number of characters of the target recommended search content determined based on the first character number requirement corresponding to the target recommended search content.

[0179] As mentioned above, when executing step S702, the spliced ​​target search content, target reply content and target recommended search content can be decomposed to obtain a text unit sequence, and the text unit sequence can be vectorized to obtain vectorized features, and then the vectorized features can be fusion-encoded to obtain feature fusion results.

[0180] Since the most fragmented decomposition method is that one character corresponds to one first text unit when decomposing the spliced ​​target search content, target reply content and target recommended search content, in the disclosed embodiment, the number of first text units, the number of first vectorized feature units and the number of first fusion feature units must be less than or equal to the first internal number of characters, and less than or equal to the upper limit of the number of characters of the target recommended search content. Among them, the first internal number of characters can be the number of internal characters of the target recommended search content; the first text unit number can be the number of first text units corresponding to the target recommended search content determined from the text unit sequence; the first vectorized feature unit number can be the number of first vectorized feature units corresponding to the target recommended search content determined from the vectorized features; the first fusion feature unit number can be the number of first fusion feature units corresponding to the target recommended search content determined from the feature fusion result. Therefore, in an embodiment of the present disclosure, the first internal character number of the target recommended search content or the upper limit of the character number of the target recommended search content determined based on the first character number requirement corresponding to the target recommended search content can be used as the first reference character number corresponding to the target recommended search content, and based on the specific order and the first reference character number, the target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

[0181] For example, there are target data:

[0182] Target search content: 20 words about winter outfits;

[0183] Target reply content: Keep warm and fashionable in cold winter, down sweater with long boots, simple but stylish;

[0184] Target recommended search content: Are there any other 20-word suggestions for winter outfits?

[0185] The specific order of target search content, target reply content and target recommended search content is: target search content—target reply content—target recommended search content.

[0186] Among them, the number of internal characters of the target recommended search content is 14, that is, the first internal character number is 14. Therefore, based on a specific order (that is, target search content—target reply content—target recommended search content) and the first internal character number (that is, 14), the 14 first fusion feature units corresponding to the target recommended search content (that is, the 14 first fusion feature units located at the last position in the feature fusion result) can be determined from the feature fusion results and collectively used as the target feature units.

[0187] Assuming that the number of characters corresponding to the target recommended search content is required to be that the internal number of characters of the target recommended search content is less than 20 characters, that is, the upper limit of the number of characters of the target recommended search content is 20, then, the 20 first fusion feature units corresponding to the target recommended search content (that is, the 20 first fusion feature units located at the last position in the feature fusion result) can also be determined from the feature fusion results and used together as the target feature units.

[0188] Through the above two methods, in the embodiment of the present disclosure, the first number of text units corresponding to the target recommended search content can be determined from the text unit sequence, and based on the specific order and the first number of text units, the target feature unit corresponding to the target recommended search content can be determined from the feature fusion result, or the first reference character number corresponding to the target recommended search content can be determined (the first reference character number is the first internal character number of the target recommended search content, or the first reference character number is determined based on the first character number requirement corresponding to the target recommended search content), and based on the specific order and the first reference character number, the target feature unit corresponding to the target recommended search content can be determined from the feature fusion result. In this way, it is possible to more accurately focus on the target feature unit related to the target recommended search content, so as to reduce the noise interference caused by other features to a greater extent, thereby further improving the acquisition efficiency and accuracy of the output results of the trained recommendation model (that is, the model recommendation degree for the target recommended search content).

[0189] In some optional embodiments, after obtaining the target feature unit, the recommendation degree acquisition module included in the recommendation model can be used to convert the target feature unit sample to obtain the model recommendation degree for the target recommended search content sample, and based on the model recommendation degree for the target recommended search content sample, the final recommendation score for the target recommended search content can be obtained.

[0190] In one example, the model recommendation degree for the target recommended search content can be used as the final recommendation score for the target recommended search content. In another example, the final recommendation score for the target recommended search content can be obtained by:

[0191] Obtaining a reference recommendation degree for a target recommended search content;

[0192] The model recommendation degree and the reference recommendation degree are integrated to obtain the final recommendation score for the target recommended search content.

[0193] The reference recommendation degree may be calculated based on the target data.

[0194] In a specific example, the reference recommendation degree for the target recommended search content may be obtained in the following manner:

[0195] Based on at least one of the first similarity and the second similarity, a reference recommendation degree for the target recommended search content is acquired.

[0196] Among them, the first similarity is the similarity between the target search content and the target recommended search content, and the second similarity is the similarity between the target reply content and the target recommended search content.

[0197] In a more specific example, the reference recommendation degree for the target recommended search content may be obtained in any of the following three ways:

[0198] First way:

[0199] Calculate the first similarity between the target search content and the target recommended search content;

[0200] Based on the first similarity, a reference recommendation degree for the target recommended search content is obtained.

[0201] In actual implementation, the cosine similarity between the target search content and the target recommended search content may be calculated as the first similarity, and the first similarity may be normalized to obtain a reference recommendation degree for the target recommended search content.

[0202] Second way:

[0203] Calculate the second similarity between the target reply content and the target recommended search content;

[0204] Based on the second similarity, a reference recommendation degree for the target recommended search content is obtained.

[0205] In actual implementation, the cosine similarity between the target reply content and the target recommended search content can be calculated as the second similarity, and the second similarity can be normalized to obtain a reference recommendation degree for the target recommended search content.

[0206] The third way:

[0207] Calculate the first similarity between the target search content and the target recommended search content;

[0208] Calculate the second similarity between the target reply content and the target recommended search content;

[0209] Based on the first similarity and the second similarity, a reference recommendation degree for the target recommended search content is obtained.

[0210] In actual implementation, the cosine similarity between the target search content and the target recommended search content can be calculated as the first similarity, and the cosine similarity between the target reply content and the target recommended search content can be calculated as the second similarity, and the sum of the first similarity and the second similarity can be calculated as the similarity to be processed, and then the similarity to be processed is normalized to obtain a reference recommendation degree for the target recommended search content.

[0211] After obtaining the model recommendation degree and the reference recommendation degree, the model recommendation degree and the reference recommendation degree can be fused to obtain the final recommendation score for the target recommended search content. For example, the model recommendation degree and the reference recommendation degree can be fused through the following fusion logic to obtain the final recommendation score for the target recommended search content:

[0212] Fin_q=x1×Reward_q+x2×Sim_q

[0213] Among them, Fin_q is used to represent the final recommendation score for the target recommended search content; x1 is used to represent the first fusion weight; Reward_q is used to represent the model recommendation degree; x2 is used to represent the second fusion weight; Sim_q is used to represent the reference recommendation degree. Here, the first fusion weight and the second fusion weight can be set according to actual business needs or application scenarios. For example, the first fusion weight can be 6 and the second fusion weight can be 1. The embodiment of the present disclosure does not limit this.

[0214] Through the above method, in the embodiment of the present disclosure, the reference recommendation degree for the target recommended search content can be obtained, and the model recommendation degree and the reference recommendation degree can be fused to obtain the final recommendation score for the target recommended search content, thereby avoiding directly using the model recommendation degree for the target recommended search content as the final recommendation score for the target recommended search content, so as to improve the accuracy of the final recommendation score for the target recommended search content.

[0215] Further, in the embodiment of the present disclosure, after obtaining multiple final recommendation scores corresponding to multiple candidate recommended search contents one by one, based on the multiple final recommendation scores corresponding to multiple candidate recommended search contents one by one, a target number of candidate recommended search contents with the highest final recommendation scores can be selected from the multiple candidate recommended contents as recommended search contents to be sent, and then the target number of recommended search contents to be sent can be sent to the terminal device for display. Among them, the target number can be set according to actual business needs or application scenarios, and the embodiment of the present disclosure does not limit this.

[0216] See also Figure 8 , which is a schematic diagram of an application scenario of a training method for a recommendation model provided in an embodiment of the present disclosure.

[0217] The training method of the recommendation model provided in the embodiment of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. The terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing device or other similar computing device.

[0218] Here, electronic devices are used to:

[0219] Obtaining a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and a target recommended search content sample;

[0220] Using the feature fusion module included in the recommendation model, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are fused to obtain a feature fusion result sample;

[0221] Based on a specific order, determine a target feature unit sample corresponding to a target recommended search content sample from the feature fusion result sample;

[0222] Using the recommendation degree acquisition module included in the recommendation model, the target feature unit sample is converted to obtain the model recommendation degree for the target recommended search content sample;

[0223] The recommendation model is trained based on the recommendation label and model recommendation degree.

[0224] It should be noted that in the embodiments of the present disclosure, Figure 8 The application scenario diagram shown is only illustrative and not restrictive. Those skilled in the art can Figure 8 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0225] See also Fig. 9 , which is a schematic diagram of an application scenario of a method for obtaining a recommendation degree provided in an embodiment of the present disclosure.

[0226] The recommendation degree acquisition method provided by the embodiment of the present disclosure is applied to an electronic device. The electronic device may be a server or a terminal device. Here, the terminal device may be a workbench, a mainframe computer, a conventional computer (e.g., a desktop computer, a laptop computer, a vehicle-mounted computer, etc.), a personal digital processing device or other similar computing devices.

[0227] Here, electronic devices are used to:

[0228] Obtaining target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content, and target recommended search content;

[0229] Using the feature fusion module included in the trained recommendation model, the spliced ​​target search content, the target reply content and the target recommendation search content are fused to obtain a feature fusion result;

[0230] Based on a specific order, determine a target feature unit corresponding to the target recommended search content from the feature fusion results;

[0231] The target feature unit is converted by using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

[0232] It should be noted that in the embodiments of the present disclosure, Fig. 9 The application scenario diagram shown is only illustrative and not restrictive. Those skilled in the art can Fig. 9 Various obvious changes and / or substitutions are made to the examples, and the obtained technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0233] In order to better implement the training method of the recommendation model, the embodiment of the present disclosure also provides a training device for the recommendation model, which can be integrated into an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a mainframe computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device. Fig.10 The schematic structural block diagram shown illustrates a training device 1000 for a recommendation model provided in the disclosed embodiment.

[0234] The training device 1000 of the recommendation model includes:

[0235] The target sample acquisition unit 1001 is used to acquire a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and a target recommended search content sample;

[0236] The feature fusion result sample acquisition unit 1002 is used to use the feature fusion module included in the recommendation model to fuse the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample to obtain a feature fusion result sample;

[0237] A target feature sample determination unit 1003 is used to determine a target feature unit sample corresponding to a target recommended search content sample from the feature fusion result sample based on a specific order;

[0238] The second model recommendation degree acquisition unit 1004 is used to convert the target feature unit sample using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content sample;

[0239] The model training unit 1005 is used to train the recommendation model based on the recommendation degree label and the model recommendation degree.

[0240] In some optional implementations, the feature fusion result sample acquisition unit 1002 is used to:

[0241] Using the content decomposition module included in the feature fusion module, the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample are decomposed to obtain a text unit sequence sample;

[0242] Using the vectorization module included in the feature fusion module, the text unit sequence samples are vectorized to obtain vectorized feature samples;

[0243] The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized feature samples to obtain feature fusion result samples.

[0244] In some optional implementations, the target feature sample determination unit 1003 is used to:

[0245] Determining the number of second text units corresponding to the target recommended search content sample from the text unit sequence sample;

[0246] Based on the specific order and the number of the second text units, a target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

[0247] In some optional implementations, the target feature sample determination unit 1003 is used to:

[0248] Determine a second reference number of characters corresponding to the target recommended search content sample; wherein the second reference number of characters is a second internal number of characters of the target recommended search content; or, the second reference number of characters is determined based on a second number of characters required to correspond to the target recommended search content sample;

[0249] Based on the specific order and the second reference character quantity, a target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

[0250] In some optional implementations, the target sample acquisition unit 1001 is used to:

[0251] Determine the click ratio of multiple historical recommended search contents displayed in the real recommendation scenario;

[0252] Obtain multiple candidate samples according to the click ratio; wherein the ratio of positive and negative samples in the multiple candidate samples is the same as the click ratio;

[0253] Based on multiple candidate samples, a target sample is obtained.

[0254] In some optional implementations, the target sample acquisition unit 1001 is used to:

[0255] Based on multiple candidate samples, samples to be tested are obtained; wherein the samples to be tested include search content samples to be tested, reply content samples to be tested, and recommended search content samples to be tested;

[0256] Acquire a sample of an object to be checked; wherein the sample of an object to be checked includes one of a search content sample to be checked, a reply content sample to be checked, and a recommended search content sample to be checked;

[0257] Determine the second character quantity requirement corresponding to the sample of the object to be tested;

[0258] When the number of internal characters of the sample of the object to be detected meets the second character number requirement, obtaining a target sample based on the sample of the object to be detected;

[0259] When the number of internal characters of the sample of the object to be detected does not meet the second character number requirement, the number of characters of the sample of the object to be detected is adjusted to obtain an adjusted sample of the object to be detected, and a target sample is obtained based on the adjusted sample of the object to be detected.

[0260] In the embodiment of the present disclosure, the specific functions and examples of each unit in the training device 1000 for the recommendation model can be found in the relevant descriptions of the corresponding steps in the aforementioned training method embodiment for the recommendation model, and will not be repeated here.

[0261] In order to better implement the aforementioned recommendation degree acquisition method, the embodiment of the present disclosure also provides a recommendation degree acquisition device, which can be integrated into an electronic device. The electronic device can be a server or a terminal device. Here, the terminal device can be a workbench, a large computer, a conventional computer (for example, a desktop computer, a laptop computer, a car computer, etc.), a personal digital processing or other similar computing device 1100. Fig.11 The schematic structural block diagram shown is used to illustrate a recommendation degree acquisition device 1100 provided in the disclosed embodiment.

[0262] The recommendation degree obtaining device 1000 comprises:

[0263] The target data acquisition unit 1101 is used to acquire target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content and target recommended search content;

[0264] The feature fusion result acquisition unit 1102 is used to fuse the spliced ​​target search content, target reply content and target recommendation search content using the feature fusion module included in the trained recommendation model to obtain a feature fusion result;

[0265] A target feature determination unit 1103 is used to determine a target feature unit corresponding to the target recommended search content from the feature fusion result based on a specific order;

[0266] The first model recommendation degree acquisition unit 1104 is used to convert the target feature unit using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

[0267] In some optional implementations, the feature fusion result acquisition unit 1102 is used to:

[0268] Using the content decomposition module included in the feature fusion module, the spliced ​​target search content, target reply content and target recommendation search content are decomposed to obtain a text unit sequence;

[0269] Using the vectorization module included in the feature fusion module, the text unit sequence is vectorized to obtain vectorized features;

[0270] The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized features to obtain the feature fusion result.

[0271] In some optional implementations, the target feature determination unit 1103 is used to:

[0272] Determining, from the text unit sequence, the number of first text units corresponding to the target recommended search content;

[0273] Based on the specific order and the number of the first text units, a target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

[0274] In some optional implementations, the target feature determination unit 1103 is used to:

[0275] Determine a first reference number of characters corresponding to the target recommended search content; wherein the first reference number of characters is a first internal number of characters of the target recommended search content; or, the first reference number of characters is determined based on a first number of characters required to correspond to the target recommended search content;

[0276] Based on the specific order and the first reference character quantity, a target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

[0277] In some optional implementations, the target data acquisition unit 1101 is used to:

[0278] Acquire an object to be inspected; wherein the object to be inspected includes one of initial search content, initial reply content, and initial recommended search content;

[0279] Determine the first character quantity requirement corresponding to the object to be inspected;

[0280] When the number of internal characters of the object to be inspected meets the first character number requirement, acquiring target data based on the object to be inspected;

[0281] When the number of internal characters of the object to be inspected does not meet the first character number requirement, the number of characters of the object to be inspected is adjusted to obtain an adjusted object to be inspected, so as to obtain target data based on the adjusted object to be inspected.

[0282] In some optional implementations, the recommendation degree acquisition device 1100 further includes a recommendation score acquisition unit, which is used to:

[0283] Based on at least one of the first similarity and the second similarity, obtaining a final recommendation score for the target recommended search content;

[0284] Among them, the first similarity is the similarity between the target search content and the target recommended search content, and the second similarity is the similarity between the target reply content and the target recommended search content.

[0285] In the embodiment of the present disclosure, the specific functions and examples of each unit in the recommendation degree acquisition device 1100 can refer to the relevant descriptions of the corresponding steps in the recommendation degree acquisition method embodiment, and will not be repeated here.

[0286] 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.

[0287] 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.

[0288] Fig.12A schematic block diagram of an example electronic device 1200 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as vehicle-mounted computing devices, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may 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.

[0289] like Fig.12 As shown, the electronic device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from a storage unit 1208 to a random access memory (RAM) 1203. In RAM 1203, various programs and data required for the operation of the electronic device 1200 can also be stored. The computing unit 1201, ROM 1202, and RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0290] A number of components in the electronic device 1200 are connected to the I / O interface 1205, including: an input unit 1206, such as a keyboard, a mouse, etc.; an output unit 1207, such as various types of renderers, speakers, etc.; a storage unit 1208, such as a disk, an optical disk, etc.; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1209 allows the electronic device 1200 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0291] The computing unit 1201 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSP), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1201 performs the various methods and processes described above, such as the training method of the recommendation model and / or the recommendation degree acquisition method. For example, in some embodiments, the training method of the recommendation model and / or the recommendation degree acquisition method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1208. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1200 via the ROM 1202 and / or the communication unit 1209. When the computer program is loaded into RAM 1203 and executed by the computing unit 1201, one or more steps of the training method for the recommendation model and / or the method for obtaining the recommendation degree described above may be performed. Alternatively, in other embodiments, the computing unit 1201 may be configured as the training method for the recommendation model and / or the method for obtaining the recommendation degree by any other appropriate means (e.g., by means of firmware).

[0292] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0293] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data optimization device so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine or completely on a remote machine or server.

[0294] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM) or a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0295] To provide interaction with a user, the systems and techniques described herein may be implemented on a computer having: a rendering device (e.g., a cathode ray tube (CRT) renderer or a liquid crystal display (LCD)) for rendering information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices are also used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0296] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: Local Area Network (LAN), Wide Area Network (WAN), and the Internet.

[0297] The computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0298] The embodiments of the present disclosure also provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute a training method for a recommendation model and / or a method for obtaining a recommendation degree.

[0299] The embodiments of the present disclosure also provide a computer program product, including a computer program, which, when executed by a processor, implements a training method for a recommendation model and / or a method for obtaining a recommendation degree.

[0300] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and this document does not limit this. In addition, in the present disclosure, relational terms such as "first", "second", "third", etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. In addition, "multiple" in the present disclosure can be understood as at least two.

[0301] 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 obtaining a recommendation degree, comprising: Acquire target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content, and target recommended search content; Using a feature fusion module included in the trained recommendation model, the target search content, the target reply content and the target recommendation search content after splicing are fused to obtain a feature fusion result; Based on the specific order, determining a target feature unit corresponding to the target recommended search content from the feature fusion result; The target feature unit is converted by utilizing the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

2. The method according to claim 1, wherein: The feature fusion module included in the trained recommendation model is used to fuse the spliced ​​target search content, the target reply content and the target recommendation search content to obtain a feature fusion result, including: Using the content decomposition module included in the feature fusion module, the target search content, the target reply content and the target recommended search content after splicing are decomposed to obtain a text unit sequence; Using the vectorization module included in the feature fusion module, the text unit sequence is vectorized to obtain vectorized features; The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized features to obtain the feature fusion result.

3. The method according to claim 2, wherein: The determining, based on the specific order, from the feature fusion result, a target feature unit corresponding to the target recommended search content includes: Determining, from the text unit sequence, the number of first text units corresponding to the target recommended search content; Based on the specific order and the first number of text units, the target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

4. The method according to claim 1, wherein: The determining, based on the specific order, from the feature fusion result, a target feature unit corresponding to the target recommended search content includes: Determine a first reference number of characters corresponding to the target recommended search content; wherein the first reference number of characters is a first internal number of characters of the target recommended search content; or, the first reference number of characters is determined based on a first character number requirement corresponding to the target recommended search content; Based on the specific sequence and the first reference character quantity, the target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

5. The method according to claim 1, wherein: The acquiring of target data comprises: Acquire an object to be inspected; wherein the object to be inspected includes one of initial search content, initial reply content, and initial recommended search content; Determining a first character quantity requirement corresponding to the object to be inspected; When the number of internal characters of the object to be inspected meets the first character number requirement, acquiring the target data based on the object to be inspected; When the number of internal characters of the object to be inspected does not meet the first character number requirement, the number of characters of the object to be inspected is adjusted to obtain an adjusted object to be inspected, so as to obtain the target data based on the adjusted object to be inspected.

6. The method according to any one of claims 1 to 5, further comprising: Obtaining a reference recommendation degree for the target recommended search content; The model recommendation degree and the reference recommendation degree are integrated to obtain a final recommendation score for the target recommended search content.

7. The method according to claim 6, wherein: The obtaining of a reference recommendation degree for the target recommended search content includes: Based on at least one of the first similarity and the second similarity, obtaining the reference recommendation degree for the target recommended search content; The first similarity is the similarity between the target search content and the target recommended search content, and the second similarity is the similarity between the target reply content and the target recommended search content.

8. A training method for a recommendation model, comprising: Obtain a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and the target recommended search content sample; Using a feature fusion module included in the recommendation model, the target search content sample, the target reply content sample and the target recommendation search content sample are fused after being spliced ​​to obtain a feature fusion result sample; Based on the specific order, determining a target feature unit sample corresponding to the target recommended search content sample from the feature fusion result sample; Using the recommendation degree acquisition module included in the recommendation model, the target feature unit sample is converted to obtain a model recommendation degree for the target recommended search content sample; The recommendation model is trained based on the recommendation degree label and the model recommendation degree.

9. The method according to claim 8, wherein: The feature fusion module included in the recommendation model is used to fuse the spliced ​​target search content sample, the target reply content sample and the target recommendation search content sample to obtain a feature fusion result sample, including: Using the content decomposition module included in the feature fusion module, the target search content sample, the target reply content sample and the target recommendation search content sample after the splicing are decomposed to obtain a text unit sequence sample; Using the vectorization module included in the feature fusion module, the text unit sequence samples are vectorized to obtain vectorized feature samples; The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized feature samples to obtain the feature fusion result samples.

10. The method according to claim 9, wherein: The determining, based on the specific order, from the feature fusion result samples, a target feature unit sample corresponding to the target recommended search content sample comprises: Determining the number of second text units corresponding to the target recommended search content sample from the text unit sequence sample; Based on the specific order and the second text unit quantity, the target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

11. The method according to claim 8, wherein: The determining, based on the specific order, from the feature fusion result samples, a target feature unit sample corresponding to the target recommended search content sample comprises: Determine a second reference number of characters corresponding to the target recommended search content sample; wherein the second reference number of characters is a second internal number of characters of the target recommended search content; or, the second reference number of characters is determined based on a second number of characters required to correspond to the target recommended search content sample; Based on the specific sequence and the second reference character quantity, the target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

12. The method according to any one of claims 8 to 11, wherein: The obtaining of the target sample comprises: Determine the click ratio of multiple historical recommended search contents displayed in the real recommendation scenario; Acquire multiple candidate samples according to the click ratio; wherein the ratio of positive and negative samples in the multiple candidate samples is the same as the click ratio; Based on the multiple candidate samples, a target sample is obtained.

13. The method according to claim 12, wherein: The step of obtaining a target sample based on the multiple candidate samples includes: Based on the multiple candidate samples, samples to be tested are obtained; wherein the samples to be tested include search content samples to be tested, reply content samples to be tested, and recommended search content samples to be tested; Acquire a sample of an object to be checked; wherein the sample of an object to be checked includes one of the sample of search content to be checked, the sample of reply content to be checked, and the sample of recommended search content to be checked; Determining a second character quantity requirement corresponding to the sample of the object to be tested; When the number of internal characters of the sample of the object to be detected meets the second character number requirement, acquiring the target sample based on the sample of the object to be detected; When the number of internal characters of the sample of the object to be detected does not meet the second character number requirement, the number of characters of the sample of the object to be detected is adjusted to obtain an adjusted sample of the object to be detected, so as to obtain the target sample based on the adjusted sample of the object to be detected.

14. A device for obtaining a recommendation degree, comprising: A target data acquisition unit, used to acquire target data; wherein the target data includes the following contents spliced ​​in a specific order: target search content, target reply content and target recommended search content; A feature fusion result acquisition unit, used to fuse the spliced ​​target search content, the target reply content and the target recommended search content using a feature fusion module included in the trained recommendation model to obtain a feature fusion result; a target feature determination unit, configured to determine, based on the specific order, a target feature unit corresponding to the target recommended search content from the feature fusion result; The first model recommendation degree acquisition unit is used to convert the target feature unit using the recommendation degree acquisition module included in the recommendation model to obtain the model recommendation degree for the target recommended search content.

15. The device according to claim 14, wherein: The feature fusion result acquisition unit is used for: Using the content decomposition module included in the feature fusion module, the target search content, the target reply content and the target recommended search content after splicing are decomposed to obtain a text unit sequence; Using the vectorization module included in the feature fusion module, the text unit sequence is vectorized to obtain vectorized features; The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized features to obtain the feature fusion result.

16. The device according to claim 15, wherein: The target feature determination unit is used for: Determining, from the text unit sequence, the number of first text units corresponding to the target recommended search content; Based on the specific order and the first number of text units, the target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

17. The device according to claim 14, wherein: The target feature determination unit is used for: Determine a first reference number of characters corresponding to the target recommended search content; wherein the first reference number of characters is a first internal number of characters of the target recommended search content, or the first reference number of characters is determined based on a first character number requirement corresponding to the target recommended search content; Based on the specific sequence and the first reference character quantity, the target feature unit corresponding to the target recommended search content is determined from the feature fusion result.

18. The device according to claim 14, wherein: The target data acquisition unit is used for: Acquire an object to be inspected; wherein the object to be inspected includes one of initial search content, initial reply content, and initial recommended search content; Determining a first character quantity requirement corresponding to the object to be inspected; When the number of internal characters of the object to be inspected meets the first character number requirement, acquiring the target data based on the object to be inspected; When the number of internal characters of the object to be inspected does not meet the first character number requirement, the number of characters of the object to be inspected is adjusted to obtain an adjusted object to be inspected, so as to obtain the target data based on the adjusted object to be inspected.

19. The device according to any one of claims 14 to 18, further comprising a recommendation score obtaining unit, configured to: Obtaining a reference recommendation degree for the target recommended search content; The model recommendation degree and the reference recommendation degree are integrated to obtain a final recommendation score for the target recommended search content.

20. The device according to claim 19, wherein The recommendation score acquisition unit is used for: Based on at least one of the first similarity and the second similarity, obtaining the reference recommendation degree for the target recommended search content; The first similarity is the similarity between the target search content and the target recommended search content, and the second similarity is the similarity between the target reply content and the target recommended search content.

21. A training device for a recommendation model, comprising: A target sample acquisition unit, used to acquire a target sample; wherein the target sample includes a recommendation degree label of a target recommended search content sample, and the following contents spliced ​​in a specific order: a target search content sample, a target reply content sample, and the target recommended search content sample; A feature fusion result sample acquisition unit is used to use the feature fusion module included in the recommendation model to fuse the spliced ​​target search content sample, the target reply content sample and the target recommended search content sample to obtain a feature fusion result sample; a target feature sample determining unit, configured to determine, based on the specific order, a target feature unit sample corresponding to the target recommended search content sample from the feature fusion result sample; A second model recommendation degree acquisition unit, configured to convert the target feature unit sample using the recommendation degree acquisition module included in the recommendation model to obtain a model recommendation degree for the target recommended search content sample; A model training unit is used to train the recommendation model based on the recommendation degree label and the model recommendation degree.

22. The device according to claim 21, wherein The feature fusion result sample acquisition unit is used for: Using the content decomposition module included in the feature fusion module, the target search content sample, the target reply content sample and the target recommendation search content sample after splicing are decomposed to obtain a text unit sequence sample; Using the vectorization module included in the feature fusion module, the text unit sequence samples are vectorized to obtain vectorized feature samples; The feature encoding module included in the feature fusion module is used to perform fusion encoding on the vectorized feature samples to obtain the feature fusion result samples.

23. The device according to claim 22, wherein: The target feature sample determination unit is used for: Determining the number of second text units corresponding to the target recommended search content sample from the text unit sequence sample; Based on the specific order and the second text unit quantity, the target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

24. The device according to claim 21, wherein The target feature sample determination unit is used for: Determine a second reference number of characters corresponding to the target recommended search content sample; wherein the second reference number of characters is a second internal number of characters of the target recommended search content, or the second reference number of characters is determined based on a second number of characters required to correspond to the target recommended search content sample; Based on the specific sequence and the second reference character quantity, the target feature unit sample corresponding to the target recommended search content sample is determined from the feature fusion result sample.

25. The device according to any one of claims 21 to 24, wherein: The target sample acquisition unit is used for: Determine the click ratio of multiple historical recommended search contents displayed in the real recommendation scenario; Acquire multiple candidate samples according to the click ratio; wherein the ratio of positive and negative samples in the multiple candidate samples is the same as the click ratio; Based on the multiple candidate samples, a target sample is obtained.

26. The device according to claim 25, wherein The target sample acquisition unit is used for: Based on the multiple candidate samples, samples to be tested are obtained; wherein the samples to be tested include search content samples to be tested, reply content samples to be tested, and recommended search content samples to be tested; Acquire a sample of an object to be checked; wherein the sample of an object to be checked includes one of the sample of search content to be checked, the sample of reply content to be checked, and the sample of recommended search content to be checked; Determining a second character quantity requirement corresponding to the sample of the object to be tested; When the number of internal characters of the sample of the object to be detected meets the second character number requirement, acquiring the target sample based on the sample of the object to be detected; When the number of internal characters of the sample of the object to be detected does not meet the second character number requirement, the number of characters of the sample of the object to be detected is adjusted to obtain an adjusted sample of the object to be detected, so as to obtain the target sample based on the adjusted sample of the object to be detected.

27. An electronic device comprising: at least one processor; a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 13.

28. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 13.

29. 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 13.

Images